Molecular Property Prediction Based on Dual-Channel Feature Separation Network and Contrastive Learning

Molecular Property Prediction Based on Dual-Channel Feature Separation Network and Contrastive Learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Molecular Property Prediction Based on Dual-Channel Feature Separation Network and Contrastive Learning Yishan Zhu, Qian Zhou, Faming Lu, Guanglei Cui, Cong Liu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8846431/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Contrastive learning, a prominent self-supervised technique, has shown great potential in molecular property prediction by reducing reliance on labeled data. However, existing methods face challenges such as disrupting molecular structures through aggressive graph augmentations and feature interference from extracting all features within a single network. To address these issues, this paper propose DCFS-CL (Dual-Channel Feature Separation with Contrastive Learning). First, a property-preserving augmentation strategy modifies non-essential structures while retaining key scaffolds. Then, a dual-channel network is introduced to separately extract inherent and auxiliary features, enhancing interpretability and task adaptability. As a result, DCFS-CL achieves outstanding accuracy and robustness across seven biological classification datasets and demonstrates high predictive performance on an electrochemical regression dataset. This framework offers strong generalization, making it well-suited for molecular screening and property prediction in data-scarce scenarios. Molecular Property Prediction Self-Supervised Learning Molecular Contrastive Learning Dual-Channel Feature Separation Network Intrinsic Features Auxiliary Features Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1 Introduction Materials science, electrochemistry and nanotechnology are important fields in modern research[ 23 , 24 , 25 ].Molecular property prediction is a core task in modern chemistry, materials science, and drug development. Accurate prediction models typically require large amounts of labeled data to achieve satisfactory performance. However, in many real-world scenarios, obtaining large-scale annotated datasets for molecular properties is infeasible—especially for electrochemical properties, which often rely on complex experiments or high-precision quantum chemical calculations, making data acquisition costly. Moreover, some of these properties are inherently rare. Self-Supervised Learning (SSL) can automatically learn useful feature representations from a large volume of unlabeled data by designing appropriate pretext tasks. It has attracted widespread attention in molecular property prediction and has demonstrated great potential in electrochemical molecular modeling tasks. Contrastive learning, as a typical self-supervised learning method, is a widely adopted technique for chemistry. The core idea of contrastive learning is to construct similar pairs and learn feature representations of data by maximizing the consistency within these similar pairs. In molecular representation tasks, current contrastive learning methods generally rely on graph augmentation techniques to generate similar pairs, including node deletion, edge perturbation, and subgraph extraction [ 4 ], whose goal is to generate multiple views that guide the model to learn more robust feature representations. Hu et al. introduced GraphCL[ 1 ] by generating multiple views (e.g., node dropping, edge perturbation) to strengthen the robustness of molecular graph representations. GraphMAE employs a masking strategy (i.e., randomly masking node or edge information) to enable the model to learn both global and local structural features of molecular graphs, enhancing its self-supervised representation learning ability [ 2 ]. However, such graph augmentation strategies can easily lead to information loss, introduce noise or bias, neglect local-global relationships within molecules, and demonstrate insufficient adaptability to downstream tasks. For example, node deletion and edge removal strategies may disrupt important chemical groups or functional groups in the molecule, causing the model to lose critical semantic information. Specifically, in chemical property prediction tasks, the loss of certain atoms or chemical bonds can significantly affect the molecular properties. Moreover, classic methods for molecular graph contrastive learning often extract all features within the same network [ 7 , 8 , 9 , 10 ]. DGI (Deep Graph Infomax) proposed learning global graph representations by maximizing the mutual information between local and global feature[ 12 ]. MGSSL (Multi-Granularity Self-Supervised Learning) uses a multi-granularity graph augmentation strategy to capture multi-scale features of molecular graphs through global and local view contrasts, thereby improving the accuracy of molecular property prediction [ 3 ]. GraphMVP proposes a multi-view pretraining framework, combining the 2D topology of molecular graphs and 3D molecular conformations, capturing the physicochemical properties of molecules and performing excellently in molecular property prediction [ 11 ].MoleMCL captures key chemical features of molecules by comparing augmented views at different structural levels, such as scaffolds and functional groups, thereby improving molecular property prediction performance [ 7 ]. In molecular graphs, nodes, edges, and global structures contain rich information, such as atom types, bond types, and spatial conformations. This information is progressively compressed into a low-dimensional embedding vector, when the features are extracted by neural networks. When all features of molecular graphs are generally extracted by one single neural network model, it may result in the loss of certain details due to information compression. For example, atomic information and stereoconformational information could be mixed into the same embedding vector, leading to the dilution of key features. For complex molecular property prediction tasks, this information loss can hinder the model's ability to capture essential chemical characteristics. In addition, different molecular property prediction tasks exhibit varying levels of dependence on atomic information and spatial conformations data. For instance, toxicity prediction of a molecule places greater emphasis on the local chemical environment, whereas solubility prediction may focus more on the overall shape of the molecule. If the single model fails to adapt to the task requirements, it may result in suboptimal performance. Jiang et al. [ 13 ] designed a property-preserving molecular graph augmentation method, which does not remove edges or nodes from the molecular graph but instead changes their positions to generate augmented graphs, thereby preserving the fundamental features. This method converts the original molecular graph into two isomers, selectively retaining the molecular scaffold or functional groups, emphasizing the preservation of key structures that significantly influence the chemical properties of the molecule. By adjusting nonessential structures, the method generates diverse augmented samples while minimizing disruption to the original molecular semantics. However, previous methods only used scaffold augmentation or functional group augmentation individually for transformations and applied them in supervised learning tasks. In this study, this work present DCFS-CL(Dual-Channel Feature Separation with Contrastive Learning is proposed for molecular property prediction. Firstly, this work combine scaffold augmentation and functional group augmentation to flexibly adjust the layout of functional groups while ensuring the stability of the key scaffold structure, generating diverse molecular graphs without losing critical information. Then, a dual-channel feature separation network is proposed to extract inherent and auxiliary features, which can decouple these features and avoid interference caused by the mixing of feature types. This decoupling allows for better control over the feature extraction process, enabling the model to focus on the correct feature type when handling molecular structural changes. Additionally, semantic changes in molecular graphs often affect the feature extraction performance of traditional GNNs. By using feature extraction networks, these semantic changes can be more sensitively captured. This improves the model’s sensitivity to small molecular changes, thereby more accurately describing the differences between molecules. As a result, DCFS-CL demonstrates outstanding accuracy and robustness in regression tasks on an electrochemical dataset and classification tasks on seven biological datasets.In the following sections, this work will provide a detailed introduction to each aspect. Section 2 will focus on the graph augmentation strategy, contrastive learning with the dual-channel feature separation network, and the overall framework of the model. In Sections 3 and 4 , this work will discuss the performance differences between DCFS-CL and other models on electrochemical and biological datasets, along with the underlying reasons. Finally, this work will conclude the paper and explore future directions for optimization. 2 Methods This section systematically introduces the molecular property prediction model framework and its theoretical support. First, this work analyze the accompanying features of molecular properties and describe what these features represent. Next, a molecular graph-preserving augmentation strategy is introduced to generate new data samples by enhancing the structure of molecular graphs with chemical plausibility. Finally, the molecular property prediction model framework is explained in detail, which is based on a dual-channel feature separation network and contrastive learning. 2.1Accompanying features of molecular properties In molecular property prediction, the associated features in the molecular structure play a crucial role in describing molecular behavior, reaction mechanisms, and properties. These associated features include atomic type, bond type, atomic chirality, and bond stereochemistry. Atomic Type refers to the element type in molecule, such as carbon, hydrogen, oxygen, etc. As the basic units of functional groups, atomic type plays an important role in the chemical properties of molecule, such as polarity, reactivity, and stability. In a molecular graph, each node is typically considered an atom, whose type is used as an inherent property of the node in the model. Bond Type describes the chemical bonds connecting different atoms, including single, double, and triple bonds. These bond types affect the molecule's geometric structure and energy state, significantly impacting on the molecule's reaction pathways and physicochemical properties. Bond types are typically input as edge properties in the molecular graph. Atomic Chirality refers to the spatial arrangement of atoms around a chiral center (such as R/S configurations), which leads to different stereoisomers of the molecule. These isomers may exhibit vastly different properties in biological activity and pharmacological effects. Atomic chirality is an auxiliary feature of the molecule, with high discriminative ability for certain molecular prediction tasks. Bond Stereochemistry describes the spatial configuration of bonds (e.g., Z/E configurations), which significantly impacts on the three-dimensional structure and spatial arrangement of the molecule. Molecules often exhibit differences in solubility, polarity, and pharmacokinetics due to these features. 2.2 Molecular graph augmentation In molecular representation tasks, the generation of positive pairs often relies on graph augmentation techniques. Typical augmentation methods include node deletion, edge perturbation, and subgraph extraction, aimed at generating multiple views to help the model learn robust feature representations. However, these augmentation methods may disrupt the overall structure of the molecule or important chemical information on the molecular graph. For example, deleting nodes or edges may destroy key functional groups in the molecule, causing the model to lose crucial semantic information. This is especially problematic in certain chemical property prediction tasks, where the loss of specific atoms or bonds can significantly impact the molecular performance. This work introduce a novel graph augmentation strategy. It does not remove elements but instead swaps the positions of nodes or edges, selectively preserving the molecular scaffold and functional groups. This approach emphasizes retaining key structures that have a significant impact on the chemical properties of the molecule. First, this work extract the scaffold graph of the molecule and mask its edges. Since functional groups are more dispersed than the scaffold, this work do not construct a line graph here, but instead directly reconnect the edges of the original graph. This work randomly select an edge \(\:{\text{e}}_{1}\) =( \(\:{\text{v}}_{1}\) , \(\:{\text{v}}_{2}\) , ω)and disconnect it. Modifying edges in the molecular graph leads to changes in the molecular topology, which in turn affects the properties of the molecule, including the number of hydrogens and the number of charges. Therefore, here, this work update the number of hydrogens as follows: $$\:{\text{h}}_{\text{i}}={\text{h}}_{\text{i}}+{\omega\:}$$ 1 In the equation, ω represents the valency of the broken edge. Then, two randomly selected non-connected nodes are connected to form a new edge \(\:{\text{e}}_{1}^{{\prime\:}}\) =( \(\:{\text{v}}_{3}\) , \(\:{\text{v}}_{4}\) , ω). Similarly, the number of hydrogen atoms in these nodes should also be updated here. $$\:{\text{h}}_{\text{j}}={\text{h}}_{\text{j}}-{\omega\:}$$ 2 In addition, when \(\:{\text{h}}_{\text{j}}\) is less than 0, this work need to update the number of electrons. $$\:{\text{c}}_{\text{j}}={\text{c}}_{\text{j}}+{\text{h}}_{\text{j}}$$ 3 Let \(\:{\text{h}}_{\text{j}}\) =0. After that apply ScaffoldAug. ScaffoldAug is a method that generates isomers by swapping edges in the molecular scaffold without changing the functional groups of the molecule. Functional groups are often structural units closely related to the chemical properties of the molecule. To reduce the search space of edge-swapping operations, this work convert node selection into edge selection. ScaffoldAug first extracts the scaffold graph from the molecule and converts it into a line graph. Then, two nodes are selected from the line graph and mapped to two edges in the scaffold graph. Finally, this work perform the edge-swapping operation and use chemical rules to filter out invalid isomers. Figure 1 illustrates the main architecture of ScaffoldAug, with the specific details of each generation step shown below. • Line graph construction As shown in Fig. 2 , this work first describes how to construct the line graph of a molecular scaffold. Given a graph \(\:G\) , this work denote its line graph as \(\:{L}_{G}\) = ( \(\:{A}_{L}\) , \(\:{X}_{L}\) , \(\:{W}_{L}\) ). The construction of the line graph is as follows: each edge in the original graph \(\:\:G\) corresponds to a node in the line graph.If two edges in \(\:G\:\) share a common node.Then,in \(\:{L}_{\text{G}}\) , an edge will be created between the corresponding nodes \(\:{v}_{ij}\) and \(\:{v}_{jk}\) .Specifically, this work first use the open-source toolkit RDKit to extract the scaffold graph and transform it into a line graph \(\:{L}_{{G}_{S}}\) .Then, to prepare for the next step of node selection, this work represent each element \(\:{a}_{ij,jk}\) in the adjacency matrix of \(\:{L}_{{G}_{S}}\) as: $$\:{a}_{ij,jk}\:=\left\{\begin{array}{c}1,\:\:\:if\:{v}_{ij},\:{v}_{jk}\:are\:connected\:\\\:0,\:\:otherwise\end{array}\right.$$ 4 In this way, the operation of selecting four nodes can be simplified to the operation of selecting two edges when swapping edges. • Node selection After constructing the line graph, this work first randomly select a node \(\:{\text{v}}_{\text{i}}\) in \(\:{\text{L}}_{{\text{G}}_{\text{S}}}\) . Then, to reduce the search space as described above, this work introduce the adjacency matrix of \(\:{\text{L}}_{{\text{G}}_{\text{S}}}\) . Specifically, the operation of reducing the search space is divided into the following steps. First, to avoid selecting edges with common nodes during the edge-swapping process, or generating edges that already exist in the original graph, this work perform some masking operations on the nodes using the adjacency matrix. Let \(\:\text{i}\) be the index of the first node selected for edge swapping. Here, this work introduce a mask vector \(\:{M}_{1}\) = { \(\:{m}_{1}\) , \(\:{m}_{2}\) , ... , \(\:{m}_{\left|{V}_{L}\right|}\) } . To ensure that the same node is not selected, this work set each element of \(\:{\text{M}}_{1}\) as follows: $$\:{m}_{u}\:=\left\{\begin{array}{c}1,\:if\:{a}_{ui}=0,u=1,...,\left|{V}_{L}\right|\:\\\:0,\:otherwise\end{array}\right.$$ 5 Then, to avoid reconnecting to edges that already exist, this work perform a logical AND operation between the \(\:i\) -th row of the adjacency matrix and every other row. Each element of the matrix is represented as follows: $$\:{a}_{uv}^{{\prime\:}}={a}_{uv}·{a}_{iv},\:u,\:v\:=\:1,\:...,\:\left|{V}_{L}\right|$$ 6 Where \(\:|{V}_{L}\) | is the number of nodes in \(\:{L}_{{G}_{S}}\) . After this operation, if the \(\:v\) -th element of the \(\:u\) -th row is 1, the \(\:v\) -th row will be logically ANDed with both the \(\:u\) -th row and the \(\:\:i\) -th row. If any row contains an element equal to 1, the corresponding node will be masked, i.e., \(\:{\:M}_{1}\) will have \(\:{m}_{u}\) = 0. In addition, to prevent the generated graph from deviating too much from the original graph due to significant changes in the scaffold, this work also mask the corresponding nodes in the line graph where entire rows become zero after the first logical AND operation. This work can restrict the edge-swapping operation to nodes within its 2-hop neighborhood. Finally, to further reduce the probability of edge-swapping failure, this work also mask the corresponding edges with different valency values. To achieve this, this work construct another mask vector \(\:{M}_{2}\) = { \(\:{m}_{1}\) , \(\:{m}_{2}\) , ... , \(\:{m}_{\left|{V}_{L}\right|}\) } , and set each element of \(\:{M}_{2}\) as follows: $$\:{m}_{u}\:=\left\{\begin{array}{c}1,\:if\:{w}_{u}\:={w}_{i}\\\:0,\:otherwise\end{array}\right.$$ 7 After all the above masking operations, this work randomly select an index \(\:\text{j}\) where both elements in \(\:{\text{M}}_{1}\) and \(\:{\text{M}}_{2}\) are 1 as the second node \(\:{\text{v}}_{\text{j}}\) . • Swap edges In the above method, this work have selected two nodes \(\:{\text{v}}_{\text{i}}\) and \(\:{\text{v}}_{\text{j}}\) in the line graph to prepare for edge swapping. Then convert these two nodes into the corresponding edges in the scaffold graph to obtain \(\:{\text{e}}_{\text{s}1}\) and \(\:{\text{e}}_{\text{s}2}\) . This work ultimate goal is to swap edges in the original molecular graph, so here this work need to map \(\:{\text{e}}_{\text{s}1}\) and \(\:{\text{e}}_{\text{s}2}\) in the scaffold graph to the corresponding edges in the original graph \(\:\text{G}\) , i.e., this work obtain \(\:{\text{e}}_{1}\) =( \(\:{\text{v}}_{1}\) , \(\:{\text{v}}_{2}\) , ω) and \(\:{\text{e}}_{2}\) =༈ \(\:{\text{v}}_{3}\) , \(\:{\text{v}}_{4}\) , ω༉.Then, disconnect \(\:{\text{e}}_{1}\) , \(\:{\text{e}}_{2}\) and reconnect them as \(\:{\text{e}}_{1}\) =༈ \(\:{\text{v}}_{1}\) , \(\:{\text{v}}_{4}\) , ω༉and \(\:{\text{e}}_{2}\) =༈ \(\:{\text{v}}_{2}\) , \(\:{\text{v}}_{3}\) , ω༉.During the generation process, multiple modifications can be made until the modification rate or the number of failures exceeds the number of edges in the original graph. 2.3 The framework of the molecular property prediction model based on a dual-channel feature separation network and contrastive learning DCFS-CL is developed based on MolCLR [ 4 ]. The latent representations of positive augmented molecular graph pairs are contrasted with those of negative augmented molecular graph pairs. The entire model (Fig. 3 ) consists of four components: data processing and augmentation, a GNN-based feature extractor, a nonlinear projection head, and the normalized temperature-scaled cross-entropy (NT-Xent) [ 6 ] contrastive loss. For molecular property prediction, this work first construct the SMILES data \(\:{\text{s}}_{\text{n}}\) into the corresponding molecular graph \(\:{\text{G}}_{\text{n}}\) . In these molecular graphs, each node represents an atom, and each edge represents a chemical bond between atoms. Next, this work apply molecular graph augmentation strategies to transform the original molecular graph \(\:{\text{G}}_{\text{n}}\) into two different but related molecular graphs \(\:{\text{G}}_{\text{n}\_1}\) and \(\:{\text{G}}_{\text{n}\_2}\) . When constructing positive and negative sample pairs, if the two augmented graphs come from the same molecule (i.e., \(\:{\text{G}}_{\text{n}\_1}\) and \(\:{\text{G}}_{\text{n}\_2}\) are generated from the same \(\:{\text{G}}_{\text{n}}\) ), they are considered a positive sample pair. If the augmented graphs come from different molecules, they are considered a negative sample pair. This work can train the model using contrastive learning, allowing it to learn the similarities and differences between molecules. To extract features from molecular graphs, this work use a pair of Graph Neural Networks (GNNs) as feature extractors, each responsible for extracting different types of features. The intrinsic feature extractor mainly extracts the inherent features \(\:{\text{h}}_{\text{n}\_1}^{\text{i}\text{n}\text{h}}\) , \(\:{\text{h}}_{\text{n}\_2}^{\text{i}\text{n}\text{h}}\) of the molecules, such as atom types, bond types, etc. These are the foundational features of the molecular graph, and after graph augmentation, these features do not change for positive pairs. The auxiliary feature extractor mainly extracts the auxiliary features \(\:{\text{h}}_{\text{n}\_1}^{\text{a}\text{u}\text{x}}\) , \(\:{\text{h}}_{\text{n}\_2}^{\text{a}\text{u}\text{x}}\) of the molecules, such as atomic chirality information, bond stereochemistry information, etc. These features capture finer structural differences between molecules, and after graph augmentation, these features will change. After feature extraction, this work further process the representations \(\:{\text{h}}_{\text{n}\_1}^{\text{i}\text{n}\text{h}}\) , \(\:{\text{h}}_{\text{n}\_2}^{\text{i}\text{n}\text{h}}\) , \(\:{\text{h}}_{\text{n}\_1}^{\text{a}\text{u}\text{x}}\) , \(\:{\text{h}}_{\text{n}\_2}^{\text{a}\text{u}\text{x}}\) through a nonlinear projection head g(·). This projection head consists of a multi-layer perceptron (MLP) with one hidden 7layer, which maps the feature vectors \(\:{\text{h}}_{\text{n}\_1}^{\text{i}\text{n}\text{h}}\) , \(\:{\text{h}}_{\text{n}\_2}^{\text{i}\text{n}\text{h}}\) , \(\:{\text{h}}_{\text{n}\_1}^{\text{a}\text{u}\text{x}}\) , \(\:{\text{h}}_{\text{n}\_2}^{\text{a}\text{u}\text{x}}\) to vectors in the latent space \(\:{\text{z}}_{\text{n}\_1}^{\text{i}\text{n}\text{h}}\) , \(\:{\text{z}}_{\text{n}\_2}^{\text{i}\text{n}\text{h}}\) , \(\:{\text{z}}_{\text{n}\_1}^{\text{a}\text{u}\text{x}}\) , \(\:{\text{z}}_{\text{n}\_2}^{\text{a}\text{u}\text{x}}\) , respectively. Next, the two latent vectors are combined to generate a total feature. The representation of the total feature can be considered as a fusion of intrinsic and auxiliary features. The goal of this mapping is to further enhance the consistency of positive sample pairs and the difference of negative sample pairs in the representation space during contrastive learning. $$\:{\text{z}}_{\text{n}\_1}^{\text{t}\text{o}\text{t}}={\text{z}}_{\text{n}\_1}^{\text{i}\text{n}\text{h}}+{\text{z}}_{\text{n}\_1}^{\text{a}\text{u}\text{x}}$$ 8 $$\:{\text{z}}_{\text{n}\_2}^{\text{t}\text{o}\text{t}}={\text{z}}_{\text{n}\_2}^{\text{i}\text{n}\text{h}}+{\text{z}}_{\text{n}\_2}^{\text{a}\text{u}\text{x}}$$ 9 To achieve this goal, this work use the Normalized Temperature Cross-Entropy (NT-Xent) loss function. This loss function is specifically designed for contrastive learning. For positive sample pairs, this work aim to maximize the similarity between the shared features and the total features, while minimizing the similarity of the private features. For negative sample pairs, this work aim to reduce the similarity between the shared features and the total features, while enhancing the difference of the private features. After the contrastive learning pretraining is completed, as shown in Fig. 4 , this work further fine-tune the pretrained model for molecular property prediction tasks. In this stage, the structure of the prediction model is similar to that of the pretrained model, consisting of a GNN backbone and an MLP head. The GNN backbone uses the same architecture as the pretrained feature extractor, while the MLP head is used to map the features extracted by the GNN to the predicted molecular properties. During the fine-tuning process, the GNN backbone is initialized by sharing parameters from the pretrained model, leveraging the rich structural information learned during pretraining. The MLP head is randomly initialized, allowing it to adapt more flexibly to different molecular property prediction tasks. Subsequently, this work train the entire fine-tuned model on the target molecular property dataset using supervised learning. In this way, the model can be fine-tuned on new molecular data for specific tasks, further improving prediction performance. 3 Experimental evaluation with molecular property In the field of molecular property prediction, classification tasks play a crucial role. Different datasets are designed around specific biological or pharmacological questions. For example, BBBP, BACE, and HIV focus on single-label classification problems, such as determining whether a molecule can cross the blood-brain barrier, inhibit BACE-1, or exhibit anti-HIV activity. In contrast, Clintox, Tox21, Sider, and MUV are multilabel classification datasets, requiring the simultaneous prediction of molecular behavior across multiple toxicity pathways, side effects, or activity targets. These benchmark datasets provide a rich foundation for developing and evaluating molecular classification models, driving advances in drug discovery and toxicity risk assessment. 3.1 Dataset and Preprocession This section provides implementation details. To demonstrate the effectiveness of the approach across various datasets, this work conducted experiments on seven molecular datasets from MoleculeNet [ 5 ], including BACE and BBBP for single-task classification and MUV, HIV, SIDER, ClinTox, and Tox21 for multi-task classification. BBBP (Blood-Brain Barrier Penetration) categorizes compounds based on their ability to penetrate or fail to penetrate the blood-brain barrier (BBBP). The BBBP is critical in drug development as it determines whether a drug can access the central nervous system. BACE (Beta-secretase 1) is designed to predict the inhibitory activity of compounds on the enzyme beta-secretase 1 (BACE-1), a key target in Alzheimer’s drug development. HIV predicts whether a compound has the ability to inhibit HIV virus replication. It contains anti-HIV screening results published by the Drug Therapeutics Program (DTP). ClinTox (Clinical Toxicity) predicts whether a compound is safe or toxic based on FDA-approved drug safety data. It includes labeled information on toxic and safe compounds. Tox21 predicts 12 types of toxicity associated with compounds. Tox21 aims to assess the potential impact of drugs and chemicals on human health. SIDER (Side Effect Resource) contains information on the side effects of FDA-approved drugs, categorized into 27 common side effect types. MUV (Maximum Unbiased Validation) is a high-quality dataset used for benchmarking virtual screening. It includes activity and inactivity data for compounds targeting 17 biological targets, designed to validate virtual screening models. Table 1 Dataset Summary Dataset Graphs Tasks Metric Split Classification task type bbbp 2039 1 ROC-AUC Scaffold Single-label classification bace 1513 1 ROC-AUC Scaffold Single-label classification hiv 41127 1 ROC-AUC Random Single-label classification clintox 1478 2 ROC-AUC Random Multi-label classification tox21 7831 12 ROC-AUC Random Multi-label classificatio sider 1427 27 ROC-AUC Random Multi-label classificatio 3.2 Evaluation metrics ROC-AUC (Receiver Operating Characteristic - Area Under the Curve) is a widely used metric for evaluating the performance of classification models, particularly for measuring their ability to distinguish between positive and negative samples. It combines the ROC curve and the AUC value: 1.The ROC curve illustrates the model’s performance at different threshold values. It is a plot where the False Positive Rate (FPR) is on the x-axis and the True Positive Rate (TPR) is on the y-axis. True Positive Rate (TPR) is proportion of actual positive samples correctly predicted as positive. It is calculated as: $$\:TPR=\:\frac{TP}{TP+FN}$$ 10 False Positive Rate (FPR) is proportion of samples predicted as positive that are actually negative. It is calculated as: $$\:FPR=\frac{FP}{FP+TN}$$ 11 2. AUC (Area Under the Curve) is the area under the ROC curve. Its value ranges from 0 to 1, with values closer to 1 indicating a stronger ability to distinguish between classes. The specific meaning of AUC is the probability that a randomly chosen positive sample has a higher predicted probability than a randomly chosen negative sample. AUC = 1: The model's predictions are perfect. AUC = 0.5: The model's predictions are no better than random guessing. AUC < 0.5: The model's predictions are worse than random guessing, indicating that there might be an issue with the model. 3.3 Loss Function SimCLR demonstrated that contrastive learning for images can greatly benefit from the combination of data augmentation and large batch sizes. Based on the InfoNCE loss, SimCLR introduced the NT-Xent loss. In this study, for each feature type (public, private, and total features), this work calculate the similarity of the augmented samples for each instance, denoted as \(\:{\text{s}\text{i}\text{m}}^{\text{i}\text{n}\text{h}}({\text{z}}_{\text{i}}^{\text{i}\text{n}\text{h}},{\text{z}}_{\text{j}}^{\text{i}\text{n}\text{h}})\) and \(\:{\text{s}\text{i}\text{m}}^{\text{t}\text{o}\text{t}}({\text{z}}_{\text{i}}^{\text{t}\text{o}\text{t}},{\text{z}}_{\text{j}}^{\text{t}\text{o}\text{t}})\) , where \(\:{\text{z}}_{\text{i}}\) and \(\:{\text{z}}_{\text{j}}\) are the latent vectors extracted from positive data pairs, N is the batch size, sim(·) measures the similarity between two vectors, and τ is the temperature parameter. In DCFS-CL, this work follow the NT-Xent loss to pre-train the GNN encoder, and the cosine similarity is implemented as: $$\:sim({z}_{i},{z}_{j})=\frac{{z}_{i}^{T}{z}_{j}}{{\left|\right|{z}_{i}\left|\right|}_{2}{\left|\right|{z}_{j}\left|\right|}_{2}}$$ 12 The loss for intrinsic features \(\:{\text{L}}_{\text{i}}\) : $$\:{L}_{i}=-log\frac{exp\left({sim}^{inh}\right({z}_{i}^{inh},{z}_{j}^{inh})/{\tau\:}^{inh})}{{\sum\:}_{k=1}^{2N}{1}_{i\ne\:k}exp\left({sim}^{inh}\right({z}_{i}^{inh},{z}_{k}^{inh})/{\tau\:}^{inh})}$$ 13 The loss for auxiliary features \(\:{L}_{a}\) : $$\:{L}_{a}=-log\frac{exp\left({sim}^{aux}\right({z}_{i}^{aux},{z}_{j}^{aux})/{\tau\:}^{aux})}{{\sum\:}_{k=1}^{2N}{1}_{i\ne\:k}exp\left({sim}^{aux}\right({z}_{i}^{aux},{z}_{k}^{aux})/{\tau\:}^{aux})}\:$$ 14 The loss for auxiliary features \(\:{L}_{t}\) : $$\:{L}_{t}=-log\frac{exp\left({sim}^{tot}\right({z}_{i}^{tot},{z}_{j}^{tot})/{\tau\:}^{tot}}{{\sum\:}_{k=1}^{2N}{1}_{i\ne\:k}exp\left({sim}^{tot}\right({z}_{i}^{tot},{z}_{k}^{tot})/{\tau\:}^{tot})}$$ 15 The final loss combines the three types of feature losses with weighting coefficients a and b : $$\:L={a\ast\:L}_{i}+b\ast\:{L}_{a}+{L}_{t}$$ 16 3.4 Results and discussion Table 2 Performance of different models tested on seven classification benchmarks. Methods Classification ROC-AUC bbbp bace hiv clintox tox21 sider muv GCN[ 14 ] 71.8(0.9) 71.6(2.0) 74.0(3.0) 62.5(2.8) 70.9(2.6) 53.6(3.2) 71.6(4.0) GIN[ 15 ] 65.8(4.5) 70.1(5.4) 75.3(1.9) 58.0(4.4) 74.0(0.8) 57.3(1.6) 71.8(2.5) SVM[ 16 ] 72.9(0.0) 86.2(0.0) 79.2(0.0) 66.9(9.2) 81.8(1.0) 68.2(1.3) 67.3(1.3) RF[ 17 ] 71.4(0.1) 86.7(0.8) 78.1(0.6) 71.3(5.6) 76.9(1.5) 68.4(0.9) 63.2(2.3) JOAO[ 18 ] 71.4(0.0) 68.0 (1.6) 76.6 (1.0) 78.1 (3.9) 74.7 (0.2) 60.6 (1.1) 74.2 (1.6) JOAOv2[ 19 ] 68.0 (1.0) 75.3 (3.7) 76.9 (0.8) 78.6 (3.3) 74.9 (0.7) 61.1 (0.7) 75.3 (1.2) MGSSL[ 3 ] 70.5 (1.1) 79.7 ( 0.8) 79.5 (1.1) 80.7 (2.1) 76.4 (0.4) 61.8 (0.8) 78.7 (1.5) GraphCL[ 1 ] 68.9(1.3) 74.4 (2.5) 77.3 (1.1) 70.1 (7.8) 74.2 (0.5) 59.5 (0.6) 71.9 (2.7) GraphMAE[ 2 ] 69.0 (2.0) 81.3 (1.4) 76.3 (1.1) 73.1 (5.0) 73.6 (2.2) 59.7 (0.8) 76.2 (1.8) GraphMVP[ 11 ] 67.9 (2.0) 78.6 (3.5) 76.1 (1.7) 76.1 (4.8) 75.4 (0.5) 59.3 (1.1) 74.4 (2.2) MoleMCL[ 7 ] 67.8 (3.5) 80.3 (0.8) 79.6 (0.5) 80.2 (1.6) 76.2 (0.6) 59.0 (0.8) 78.2 (1.0) MolCLR[ 4 ] 73.3(1.0) 82.8(0.7) 77.4(0.6) 89.8(2.7) 74.1(5.3) 61.2(3.6) 78.9(2.3) DCFS-CL 76.7(0.9) 81.5(0.5) 80.8(0.4) 91.4(1.6) 75.3(0.7) 65.9(1.3) 81.3(1.7) As shown in Table 2 . The first four models are supervised learning methods, while the latter eight are self-supervised/pre-trained methods. The table reports the average and standard deviation of the test ROC-AUC (%) for each benchmark. This work proposed a self-supervised learning framework based on a dual-channel feature separation network and contrastive learning for molecular property prediction. To validate the effectiveness of the model, this work conducted extensive experiments on 661 classification tasks across seven biological datasets and compared the results with existing supervised learning methods and self-supervised learning approaches. Table 2 presented the average ROC-AUC (%) values and standard deviations of different models tested on these datasets. As shown in Table 2 , DCFS-CL achieved an average ROC-AUC score of 81.9% across all seven classification datasets, which was significantly higher than that of other supervised learning and self-supervised/pretrained methods. In particular, DCFS-CL achieved a ROC-AUC score of 91.4% on the ClinTox dataset, surpassing the next-best method (MolCLR) by 1.6%. Additionally, DCFS-CL achieved a ROC-AUC score of 81.3% on the challenging dataset known for its imbalanced samples (MUV), outperforming MGSSL (78.7%) and other models. These above results indicated that DCFS-CL not only performed exceptionally well on simpler tasks but also demonstrated strong robustness and generalization capability on complex and imbalanced datasets. The first four methods in Table 2 (GCN, GIN, SVM, RF) represented traditional supervised learning approaches. While they performed well on certain datasets, their overall performance was relatively limited. For instance, although GIN and GCN could capture the local structural information in molecular graphs, they exhibited poor performance on complex tasks (such as ClinTox and MUV). Especially, they were prone to the over-smoothing problem in the large-scale data. As traditional machine learning methods, SVM and RF relied on manual feature extraction and excel in certain specific tasks, such as BACE and HIV. However, they lacked a deep understanding of molecular graph structures, resulting in insufficient generalization ability on other tasks. In contrast, DCFS-CL overcame the limitations of traditional methods by integrating a dualchannel feature separation network with contrastive learning, achieving consistently high performance across multiple tasks. The last eight methods in Table 2 (e.g., JOAO, GraphCL, MolCLR, etc.) were popular self-supervised learning or pre-training approaches, whose initialization representation capacity were enhanced through an unsupervised learning phase. However, these methods still exhibited certain performance limitations. For instance, JOAO and its improved version, JOAOv2, performed relatively well on some datasets (such as MUV), but they demonstrated weak generalization abilities, particularly under complex tasks (such as ClinTox, SIDER). GraphCL and GraphMAE employed graph augmentation techniques for contrastive learning, but their performance was limited by the choice of augmentation strategies. These strategies could lead to the loss of critical chemical information on molecule. MolCLR, as a state-of-the-art contrastive learning-based method, performed notably well on ClinTox dataset (89.8%), but its performance tended to be relatively unstable on other tasks, such as SIDER and MUV. DCFS-CL effectively addressed the limitations of the aforementioned methods by introducing graph augmentation strategies that align more closely with chemical logic and a dual-channel feature extraction mechanism. As a result, it demonstrated good generalization ability across multiple datasets. 3.4 Ablation experiment DCFS-CL was developed based on the MolCLR framework. The primary improvements in DCFS-CL were the graph augmentation strategy, which involves modifying the scaffold structure and functional group placement, and the dual-channel feature separation network, which separately extracts inherent and auxiliary features. This work denoted the operation of using new graph augmentation strategies as R, and the operation of employing the dual-channel feature separation network as G. To verify the effectiveness of the components in DCFS-CL, This work conducted an ablation study based on the MolCLR model. Table 3 Ablation Study Results Methods Classification ROC-AUC bbbp bace hiv clintox tox21 sider muv MolCLR 73.3(1.0) 82.8(0.7) 77.4(0.6) 89.8(2.7) 74.1(5.3) 61.2(3.6) 78.9(2.3) MolCLR + R 74.4(0.3) 81.3(1.1) 78.6(0.4) 90.6(1.3) 74.8(0.9) 63.5(2.3) 80.8(1.5) MolCLR + G 73.5(0.6) 80.4(0.3) 78.4(0.9) 90.1(1.2) 75.1(0.4) 63.1(1.8) 79.6(0.9) DCFS-CL 76.7(0.9) 81.5(0.5) 80.8(0.4) 91.4(1.6) 75.3(0.7) 65.9(1.3) 81.3(1.7) This work designed multiple ablation experiments to evaluate the performance of each modification. The ROC-AUC values for the HIV dataset prediction are 77.4, 78.6, and 78.4 for the MolCLR, MolCLR + R, and MolCLR + G algorithms, respectively. These results indicate that both the inclusion of the graph augmentation strategy and the dual-channel feature separation network individually contribute to improving the model’s accuracy. In comparation with MolCLR, the average ROC-AUC of MolCLR + R on classification tasks increased from 74.1% to 75.8%, demonstrating an improvement of approximately 1.9%. The improved ROC-AUC was attributed to the property-preserving molecular graph augmentation strategy in MolCLR + R. This strategy focused on preserving the key structural information of molecules by adjusting the positions of nodes and edges rather than simply removing them, while fine-tuning the non-essential parts. The augmented molecular graphs retained the chemical properties of the original molecules and the data diversity was enhanced by this strategy. Consequently, the unlabeled data was effectively used in MolCLR + R, thereby reducing reliance on labeled data. Furthermore, the average ROC-AUC score of MolCLR + G on classification tasks increased from 74.1% (MolCLR) to 75.5%. This improvement was attributed to the dual-channel feature separation network. This network could extract both inherent and auxiliary features of molecules and decouple these features to avoid the mixing of feature types. This decoupling allowed for better control over the feature extraction process, enabling the model to focus on the correct feature type when handling molecular structural changes. Additionally, the semantic changes in molecular graphs could be sensitively captured by the dedicated auxiliary feature extraction network. As a result, it was found that both the graph augmentation strategy (R) and the dual-channel feature separation network (G) make significant contributions to the model’s performance, with neither being dispensable. This work complete model outperformed MolCLR across all datasets, significantly enhancing the model’s interpretability and task adaptability. 4 experimental evaluation with performance prediction of flow batteries In this work, the Single Point HOMO Solution denotes the highest occupied molecular orbital (HOMO) energy calculated from a fixed molecular geometry using quantum chemical methods. This approach does not involve any additional geometry optimization steps, thereby enabling efficient and rapid estimation of electronic properties. The HOMO energy is directly related to the molecule’s oxidation potential and overall electrochemical stability, making it a key descriptor in the screening of redox-active materials for flow batteries. Accurate prediction of the Single Point HOMO Solution allows for high-throughput computational evaluation of candidate molecules, significantly accelerating the material discovery process. 4.1 Dataset and its processing RedDB is a computational database specifically designed for aqueous redox flow batteries (ARFBs), encompassing 31,618 organic electroactive molecules. The database primarily includes two prominent classes of compounds: quinones and aza-aromatics.This work tested single point homo solution. Table 4 RedDB Dataset Summary Dataset Graphs Tasks Metric Split RedDB 31618 1 MAE, RMSE Random 4.2 Evaluation Metrics 1.Mean Absolute Error (MAE) is a widely used metric in statistics and machine learning for evaluating the accuracy of predictive models, particularly in regression tasks. It quantifies the average magnitude of errors between predicted and actual values, without considering their direction. MAE is calculated as the average of the absolute differences between predicted values \(\:{\text{y}}_{\text{i}}\) and \(\:{\text{y}}_{\text{i}}^{{\prime\:}}\) actual values across all observations: $$\:MAE=\frac{1}{n}{\sum\:}_{i=1}^{n}|{y}_{i}-{y}_{i}^{{\prime\:}}|$$ 17 2. Root Mean Square Error (RMSE) is a widely used metric to evaluate the performance of regression models. It measures the average magnitude of the errors between predicted values and actual values. The RMSE is calculated as follows: $$\:RMSE=\sqrt{\frac{1}{n}{\sum\:}_{i=1}^{n}|{y}_{i}-{y}_{i}^{{\prime\:}}|\:}$$ 18 4.3 Experimental Results Table 5 Model Performance Comparison Using the RedDB Dataset for single_point_homo_solution model MAE RMSE MPNN[ 20 ] 0.083 0.117 GCN 0.075 0.101 GAT[ 21 ] 0.061 0.086 AttentiveFP[ 22 ] 0.060 0.086 DCFS-CL 0.037 0.044 The results shows the MAE and RMSE for different models, namely MPNN, GCN, GAT, AttentiveFP, and DCFS-CL. It is evident that DCFS-CL has lower MAE and RMSE values, indicating its superiority over all of the other models. Therefore, the results indicate that DCFS-CL is an effective tool for predicting electrochemical properties. As shown in Fig. 6 , the prediction performance of DCFS-CL is relatively accurate, which may be attributed to the proposed method's ability to leverage large-scale unlabeled data for pretraining in advance. Moreover, the graph augmentation strategy and dual-channel feature separation network employed by DCFS-CL demonstrate significant advantages over other baseline models in terms of generalization across chemical space and various molecular properties. 5 Conclusions This paper propose DCFS-CL(Dual-Channel Feature Separation with Contrastive Learningfor molecular property prediction based on feature decoupling and molecular contrastive learning. This work introduce a novel graph augmentation strategy and a dual-path feature separation network to extract inherent and auxiliary features of molecules separately. Experimental results demonstrate that DCFS-CL achieves significant improvements across multiple molecular benchmark datasets, even outperforming some traditional supervised learning methods. This framework not only enhances the accuracy of molecular representations but also improves the robustness of the model, offering new insights and approaches for molecular property prediction tasks. To summarize, this work introduce an innovative approach to molecular property prediction by integrating two key strategies: (1) This work combine scaffold augmentation with functional group augmentation and apply them to self-supervised learning. This combination allows for flexible manipulation of molecular graphs, ensuring that the essential scaffold structure remains intact while simultaneously enhancing the functional groups to generate diverse molecular representations. This strategy not only preserves the fundamental molecular features but also introduces variability, which is crucial for improving model generalization and performance. (2) This work propose a dual-channel feature separation network that separates inherent features, such as atomic types and bond types, from auxiliary features, such as atomic chirality and bond stereochemistry. This separation prevents interference between different types of features, allowing for more focused and effective feature extraction. By decoupling these features, the approach enhances the model’s ability to accurately capture the nuances of molecular structures, leading to improved predictions of molecular properties. Together, these strategies contribute to a more robust and flexible model that can be applied to various molecular property prediction tasks, offering new insights into molecular graph learning. Furthermore, in future work, this work incorporating the domain knowledge embedded in large models can further improve the accuracy of molecular property prediction. Additionally, this work plans to apply DCFS-CL to the field of electrochemistry. Specifically, the model can be used to predict the properties of energy storage and conversion materials, such as batteries, supercapacitors, and fuel cells. By leveraging the molecular-level insights provided by framework, this work expects to make progress in material discovery, electrochemical performance optimization, and the development of more efficient and sustainable energy systems. Future research can focus on integrating electrochemical-specific data and domain knowledge into the model to enhance its predictive capabilities and accelerate the advancement of electrochemical technologies. Declarations Competing financial interests The authors declare no competing financial interests. Data availibility The molecular SMILES strings and corresponding datasets are available https://moleculenet.org/ datasets-1. Funding This work was supported in part by the National Science and Technology Major Project of China under grant 2022ZD0119501, the National Natural Science Foundation of China under grant 52374221, the Natural Science Foundation of Shandong Province of China under grant ZR2023MF097 and ZR2024QF107, the Major Basic Research Project of the Shandong Provincial Natural Science Foundation under grant ZR2024ZD22.Qingdao West Coast New Area Science and Technology Project (Science and Technology Tackling Key Problems with Open Bidding) under grant 2022-5. Author Contribution Faming Lu proposed the core methodology and technical pipeline, conducted the main experimental design, model development, and result analysis, and drafted the initial manuscript. Yishan Zhu contributed to model implementation, experimental validation, and data analysis, and participated in revising and improving the manuscript. Zedong Lin provided overall supervision, participated in research discussions, and offered critical academic insights. Yunxia Bao, as one of the corresponding authors, guided the research direction, supervised the experimental design, and provided important revisions to the manuscript structure and academic presentation. Wanpeng Huang contributed to the analysis and discussion of experimental results and assisted in manuscript polishing and revision. All authors reviewed and approved the final manuscript. Data Availability The molecular SMILES strings and corresponding datasets are available https:// moleculenet. org/ datasets-1. References You Y, Chen T, Sui Y et al. Graph Contrastive Learning with Augmentations[J]. arXiv preprint arXiv:2006.08242, 2020. Hou Z, Wu Q, Zhang Z et al. GraphMAE: Self-Supervised Masked Graph Autoencoders[J]. arXiv preprint arXiv:2205.10803, 2022. Zhang S, Jiang D, Gao S et al. MGSSL: Self-Supervised Learning of Molecular Graphs with Domain Knowledge Enhanced Motif-Based Augmentation[J]. arXiv preprint arXiv:2106.13049, 2021. Wang Y, Wang J, Cao Z, Farimani AB. Molecular contrastive learning of representations via graph neural networks. Nat Mach Intell. 2022;4:279–87. Wu Z, et al. Moleculenet: a benchmark for molecular machine learning. Chem Sci. 2018;9:513–30. Hu W, Liu Y, Yang Z. (2020). MolPre: Pretraining Molecular Graph Representations for Predicting Molecular Properties. arXiv preprint arXiv:2009.10013. Zhang X, Xu Y, Jiang C et al. MoleMCL: a multi-level contrastive learning framework for molecular pre-training.[J]. Bioinformatics (Oxford, England),2024. BROUWERS E, VAN DER SCHAAF H. Improving Graph Attention Networks with Multi-hop Attention Aggregation[J]. IEEE Trans Neural Networks Learn Syst. 2021;32(5):2046–55. XU Z, XIE W. Temporal GraphSAGE: A Graph Neural Network for Dynamic Graphs[J]. IEEE Trans Neural Networks Learn Syst. 2022;33(9):4255–66. MA T, YOU Z, LIPTON Z, et al. Multi-View Graph Autoencoders for Multimodal Integration and Retrieval[J]. IEEE Trans Pattern Anal Mach Intell. 2021;43(7):2226–37. Li Chenghong Z, Xiaofei. Dual-channel graph random convolutional network for semi-supervised node classification[J]. J Chin Comput Syst 2023,44(08):1656–64. 10.20009/j.cnki.21-1106/TP.2021-0890 VELIČKOVIĆ P, FEDUS W, HAMILTON W, L et al. Deep Graph Infomax[C]//International Conference on Learning Representations (ICLR). New Orleans, LA: ICLR, 2019. Jiang T, Wang Z, ,Yu W et al. Mix-Key: graph mixup with key structures for molecular property prediction. [J] Briefings Bioinf, 2024, 25 (3). Kipf NT, Welling M. Semi-Supervised Classification with Graph Convolutional Networks.[J]. CoRR,2016,abs/1609.02907. Xu K, Hu W, Leskovec J, Jegelka S. How powerful are graph neural networks? In International Conference onLearning Representations (2019). CORTES C. Support-vector networks[J]. Mach Learn. 1995;20(3):273–97. BREIMAN L. Random forests[J]. Mach Learn. 2001;45(1):5–32. LI C, WANG H, LI K et al. Joint optimized adversarial objective for graph representation learning[J]. Proceedings of the International Conference on Learning Representations (ICLR), 2020. LI C, WANG H, ZHU W et al. JOAOv2: Joint optimized adversarial objective for graph representation learning with improved performance[J]. Proceedings of the 37th International Conference on Machine Learning (ICML), 2020. Gilmer J, Schoenholz SS, Riley PF, Vinyals O, Dahl G. E.Neural message passing for quantum chemistry. In Proc. 34th International Conference on Machine Learning 70 (eds Doina, P. &Teh, Y. W.) 1263–1272 (PMLR, 2017). VELIČKOVIĆ P, CUCURULL G, CASANOVA A et al. Graph attention networks[J]. Proceedings of the International Conference on Learning Representations (ICLR), 2018. LIU S, WANG H, LIU W et al. AttentiveFP: attentive message passing for molecular property prediction[J]. Proceedings of the International Conference on Learning Representations (ICLR), 2020. Ocaña AP, Rodriguez M J,Ocaña MMJ et al. An AI-Powered Methodology for Atomic-Scale Analysis of Heterogenized Correlated Single-Atom Catalysts.[J]. Small methods,2025. Jang YS, Jang UJ, Yoo YG et al. Nano-Interconnected 1D/2D Boron Nitride Hybrid Networks: Unlocking Superior Thermal Conductivity in Electrically Insulating Thermal Interface Nanocomposites Based on Hybrid Thermal Percolation Model.[J]. Small methods,2025. Wen J, Ji Y, Hu Y et al. Electrolyzing Outside the Box: Non-Traditional Approaches to Electrochemical Water Splitting.[J]. Small methods,2025. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8846431","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":589961431,"identity":"866fa4a1-f0a7-4cb1-a55b-5196d54c3656","order_by":0,"name":"Yishan Zhu","email":"","orcid":"","institution":"Shandong University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Yishan","middleName":"","lastName":"Zhu","suffix":""},{"id":589961432,"identity":"ffadac16-82ca-4a07-bacb-3e556da5ebcd","order_by":1,"name":"Qian Zhou","email":"","orcid":"","institution":"Qingdao Industrial Energy Storage Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Qian","middleName":"","lastName":"Zhou","suffix":""},{"id":589961433,"identity":"eaa75b20-44cf-4aa8-976c-5dc98e5494c6","order_by":2,"name":"Faming Lu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAmElEQVRIiWNgGAWjYBAC9gYGxgMMFRJy8kRr4TnAwHCA4YyFsWEDSVoY2yoSgRqJ1SJ2gOEw7zyJBMYG5oePbhClRToBqGWbRB47A5uxcQ4xWuyhWooZG3jYpInSArFljkRiwwHStDSQquXgnGMSxobNxPoFqIXxwZuaOjl59uaHj4nSwsDA/wFCMxOnfBSMglEwCkYBMQAAa+kqqqhhPfQAAAAASUVORK5CYII=","orcid":"","institution":"Shandong University of Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Faming","middleName":"","lastName":"Lu","suffix":""},{"id":589961434,"identity":"98d0dbb6-ffb2-42bc-b27b-09b213102561","order_by":3,"name":"Guanglei Cui","email":"","orcid":"","institution":"Qingdao Industrial Energy Storage Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Guanglei","middleName":"","lastName":"Cui","suffix":""},{"id":589961435,"identity":"9814d954-27a2-4a87-b8bb-71e3f94c4d9e","order_by":4,"name":"Cong Liu","email":"","orcid":"","institution":"NOVA Information Management School","correspondingAuthor":false,"prefix":"","firstName":"Cong","middleName":"","lastName":"Liu","suffix":""}],"badges":[],"createdAt":"2026-02-11 02:54:05","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8846431/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8846431/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103035703,"identity":"fb0ea9de-0ed0-4081-bd16-39e69ddc4fb4","added_by":"auto","created_at":"2026-02-20 01:33:42","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":77871,"visible":true,"origin":"","legend":"\u003cp\u003eThe main architecture of GroupAug.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8846431/v1/db5517ec046308f7748d1bad.png"},{"id":103050488,"identity":"1935144e-c8b4-4d44-a28e-c15a7998cb8f","added_by":"auto","created_at":"2026-02-20 07:50:16","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":169349,"visible":true,"origin":"","legend":"\u003cp\u003eMain architecture of ScaffoldAug\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8846431/v1/a6d07eb54a89ac8ac436607c.png"},{"id":103049821,"identity":"246d1ca2-52bb-4dd2-a4ec-e5dff62c87d8","added_by":"auto","created_at":"2026-02-20 07:46:36","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":155622,"visible":true,"origin":"","legend":"\u003cp\u003eA Molecular Contrastive Learning Pretraining Framework Based on Dual-Channel Feature Separation Network Representation\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8846431/v1/69aed9930158c208f8ae3641.png"},{"id":103035699,"identity":"a806fb17-7353-4018-b019-0ba45717e94c","added_by":"auto","created_at":"2026-02-20 01:33:42","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":121552,"visible":true,"origin":"","legend":"\u003cp\u003eThe entire molecular contrastive learning framework based on the dual-channel feature separation network\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8846431/v1/a6c9420dfad7f25e733b7e33.png"},{"id":103050107,"identity":"147d05de-4fe2-4401-bd41-dcba5d7131bf","added_by":"auto","created_at":"2026-02-20 07:48:16","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":32672,"visible":true,"origin":"","legend":"\u003cp\u003eAblation experiment results on each dataset\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8846431/v1/b1ddabdfa7f5b26d4df4adde.png"},{"id":103035701,"identity":"b8419d86-508e-4af2-86b3-9e635cba7b72","added_by":"auto","created_at":"2026-02-20 01:33:42","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":58414,"visible":true,"origin":"","legend":"\u003cp\u003eScatter Plot of Predicted vs. True Values for Single Point HOMO Solution\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8846431/v1/058b7ba7d1dcebf73a79bb41.png"},{"id":103342286,"identity":"4c2fce80-a0bb-496c-a396-0036205b786f","added_by":"auto","created_at":"2026-02-24 15:41:37","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1630474,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8846431/v1/e58d7572-150a-44b0-bddd-9d8d0e6e1658.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Molecular Property Prediction Based on Dual-Channel Feature Separation Network and Contrastive Learning","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eMaterials science, electrochemistry and nanotechnology are important fields in modern research[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].Molecular property prediction is a core task in modern chemistry, materials science, and drug development. Accurate prediction models typically require large amounts of labeled data to achieve satisfactory performance. However, in many real-world scenarios, obtaining large-scale annotated datasets for molecular properties is infeasible\u0026mdash;especially for electrochemical properties, which often rely on complex experiments or high-precision quantum chemical calculations, making data acquisition costly. Moreover, some of these properties are inherently rare. Self-Supervised Learning (SSL) can automatically learn useful feature representations from a large volume of unlabeled data by designing appropriate pretext tasks. It has attracted widespread attention in molecular property prediction and has demonstrated great potential in electrochemical molecular modeling tasks.\u003c/p\u003e \u003cp\u003eContrastive learning, as a typical self-supervised learning method, is a widely adopted technique for chemistry. The core idea of contrastive learning is to construct similar pairs and learn feature representations of data by maximizing the consistency within these similar pairs. In molecular representation tasks, current contrastive learning methods generally rely on graph augmentation techniques to generate similar pairs, including node deletion, edge perturbation, and subgraph extraction [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e], whose goal is to generate multiple views that guide the model to learn more robust feature representations. Hu et al. introduced GraphCL[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] by generating multiple views (e.g., node dropping, edge perturbation) to strengthen the robustness of molecular graph representations. GraphMAE employs a masking strategy (i.e., randomly masking node or edge information) to enable the model to learn both global and local structural features of molecular graphs, enhancing its self-supervised representation learning ability [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. However, such graph augmentation strategies can easily lead to information loss, introduce noise or bias, neglect local-global relationships within molecules, and demonstrate insufficient adaptability to downstream tasks. For example, node deletion and edge removal strategies may disrupt important chemical groups or functional groups in the molecule, causing the model to lose critical semantic information. Specifically, in chemical property prediction tasks, the loss of certain atoms or chemical bonds can significantly affect the molecular properties.\u003c/p\u003e \u003cp\u003eMoreover, classic methods for molecular graph contrastive learning often extract all features within the same network [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. DGI (Deep Graph Infomax) proposed learning global graph representations by maximizing the mutual information between local and global feature[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. MGSSL (Multi-Granularity Self-Supervised Learning) uses a multi-granularity graph augmentation strategy to capture multi-scale features of molecular graphs through global and local view contrasts, thereby improving the accuracy of molecular property prediction [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. GraphMVP proposes a multi-view pretraining framework, combining the 2D topology of molecular graphs and 3D molecular conformations, capturing the physicochemical properties of molecules and performing excellently in molecular property prediction [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].MoleMCL captures key chemical features of molecules by comparing augmented views at different structural levels, such as scaffolds and functional groups, thereby improving molecular property prediction performance [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. In molecular graphs, nodes, edges, and global structures contain rich information, such as atom types, bond types, and spatial conformations. This information is progressively compressed into a low-dimensional embedding vector, when the features are extracted by neural networks. When all features of molecular graphs are generally extracted by one single neural network model, it may result in the loss of certain details due to information compression. For example, atomic information and stereoconformational information could be mixed into the same embedding vector, leading to the dilution of key features. For complex molecular property prediction tasks, this information loss can hinder the model's ability to capture essential chemical characteristics. In addition, different molecular property prediction tasks exhibit varying levels of dependence on atomic information and spatial conformations data. For instance, toxicity prediction of a molecule places greater emphasis on the local chemical environment, whereas solubility prediction may focus more on the overall shape of the molecule. If the single model fails to adapt to the task requirements, it may result in suboptimal performance.\u003c/p\u003e \u003cp\u003eJiang et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] designed a property-preserving molecular graph augmentation method, which does not remove edges or nodes from the molecular graph but instead changes their positions to generate augmented graphs, thereby preserving the fundamental features. This method converts the original molecular graph into two isomers, selectively retaining the molecular scaffold or functional groups, emphasizing the preservation of key structures that significantly influence the chemical properties of the molecule. By adjusting nonessential structures, the method generates diverse augmented samples while minimizing disruption to the original molecular semantics. However, previous methods only used scaffold augmentation or functional group augmentation individually for transformations and applied them in supervised learning tasks.\u003c/p\u003e \u003cp\u003eIn this study, this work present DCFS-CL(Dual-Channel Feature Separation with Contrastive Learning is proposed for molecular property prediction. Firstly, this work combine scaffold augmentation and functional group augmentation to flexibly adjust the layout of functional groups while ensuring the stability of the key scaffold structure, generating diverse molecular graphs without losing critical information. Then, a dual-channel feature separation network is proposed to extract inherent and auxiliary features, which can decouple these features and avoid interference caused by the mixing of feature types. This decoupling allows for better control over the feature extraction process, enabling the model to focus on the correct feature type when handling molecular structural changes. Additionally, semantic changes in molecular graphs often affect the feature extraction performance of traditional GNNs. By using feature extraction networks, these semantic changes can be more sensitively captured. This improves the model\u0026rsquo;s sensitivity to small molecular changes, thereby more accurately describing the differences between molecules. As a result, DCFS-CL demonstrates outstanding accuracy and robustness in regression tasks on an electrochemical dataset and classification tasks on seven biological datasets.In the following sections, this work will provide a detailed introduction to each aspect. Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e will focus on the graph augmentation strategy, contrastive learning with the dual-channel feature separation network, and the overall framework of the model. In Sections \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cspan refid=\"Sec11\" class=\"InternalRef\"\u003e4\u003c/span\u003e, this work will discuss the performance differences between DCFS-CL and other models on electrochemical and biological datasets, along with the underlying reasons. Finally, this work will conclude the paper and explore future directions for optimization.\u003c/p\u003e"},{"header":"2 Methods","content":"\u003cp\u003eThis section systematically introduces the molecular property prediction model framework and its theoretical support. First, this work analyze the accompanying features of molecular properties and describe what these features represent. Next, a molecular graph-preserving augmentation strategy is introduced to generate new data samples by enhancing the structure of molecular graphs with chemical plausibility. Finally, the molecular property prediction model framework is explained in detail, which is based on a dual-channel feature separation network and contrastive learning.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1Accompanying features of molecular properties\u003c/h2\u003e \u003cp\u003eIn molecular property prediction, the associated features in the molecular structure play a crucial role in describing molecular behavior, reaction mechanisms, and properties. These associated features include atomic type, bond type, atomic chirality, and bond stereochemistry.\u003c/p\u003e \u003cp\u003e \u003cb\u003eAtomic Type\u003c/b\u003e refers to the element type in molecule, such as carbon, hydrogen, oxygen, etc. As the basic units of functional groups, atomic type plays an important role in the chemical properties of molecule, such as polarity, reactivity, and stability. In a molecular graph, each node is typically considered an atom, whose type is used as an inherent property of the node in the model.\u003c/p\u003e \u003cp\u003e \u003cb\u003eBond Type\u003c/b\u003e describes the chemical bonds connecting different atoms, including single, double, and triple bonds. These bond types affect the molecule's geometric structure and energy state, significantly impacting on the molecule's reaction pathways and physicochemical properties. Bond types are typically input as edge properties in the molecular graph.\u003c/p\u003e \u003cp\u003e \u003cb\u003eAtomic Chirality\u003c/b\u003e refers to the spatial arrangement of atoms around a chiral center (such as R/S configurations), which leads to different stereoisomers of the molecule. These isomers may exhibit vastly different properties in biological activity and pharmacological effects. Atomic chirality is an auxiliary feature of the molecule, with high discriminative ability for certain molecular prediction tasks.\u003c/p\u003e \u003cp\u003e \u003cb\u003eBond Stereochemistry\u003c/b\u003e describes the spatial configuration of bonds (e.g., Z/E configurations), which significantly impacts on the three-dimensional structure and spatial arrangement of the molecule. Molecules often exhibit differences in solubility, polarity, and pharmacokinetics due to these features.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Molecular graph augmentation\u003c/h2\u003e \u003cp\u003eIn molecular representation tasks, the generation of positive pairs often relies on graph augmentation techniques. Typical augmentation methods include node deletion, edge perturbation, and subgraph extraction, aimed at generating multiple views to help the model learn robust feature representations. However, these augmentation methods may disrupt the overall structure of the molecule or important chemical information on the molecular graph. For example, deleting nodes or edges may destroy key functional groups in the molecule, causing the model to lose crucial semantic information. This is especially problematic in certain chemical property prediction tasks, where the loss of specific atoms or bonds can significantly impact the molecular performance.\u003c/p\u003e \u003cp\u003eThis work introduce a novel graph augmentation strategy. It does not remove elements but instead swaps the positions of nodes or edges, selectively preserving the molecular scaffold and functional groups. This approach emphasizes retaining key structures that have a significant impact on the chemical properties of the molecule.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFirst, this work extract the scaffold graph of the molecule and mask its edges. Since functional groups are more dispersed than the scaffold, this work do not construct a line graph here, but instead directly reconnect the edges of the original graph. This work randomly select an edge \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{1}\\)\u003c/span\u003e\u003c/span\u003e =(\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{1}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{2}\\)\u003c/span\u003e\u003c/span\u003e, ω)and disconnect it. Modifying edges in the molecular graph leads to changes in the molecular topology, which in turn affects the properties of the molecule, including the number of hydrogens and the number of charges. Therefore, here, this work update the number of hydrogens as follows:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{\\text{h}}_{\\text{i}}={\\text{h}}_{\\text{i}}+{\\omega\\:}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eIn the equation, ω represents the valency of the broken edge. Then, two randomly selected non-connected nodes are connected to form a new edge \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{1}^{{\\prime\\:}}\\)\u003c/span\u003e\u003c/span\u003e =(\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{3}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{4}\\)\u003c/span\u003e\u003c/span\u003e, ω). Similarly, the number of hydrogen atoms in these nodes should also be updated here.\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:{\\text{h}}_{\\text{j}}={\\text{h}}_{\\text{j}}-{\\omega\\:}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eIn addition, when \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{j}}\\)\u003c/span\u003e\u003c/span\u003e is less than 0, this work need to update the number of electrons.\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:{\\text{c}}_{\\text{j}}={\\text{c}}_{\\text{j}}+{\\text{h}}_{\\text{j}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eLet \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{j}}\\)\u003c/span\u003e\u003c/span\u003e=0.\u003c/p\u003e \u003cp\u003eAfter that apply ScaffoldAug. ScaffoldAug is a method that generates isomers by swapping edges in the molecular scaffold without changing the functional groups of the molecule. Functional groups are often structural units closely related to the chemical properties of the molecule. To reduce the search space of edge-swapping operations, this work convert node selection into edge selection. ScaffoldAug first extracts the scaffold graph from the molecule and converts it into a line graph. Then, two nodes are selected from the line graph and mapped to two edges in the scaffold graph. Finally, this work perform the edge-swapping operation and use chemical rules to filter out invalid isomers. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates the main architecture of ScaffoldAug, with the specific details of each generation step shown below.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e\u0026bull; Line graph construction\u003c/p\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, this work first describes how to construct the line graph of a molecular scaffold. Given a graph \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:G\\)\u003c/span\u003e\u003c/span\u003e, this work denote its line graph as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{L}_{G}\\)\u003c/span\u003e\u003c/span\u003e = (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{A}_{L}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{X}_{L}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{W}_{L}\\)\u003c/span\u003e\u003c/span\u003e). The construction of the line graph is as follows: each edge in the original graph\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:G\\)\u003c/span\u003e\u003c/span\u003e corresponds to a node in the line graph.If two edges in\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:G\\:\\)\u003c/span\u003e\u003c/span\u003eshare a common node.Then,in \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{L}_{\\text{G}}\\)\u003c/span\u003e\u003c/span\u003e, an edge will be created between the corresponding nodes\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{v}_{ij}\\)\u003c/span\u003e\u003c/span\u003eand \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{v}_{jk}\\)\u003c/span\u003e\u003c/span\u003e.Specifically, this work first use the open-source toolkit RDKit to extract the scaffold graph and transform it into a line graph \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{L}_{{G}_{S}}\\)\u003c/span\u003e\u003c/span\u003e.Then, to prepare for the next step of node selection, this work represent each element\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{a}_{ij,jk}\\)\u003c/span\u003e\u003c/span\u003ein the adjacency matrix of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{L}_{{G}_{S}}\\)\u003c/span\u003e\u003c/span\u003eas:\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$$\\:{a}_{ij,jk}\\:=\\left\\{\\begin{array}{c}1,\\:\\:\\:if\\:{v}_{ij},\\:{v}_{jk}\\:are\\:connected\\:\\\\\\:0,\\:\\:otherwise\\end{array}\\right.$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eIn this way, the operation of selecting four nodes can be simplified to the operation of selecting two edges when swapping edges.\u003c/p\u003e \u003cp\u003e\u0026bull; Node selection\u003c/p\u003e \u003cp\u003eAfter constructing the line graph, this work first randomly select a node \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e in \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{L}}_{{\\text{G}}_{\\text{S}}}\\)\u003c/span\u003e\u003c/span\u003e. Then, to reduce the search space as described above, this work introduce the adjacency matrix of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{L}}_{{\\text{G}}_{\\text{S}}}\\)\u003c/span\u003e\u003c/span\u003e. Specifically, the operation of reducing the search space is divided into the following steps.\u003c/p\u003e \u003cp\u003eFirst, to avoid selecting edges with common nodes during the edge-swapping process, or generating edges that already exist in the original graph, this work perform some masking operations on the nodes using the adjacency matrix. Let \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{i}\\)\u003c/span\u003e\u003c/span\u003e be the index of the first node selected for edge swapping. Here, this work introduce a mask vector \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{M}_{1}\\)\u003c/span\u003e\u003c/span\u003e\u003cem\u003e= {\u003c/em\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{m}_{1}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{m}_{2}\\)\u003c/span\u003e\u003c/span\u003e, \u003cem\u003e...\u003c/em\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{m}_{\\left|{V}_{L}\\right|}\\)\u003c/span\u003e\u003c/span\u003e\u003cem\u003e}\u003c/em\u003e. To ensure that the same node is not selected, this work set each element of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{M}}_{1}\\)\u003c/span\u003e\u003c/span\u003e as follows:\u003cdiv id=\"Equ5\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e\n$$\\:{m}_{u}\\:=\\left\\{\\begin{array}{c}1,\\:if\\:{a}_{ui}=0,u=1,...,\\left|{V}_{L}\\right|\\:\\\\\\:0,\\:otherwise\\end{array}\\right.$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThen, to avoid reconnecting to edges that already exist, this work perform a logical AND operation between the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e-th row of the adjacency matrix and every other row. Each element of the matrix is represented as follows:\u003cdiv id=\"Equ6\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ6\" name=\"EquationSource\"\u003e\n$$\\:{a}_{uv}^{{\\prime\\:}}={a}_{uv}\u0026middot;{a}_{iv},\\:u,\\:v\\:=\\:1,\\:...,\\:\\left|{V}_{L}\\right|$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eWhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:|{V}_{L}\\)\u003c/span\u003e\u003c/span\u003e\u003cem\u003e|\u003c/em\u003e is the number of nodes in \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{L}_{{G}_{S}}\\)\u003c/span\u003e\u003c/span\u003e. After this operation, if the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:v\\)\u003c/span\u003e\u003c/span\u003e-th element of the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:u\\)\u003c/span\u003e\u003c/span\u003e-th row is 1, the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:v\\)\u003c/span\u003e\u003c/span\u003e-th row will be logically ANDed with both the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:u\\)\u003c/span\u003e\u003c/span\u003e-th row and the\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\:i\\)\u003c/span\u003e\u003c/span\u003e-th row. If any row contains an element equal to 1, the corresponding node will be masked, i.e.,\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\:M}_{1}\\)\u003c/span\u003e\u003c/span\u003e will have \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{m}_{u}\\)\u003c/span\u003e\u003c/span\u003e= 0. In addition, to prevent the generated graph from deviating too much from the original graph due to significant changes in the scaffold, this work also mask the corresponding nodes in the line graph where entire rows become zero after the first logical AND operation. This work can restrict the edge-swapping operation to nodes within its 2-hop neighborhood. Finally, to further reduce the probability of edge-swapping failure, this work also mask the corresponding edges with different valency values. To achieve this, this work construct another mask vector \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{M}_{2}\\)\u003c/span\u003e\u003c/span\u003e \u003cem\u003e= {\u003c/em\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{m}_{1}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{m}_{2}\\)\u003c/span\u003e\u003c/span\u003e, \u003cem\u003e...\u003c/em\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{m}_{\\left|{V}_{L}\\right|}\\)\u003c/span\u003e\u003c/span\u003e\u003cem\u003e}\u003c/em\u003e, and set each element of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{M}_{2}\\)\u003c/span\u003e\u003c/span\u003e as follows:\u003cdiv id=\"Equ7\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ7\" name=\"EquationSource\"\u003e\n$$\\:{m}_{u}\\:=\\left\\{\\begin{array}{c}1,\\:if\\:{w}_{u}\\:={w}_{i}\\\\\\:0,\\:otherwise\\end{array}\\right.$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e7\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eAfter all the above masking operations, this work randomly select an index \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{j}\\)\u003c/span\u003e\u003c/span\u003e where both elements in \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{M}}_{1}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{M}}_{2}\\)\u003c/span\u003e\u003c/span\u003e are 1 as the second node \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{\\text{j}}\\)\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e\u0026bull; Swap edges\u003c/p\u003e \u003cp\u003eIn the above method, this work have selected two nodes \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{\\text{j}}\\)\u003c/span\u003e\u003c/span\u003e in the line graph to prepare for edge swapping. Then convert these two nodes into the corresponding edges in the scaffold graph to obtain \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{\\text{s}1}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{\\text{s}2}\\)\u003c/span\u003e\u003c/span\u003e. This work ultimate goal is to swap edges in the original molecular graph, so here this work need to map \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{\\text{s}1}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{\\text{s}2}\\)\u003c/span\u003e\u003c/span\u003e in the scaffold graph to the corresponding edges in the original graph \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{G}\\)\u003c/span\u003e\u003c/span\u003e, i.e., this work obtain \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{1}\\)\u003c/span\u003e\u003c/span\u003e=(\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{1}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{2}\\)\u003c/span\u003e\u003c/span\u003e, ω) and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{2}\\)\u003c/span\u003e\u003c/span\u003e =༈\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{3}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{4}\\)\u003c/span\u003e\u003c/span\u003e, ω༉.Then, disconnect \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{1}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{2}\\)\u003c/span\u003e\u003c/span\u003e and reconnect them as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{1}\\)\u003c/span\u003e\u003c/span\u003e =༈\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{1}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{4}\\)\u003c/span\u003e\u003c/span\u003e, ω༉and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{e}}_{2}\\)\u003c/span\u003e\u003c/span\u003e =༈\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{2}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{v}}_{3}\\)\u003c/span\u003e\u003c/span\u003e, ω༉.During the generation process, multiple modifications can be made until the modification rate or the number of failures exceeds the number of edges in the original graph.\u003c/p\u003e \u003cp\u003e \u003cb\u003e2.3 The framework of the molecular property prediction model based on a dual-channel feature separation network and contrastive learning\u003c/b\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eDCFS-CL is developed based on MolCLR [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. The latent representations of positive augmented molecular graph pairs are contrasted with those of negative augmented molecular graph pairs. The entire model (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) consists of four components: data processing and augmentation, a GNN-based feature extractor, a nonlinear projection head, and the normalized temperature-scaled cross-entropy (NT-Xent) [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e] contrastive loss.\u003c/p\u003e \u003cp\u003eFor molecular property prediction, this work first construct the SMILES data \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{s}}_{\\text{n}}\\)\u003c/span\u003e\u003c/span\u003e into the corresponding molecular graph \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{G}}_{\\text{n}}\\)\u003c/span\u003e\u003c/span\u003e. In these molecular graphs, each node represents an atom, and each edge represents a chemical bond between atoms. Next, this work apply molecular graph augmentation strategies to transform the original molecular graph \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{G}}_{\\text{n}}\\)\u003c/span\u003e\u003c/span\u003e into two different but related molecular graphs \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{G}}_{\\text{n}\\_1}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{G}}_{\\text{n}\\_2}\\)\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eWhen constructing positive and negative sample pairs, if the two augmented graphs come from the same molecule (i.e., \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{G}}_{\\text{n}\\_1}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{G}}_{\\text{n}\\_2}\\)\u003c/span\u003e\u003c/span\u003e are generated from the same \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{G}}_{\\text{n}}\\)\u003c/span\u003e\u003c/span\u003e), they are considered a positive sample pair. If the augmented graphs come from different molecules, they are considered a negative sample pair. This work can train the model using contrastive learning, allowing it to learn the similarities and differences between molecules.\u003c/p\u003e \u003cp\u003eTo extract features from molecular graphs, this work use a pair of Graph Neural Networks (GNNs) as feature extractors, each responsible for extracting different types of features. The intrinsic feature extractor mainly extracts the inherent features \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_1}^{\\text{i}\\text{n}\\text{h}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_2}^{\\text{i}\\text{n}\\text{h}}\\)\u003c/span\u003e\u003c/span\u003e of the molecules, such as atom types, bond types, etc. These are the foundational features of the molecular graph, and after graph augmentation, these features do not change for positive pairs. The auxiliary feature extractor mainly extracts the auxiliary features \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_1}^{\\text{a}\\text{u}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_2}^{\\text{a}\\text{u}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e of the molecules, such as atomic chirality information, bond stereochemistry information, etc. These features capture finer structural differences between molecules, and after graph augmentation, these features will change.\u003c/p\u003e \u003cp\u003eAfter feature extraction, this work further process the representations \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_1}^{\\text{i}\\text{n}\\text{h}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_2}^{\\text{i}\\text{n}\\text{h}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_1}^{\\text{a}\\text{u}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_2}^{\\text{a}\\text{u}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e through a nonlinear projection head g(\u0026middot;). This projection head consists of a multi-layer perceptron (MLP) with one hidden 7layer, which maps the feature vectors \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_1}^{\\text{i}\\text{n}\\text{h}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_2}^{\\text{i}\\text{n}\\text{h}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_1}^{\\text{a}\\text{u}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{h}}_{\\text{n}\\_2}^{\\text{a}\\text{u}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e to vectors in the latent space \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{z}}_{\\text{n}\\_1}^{\\text{i}\\text{n}\\text{h}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{z}}_{\\text{n}\\_2}^{\\text{i}\\text{n}\\text{h}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{z}}_{\\text{n}\\_1}^{\\text{a}\\text{u}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{z}}_{\\text{n}\\_2}^{\\text{a}\\text{u}\\text{x}}\\)\u003c/span\u003e\u003c/span\u003e, respectively. Next, the two latent vectors are combined to generate a total feature. The representation of the total feature can be considered as a fusion of intrinsic and auxiliary features. The goal of this mapping is to further enhance the consistency of positive sample pairs and the difference of negative sample pairs in the representation space during contrastive learning.\u003cdiv id=\"Equ8\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ8\" name=\"EquationSource\"\u003e\n$$\\:{\\text{z}}_{\\text{n}\\_1}^{\\text{t}\\text{o}\\text{t}}={\\text{z}}_{\\text{n}\\_1}^{\\text{i}\\text{n}\\text{h}}+{\\text{z}}_{\\text{n}\\_1}^{\\text{a}\\text{u}\\text{x}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e8\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ9\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ9\" name=\"EquationSource\"\u003e\n$$\\:{\\text{z}}_{\\text{n}\\_2}^{\\text{t}\\text{o}\\text{t}}={\\text{z}}_{\\text{n}\\_2}^{\\text{i}\\text{n}\\text{h}}+{\\text{z}}_{\\text{n}\\_2}^{\\text{a}\\text{u}\\text{x}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e9\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTo achieve this goal, this work use the Normalized Temperature Cross-Entropy (NT-Xent) loss function. This loss function is specifically designed for contrastive learning. For positive sample pairs, this work aim to maximize the similarity between the shared features and the total features, while minimizing the similarity of the private features. For negative sample pairs, this work aim to reduce the similarity between the shared features and the total features, while enhancing the difference of the private features.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAfter the contrastive learning pretraining is completed, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, this work further fine-tune the pretrained model for molecular property prediction tasks. In this stage, the structure of the prediction model is similar to that of the pretrained model, consisting of a GNN backbone and an MLP head. The GNN backbone uses the same architecture as the pretrained feature extractor, while the MLP head is used to map the features extracted by the GNN to the predicted molecular properties.\u003c/p\u003e \u003cp\u003eDuring the fine-tuning process, the GNN backbone is initialized by sharing parameters from the pretrained model, leveraging the rich structural information learned during pretraining. The MLP head is randomly initialized, allowing it to adapt more flexibly to different molecular property prediction tasks. Subsequently, this work train the entire fine-tuned model on the target molecular property dataset using supervised learning. In this way, the model can be fine-tuned on new molecular data for specific tasks, further improving prediction performance.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Experimental evaluation with molecular property","content":"\u003cp\u003eIn the field of molecular property prediction, classification tasks play a crucial role. Different datasets are designed around specific biological or pharmacological questions. For example, BBBP, BACE, and HIV focus on single-label classification problems, such as determining whether a molecule can cross the blood-brain barrier, inhibit BACE-1, or exhibit anti-HIV activity. In contrast, Clintox, Tox21, Sider, and MUV are multilabel classification datasets, requiring the simultaneous prediction of molecular behavior across multiple toxicity pathways, side effects, or activity targets. These benchmark datasets provide a rich foundation for developing and evaluating molecular classification models, driving advances in drug discovery and toxicity risk assessment.\u003c/p\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Dataset and Preprocession\u003c/h2\u003e \u003cp\u003eThis section provides implementation details. To demonstrate the effectiveness of the approach across various datasets, this work conducted experiments on seven molecular datasets from MoleculeNet [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], including BACE and BBBP for single-task classification and MUV, HIV, SIDER, ClinTox, and Tox21 for multi-task classification.\u003c/p\u003e \u003cp\u003e \u003cb\u003eBBBP (Blood-Brain Barrier Penetration)\u003c/b\u003e categorizes compounds based on their ability to penetrate or fail to penetrate the blood-brain barrier (BBBP). The BBBP is critical in drug development as it determines whether a drug can access the central nervous system.\u003c/p\u003e \u003cp\u003e \u003cb\u003eBACE (Beta-secretase 1)\u003c/b\u003e is designed to predict the inhibitory activity of compounds on the enzyme beta-secretase 1 (BACE-1), a key target in Alzheimer\u0026rsquo;s drug development.\u003c/p\u003e \u003cp\u003e \u003cb\u003eHIV\u003c/b\u003e predicts whether a compound has the ability to inhibit HIV virus replication. It contains anti-HIV screening results published by the Drug Therapeutics Program (DTP).\u003c/p\u003e \u003cp\u003e \u003cb\u003eClinTox (Clinical Toxicity)\u003c/b\u003e predicts whether a compound is safe or toxic based on FDA-approved drug safety data. It includes labeled information on toxic and safe compounds.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTox21\u003c/b\u003e predicts 12 types of toxicity associated with compounds. Tox21 aims to assess the potential impact of drugs and chemicals on human health.\u003c/p\u003e \u003cp\u003e \u003cb\u003eSIDER (Side Effect Resource)\u003c/b\u003e contains information on the side effects of FDA-approved drugs, categorized into 27 common side effect types.\u003c/p\u003e \u003cp\u003e \u003cb\u003eMUV (Maximum Unbiased Validation)\u003c/b\u003e is a high-quality dataset used for benchmarking virtual screening. It includes activity and inactivity data for compounds targeting 17 biological targets, designed to validate virtual screening models.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDataset Summary\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGraphs\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTasks\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSplit\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eClassification task type\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ebbbp\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2039\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eROC-AUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eScaffold\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSingle-label classification\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ebace\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1513\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eROC-AUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eScaffold\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSingle-label classification\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ehiv\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e41127\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eROC-AUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSingle-label classification\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eclintox\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1478\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eROC-AUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMulti-label classification\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003etox21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7831\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eROC-AUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMulti-label classificatio\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003esider\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1427\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eROC-AUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMulti-label classificatio\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Evaluation metrics\u003c/h2\u003e \u003cp\u003e \u003cb\u003eROC-AUC (Receiver Operating Characteristic - Area Under the Curve)\u003c/b\u003e is a widely used metric for evaluating the performance of classification models, particularly for measuring their ability to distinguish between positive and negative samples. It combines the ROC curve and the AUC value:\u003c/p\u003e \u003cp\u003e1.The \u003cb\u003eROC curve\u003c/b\u003e illustrates the model\u0026rsquo;s performance at different threshold values. It is a plot where the False Positive Rate (FPR) is on the x-axis and the True Positive Rate (TPR) is on the y-axis.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTrue Positive Rate (TPR)\u003c/b\u003e is proportion of actual positive samples correctly predicted as positive. It is calculated as:\u003cdiv id=\"Equ10\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ10\" name=\"EquationSource\"\u003e\n$$\\:TPR=\\:\\frac{TP}{TP+FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e10\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e \u003cb\u003eFalse Positive Rate (FPR)\u003c/b\u003e is proportion of samples predicted as positive that are actually negative. It is calculated as:\u003cdiv id=\"Equ11\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ11\" name=\"EquationSource\"\u003e\n$$\\:FPR=\\frac{FP}{FP+TN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e11\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e2.\u003cb\u003eAUC (Area Under the Curve)\u003c/b\u003e is the area under the ROC curve. Its value ranges from 0 to 1, with values closer to 1 indicating a stronger ability to distinguish between classes. The specific meaning of AUC is the probability that a randomly chosen positive sample has a higher predicted probability than a randomly chosen negative sample.\u003c/p\u003e \u003cp\u003eAUC\u0026thinsp;=\u0026thinsp;1: The model's predictions are perfect.\u003c/p\u003e \u003cp\u003eAUC\u0026thinsp;=\u0026thinsp;0.5: The model's predictions are no better than random guessing.\u003c/p\u003e \u003cp\u003eAUC\u0026thinsp;\u0026lt;\u0026thinsp;0.5: The model's predictions are worse than random guessing, indicating that there might be an issue with the model.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Loss Function\u003c/h2\u003e \u003cp\u003eSimCLR demonstrated that contrastive learning for images can greatly benefit from the combination of data augmentation and large batch sizes. Based on the InfoNCE loss, SimCLR introduced the NT-Xent loss.\u003c/p\u003e \u003cp\u003eIn this study, for each feature type (public, private, and total features), this work calculate the similarity of the augmented samples for each instance, denoted as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{s}\\text{i}\\text{m}}^{\\text{i}\\text{n}\\text{h}}({\\text{z}}_{\\text{i}}^{\\text{i}\\text{n}\\text{h}},{\\text{z}}_{\\text{j}}^{\\text{i}\\text{n}\\text{h}})\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{s}\\text{i}\\text{m}}^{\\text{t}\\text{o}\\text{t}}({\\text{z}}_{\\text{i}}^{\\text{t}\\text{o}\\text{t}},{\\text{z}}_{\\text{j}}^{\\text{t}\\text{o}\\text{t}})\\)\u003c/span\u003e\u003c/span\u003e, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{z}}_{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{z}}_{\\text{j}}\\)\u003c/span\u003e\u003c/span\u003e are the latent vectors extracted from positive data pairs, N is the batch size, sim(\u0026middot;) measures the similarity between two vectors, and τ is the temperature parameter. In DCFS-CL, this work follow the NT-Xent loss to pre-train the GNN encoder, and the cosine similarity is implemented as:\u003cdiv id=\"Equ12\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ12\" name=\"EquationSource\"\u003e\n$$\\:sim({z}_{i},{z}_{j})=\\frac{{z}_{i}^{T}{z}_{j}}{{\\left|\\right|{z}_{i}\\left|\\right|}_{2}{\\left|\\right|{z}_{j}\\left|\\right|}_{2}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e12\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe loss for intrinsic features \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{L}}_{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e:\u003cdiv id=\"Equ13\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ13\" name=\"EquationSource\"\u003e\n$$\\:{L}_{i}=-log\\frac{exp\\left({sim}^{inh}\\right({z}_{i}^{inh},{z}_{j}^{inh})/{\\tau\\:}^{inh})}{{\\sum\\:}_{k=1}^{2N}{1}_{i\\ne\\:k}exp\\left({sim}^{inh}\\right({z}_{i}^{inh},{z}_{k}^{inh})/{\\tau\\:}^{inh})}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e13\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe loss for auxiliary features \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{L}_{a}\\)\u003c/span\u003e\u003c/span\u003e:\u003cdiv id=\"Equ14\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ14\" name=\"EquationSource\"\u003e\n$$\\:{L}_{a}=-log\\frac{exp\\left({sim}^{aux}\\right({z}_{i}^{aux},{z}_{j}^{aux})/{\\tau\\:}^{aux})}{{\\sum\\:}_{k=1}^{2N}{1}_{i\\ne\\:k}exp\\left({sim}^{aux}\\right({z}_{i}^{aux},{z}_{k}^{aux})/{\\tau\\:}^{aux})}\\:$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e14\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe loss for auxiliary features \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{L}_{t}\\)\u003c/span\u003e\u003c/span\u003e:\u003cdiv id=\"Equ15\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ15\" name=\"EquationSource\"\u003e\n$$\\:{L}_{t}=-log\\frac{exp\\left({sim}^{tot}\\right({z}_{i}^{tot},{z}_{j}^{tot})/{\\tau\\:}^{tot}}{{\\sum\\:}_{k=1}^{2N}{1}_{i\\ne\\:k}exp\\left({sim}^{tot}\\right({z}_{i}^{tot},{z}_{k}^{tot})/{\\tau\\:}^{tot})}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e15\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eThe final loss combines the three types of feature losses with weighting coefficients \u003cem\u003ea\u003c/em\u003e and \u003cem\u003eb\u003c/em\u003e:\u003cdiv id=\"Equ16\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ16\" name=\"EquationSource\"\u003e\n$$\\:L={a\\ast\\:L}_{i}+b\\ast\\:{L}_{a}+{L}_{t}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e16\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Results and discussion\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance of different models tested on seven classification benchmarks.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eMethods\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"7\" nameend=\"c8\" namest=\"c2\"\u003e \u003cp\u003eClassification ROC-AUC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ebbbp\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ebace\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ehiv\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eclintox\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003etox21\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003esider\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003emuv\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGCN[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e71.8(0.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e71.6(2.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e74.0(3.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e62.5(2.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e70.9(2.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e53.6(3.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e71.6(4.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGIN[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e65.8(4.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e70.1(5.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e75.3(1.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e58.0(4.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e74.0(0.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e57.3(1.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e71.8(2.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSVM[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e72.9(0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e86.2(0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e79.2(0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e66.9(9.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e81.8(1.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e68.2(1.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e67.3(1.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e71.4(0.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e86.7(0.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e78.1(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e71.3(5.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e76.9(1.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e68.4(0.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e63.2(2.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eJOAO[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e71.4(0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e68.0\u0026nbsp;(1.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e76.6\u0026nbsp;(1.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e78.1\u0026nbsp;(3.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e74.7\u0026nbsp;(0.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e60.6\u0026nbsp;(1.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e74.2\u0026nbsp;(1.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eJOAOv2[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e68.0\u0026nbsp;(1.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e75.3\u0026nbsp;(3.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e76.9\u0026nbsp;(0.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e78.6\u0026nbsp;(3.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e74.9\u0026nbsp;(0.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e61.1\u0026nbsp;(0.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e75.3\u0026nbsp;(1.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMGSSL[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e70.5\u0026nbsp;(1.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e79.7\u0026nbsp;(\u0026nbsp;0.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e79.5\u0026nbsp;(1.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e80.7\u0026nbsp;(2.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e76.4\u0026nbsp;(0.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e61.8\u0026nbsp;(0.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e78.7\u0026nbsp;(1.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGraphCL[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e68.9(1.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e74.4\u0026nbsp;(2.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e77.3\u0026nbsp;(1.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e70.1\u0026nbsp;(7.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e74.2\u0026nbsp;(0.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e59.5\u0026nbsp;(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e71.9\u0026nbsp;(2.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGraphMAE[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e69.0\u0026nbsp;(2.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e81.3\u0026nbsp;(1.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e76.3\u0026nbsp;(1.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e73.1\u0026nbsp;(5.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e73.6\u0026nbsp;(2.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e59.7\u0026nbsp;(0.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e76.2\u0026nbsp;(1.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGraphMVP[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e67.9\u0026nbsp;(2.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e78.6\u0026nbsp;(3.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e76.1\u0026nbsp;(1.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e76.1\u0026nbsp;(4.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e75.4\u0026nbsp;(0.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e59.3\u0026nbsp;(1.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e74.4\u0026nbsp;(2.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMoleMCL[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e67.8\u0026nbsp;(3.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e80.3\u0026nbsp;(0.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e79.6\u0026nbsp;(0.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e80.2\u0026nbsp;(1.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e76.2\u0026nbsp;(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e59.0\u0026nbsp;(0.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e78.2\u0026nbsp;(1.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMolCLR[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e73.3(1.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e82.8(0.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e77.4(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e89.8(2.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e74.1(5.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e61.2(3.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e78.9(2.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDCFS-CL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e76.7(0.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e81.5(0.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e80.8(0.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e91.4(1.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e75.3(0.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e65.9(1.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e81.3(1.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAs shown in Table \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The first four models are supervised learning methods, while the latter eight are self-supervised/pre-trained methods. The table reports the average and standard deviation of the test ROC-AUC (%) for each benchmark.\u003c/p\u003e \u003cp\u003eThis work proposed a self-supervised learning framework based on a dual-channel feature separation network and contrastive learning for molecular property prediction. To validate the effectiveness of the model, this work conducted extensive experiments on 661 classification tasks across seven biological datasets and compared the results with existing supervised learning methods and self-supervised learning approaches. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presented the average ROC-AUC (%) values and standard deviations of different models tested on these datasets.\u003c/p\u003e \u003cp\u003eAs shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, DCFS-CL achieved an average ROC-AUC score of 81.9% across all seven classification datasets, which was significantly higher than that of other supervised learning and self-supervised/pretrained methods. In particular, DCFS-CL achieved a ROC-AUC score of 91.4% on the ClinTox dataset, surpassing the next-best method (MolCLR) by 1.6%. Additionally, DCFS-CL achieved a ROC-AUC score of 81.3% on the challenging dataset known for its imbalanced samples (MUV), outperforming MGSSL (78.7%) and other models. These above results indicated that DCFS-CL not only performed exceptionally well on simpler tasks but also demonstrated strong robustness and generalization capability on complex and imbalanced datasets.\u003c/p\u003e \u003cp\u003eThe first four methods in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (GCN, GIN, SVM, RF) represented traditional supervised learning approaches. While they performed well on certain datasets, their overall performance was relatively limited. For instance, although GIN and GCN could capture the local structural information in molecular graphs, they exhibited poor performance on complex tasks (such as ClinTox and MUV). Especially, they were prone to the over-smoothing problem in the large-scale data. As traditional machine learning methods, SVM and RF relied on manual feature extraction and excel in certain specific tasks, such as BACE and HIV. However, they lacked a deep understanding of molecular graph structures, resulting in insufficient generalization ability on other tasks. In contrast, DCFS-CL overcame the limitations of traditional methods by integrating a dualchannel feature separation network with contrastive learning, achieving consistently high performance across multiple tasks.\u003c/p\u003e \u003cp\u003eThe last eight methods in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (e.g., JOAO, GraphCL, MolCLR, etc.) were popular self-supervised learning or pre-training approaches, whose initialization representation capacity were enhanced through an unsupervised learning phase. However, these methods still exhibited certain performance limitations. For instance, JOAO and its improved version, JOAOv2, performed relatively well on some datasets (such as MUV), but they demonstrated weak generalization abilities, particularly under complex tasks (such as ClinTox, SIDER). GraphCL and GraphMAE employed graph augmentation techniques for contrastive learning, but their performance was limited by the choice of augmentation strategies. These strategies could lead to the loss of critical chemical information on molecule. MolCLR, as a state-of-the-art contrastive learning-based method, performed notably well on ClinTox dataset (89.8%), but its performance tended to be relatively unstable on other tasks, such as SIDER and MUV. DCFS-CL effectively addressed the limitations of the aforementioned methods by introducing graph augmentation strategies that align more closely with chemical logic and a dual-channel feature extraction mechanism. As a result, it demonstrated good generalization ability across multiple datasets.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Ablation experiment\u003c/h2\u003e \u003cp\u003eDCFS-CL was developed based on the MolCLR framework. The primary improvements in DCFS-CL were the graph augmentation strategy, which involves modifying the scaffold structure and functional group placement, and the dual-channel feature separation network, which separately extracts inherent and auxiliary features. This work denoted the operation of using new graph augmentation strategies as R, and the operation of employing the dual-channel feature separation network as G. To verify the effectiveness of the components in DCFS-CL, This work conducted an ablation study based on the MolCLR model.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAblation Study Results\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eMethods\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"7\" nameend=\"c8\" namest=\"c2\"\u003e \u003cp\u003eClassification ROC-AUC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ebbbp\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ebace\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ehiv\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eclintox\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003etox21\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003esider\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003emuv\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMolCLR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e73.3(1.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e82.8(0.7)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e77.4(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e89.8(2.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e74.1(5.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e61.2(3.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e78.9(2.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMolCLR\u0026thinsp;+\u0026thinsp;R\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e74.4(0.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e81.3(1.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e78.6(0.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90.6(1.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e74.8(0.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e63.5(2.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e80.8(1.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMolCLR\u0026thinsp;+\u0026thinsp;G\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e73.5(0.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e80.4(0.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e78.4(0.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90.1(1.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e75.1(0.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e63.1(1.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e79.6(0.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDCFS-CL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e76.7(0.9)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e81.5(0.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e80.8(0.4)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e91.4(1.6)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003e75.3(0.7)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003e65.9(1.3)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e81.3(1.7)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThis work designed multiple ablation experiments to evaluate the performance of each modification. The ROC-AUC values for the HIV dataset prediction are 77.4, 78.6, and 78.4 for the MolCLR, MolCLR\u0026thinsp;+\u0026thinsp;R, and MolCLR\u0026thinsp;+\u0026thinsp;G algorithms, respectively. These results indicate that both the inclusion of the graph augmentation strategy and the dual-channel feature separation network individually contribute to improving the model\u0026rsquo;s accuracy. In comparation with MolCLR, the average ROC-AUC of MolCLR\u0026thinsp;+\u0026thinsp;R on classification tasks increased from 74.1% to 75.8%, demonstrating an improvement of approximately 1.9%. The improved ROC-AUC was attributed to the property-preserving molecular graph augmentation strategy in MolCLR\u0026thinsp;+\u0026thinsp;R. This strategy focused on preserving the key structural information of molecules by adjusting the positions of nodes and edges rather than simply removing them, while fine-tuning the non-essential parts. The augmented molecular graphs retained the chemical properties of the original molecules and the data diversity was enhanced by this strategy. Consequently, the unlabeled data was effectively used in MolCLR\u0026thinsp;+\u0026thinsp;R, thereby reducing reliance on labeled data. Furthermore, the average ROC-AUC score of MolCLR\u0026thinsp;+\u0026thinsp;G on classification tasks increased from 74.1% (MolCLR) to 75.5%. This improvement was attributed to the dual-channel feature separation network. This network could extract both inherent and auxiliary features of molecules and decouple these features to avoid the mixing of feature types. This decoupling allowed for better control over the feature extraction process, enabling the model to focus on the correct feature type when handling molecular structural changes. Additionally, the semantic changes in molecular graphs could be sensitively captured by the dedicated auxiliary feature extraction network. As a result, it was found that both the graph augmentation strategy (R) and the dual-channel feature separation network (G) make significant contributions to the model\u0026rsquo;s performance, with neither being dispensable. This work complete model outperformed MolCLR across all datasets, significantly enhancing the model\u0026rsquo;s interpretability and task adaptability.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 experimental evaluation with performance prediction of flow batteries","content":"\u003cp\u003eIn this work, the Single Point HOMO Solution denotes the highest occupied molecular orbital (HOMO) energy calculated from a fixed molecular geometry using quantum chemical methods. This approach does not involve any additional geometry optimization steps, thereby enabling efficient and rapid estimation of electronic properties.\u003c/p\u003e \u003cp\u003eThe HOMO energy is directly related to the molecule\u0026rsquo;s oxidation potential and overall electrochemical stability, making it a key descriptor in the screening of redox-active materials for flow batteries. Accurate prediction of the Single Point HOMO Solution allows for high-throughput computational evaluation of candidate molecules, significantly accelerating the material discovery process.\u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Dataset and its processing\u003c/h2\u003e \u003cp\u003e \u003cb\u003eRedDB\u003c/b\u003e is a computational database specifically designed for aqueous redox flow batteries (ARFBs), encompassing 31,618 organic electroactive molecules. The database primarily includes two prominent classes of compounds: quinones and aza-aromatics.This work tested single point homo solution.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRedDB Dataset Summary\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGraphs\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTasks\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSplit\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRedDB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e31618\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMAE, RMSE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Evaluation Metrics\u003c/h2\u003e \u003cp\u003e \u003cb\u003e1.Mean Absolute Error (MAE)\u003c/b\u003e is a widely used metric in statistics and machine learning for evaluating the accuracy of predictive models, particularly in regression tasks. It quantifies the average magnitude of errors between predicted and actual values, without considering their direction.\u003c/p\u003e \u003cp\u003eMAE is calculated as the average of the absolute differences between predicted values \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{y}}_{\\text{i}}\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{y}}_{\\text{i}}^{{\\prime\\:}}\\)\u003c/span\u003e\u003c/span\u003e actual values across all observations:\u003cdiv id=\"Equ17\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ17\" name=\"EquationSource\"\u003e\n$$\\:MAE=\\frac{1}{n}{\\sum\\:}_{i=1}^{n}|{y}_{i}-{y}_{i}^{{\\prime\\:}}|$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e17\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003e2.\u003cb\u003eRoot Mean Square Error (RMSE)\u003c/b\u003e is a widely used metric to evaluate the performance of regression models. It measures the average magnitude of the errors between predicted values and actual values. The RMSE is calculated as follows:\u003cdiv id=\"Equ18\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ18\" name=\"EquationSource\"\u003e\n$$\\:RMSE=\\sqrt{\\frac{1}{n}{\\sum\\:}_{i=1}^{n}|{y}_{i}-{y}_{i}^{{\\prime\\:}}|\\:}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e18\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Experimental Results\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eModel Performance Comparison Using the RedDB Dataset for single_point_homo_solution\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003emodel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMAE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMPNN[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.083\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.117\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGCN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.075\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.101\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGAT[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.061\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.086\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAttentiveFP[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.060\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.086\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDCFS-CL\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.037\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.044\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe results shows the MAE and RMSE for different models, namely MPNN, GCN, GAT, AttentiveFP, and DCFS-CL. It is evident that DCFS-CL has lower MAE and RMSE values, indicating its superiority over all of the other models. Therefore, the results indicate that DCFS-CL is an effective tool for predicting electrochemical properties.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, the prediction performance of DCFS-CL is relatively accurate, which may be attributed to the proposed method's ability to leverage large-scale unlabeled data for pretraining in advance. Moreover, the graph augmentation strategy and dual-channel feature separation network employed by DCFS-CL demonstrate significant advantages over other baseline models in terms of generalization across chemical space and various molecular properties.\u003c/p\u003e \u003c/div\u003e"},{"header":"5 Conclusions","content":"\u003cp\u003eThis paper propose DCFS-CL(Dual-Channel Feature Separation with Contrastive Learningfor molecular property prediction based on feature decoupling and molecular contrastive learning. This work introduce a novel graph augmentation strategy and a dual-path feature separation network to extract inherent and auxiliary features of molecules separately. Experimental results demonstrate that DCFS-CL achieves significant improvements across multiple molecular benchmark datasets, even outperforming some traditional supervised learning methods. This framework not only enhances the accuracy of molecular representations but also improves the robustness of the model, offering new insights and approaches for molecular property prediction tasks.\u003c/p\u003e \u003cp\u003eTo summarize, this work introduce an innovative approach to molecular property prediction by integrating two key strategies: (1) This work combine scaffold augmentation with functional group augmentation and apply them to self-supervised learning. This combination allows for flexible manipulation of molecular graphs, ensuring that the essential scaffold structure remains intact while simultaneously enhancing the functional groups to generate diverse molecular representations. This strategy not only preserves the fundamental molecular features but also introduces variability, which is crucial for improving model generalization and performance. (2) This work propose a dual-channel feature separation network that separates inherent features, such as atomic types and bond types, from auxiliary features, such as atomic chirality and bond stereochemistry. This separation prevents interference between different types of features, allowing for more focused and effective feature extraction. By decoupling these features, the approach enhances the model\u0026rsquo;s ability to accurately capture the nuances of molecular structures, leading to improved predictions of molecular properties. Together, these strategies contribute to a more robust and flexible model that can be applied to various molecular property prediction tasks, offering new insights into molecular graph learning.\u003c/p\u003e \u003cp\u003eFurthermore, in future work, this work incorporating the domain knowledge embedded in large models can further improve the accuracy of molecular property prediction. Additionally, this work plans to apply DCFS-CL to the field of electrochemistry. Specifically, the model can be used to predict the properties of energy storage and conversion materials, such as batteries, supercapacitors, and fuel cells. By leveraging the molecular-level insights provided by framework, this work expects to make progress in material discovery, electrochemical performance optimization, and the development of more efficient and sustainable energy systems. Future research can focus on integrating electrochemical-specific data and domain knowledge into the model to enhance its predictive capabilities and accelerate the advancement of electrochemical technologies.\u003c/p\u003e"},{"header":"Declarations","content":" \u003cp\u003e \u003cb\u003eCompeting financial interests\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe authors declare no competing financial interests.\u003c/p\u003e\u003cp\u003e \u003ch2\u003eData availibility\u003c/h2\u003e \u003cp\u003eThe molecular SMILES strings and corresponding datasets are available \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://moleculenet.org/\u003c/span\u003e\u003cspan address=\"https://moleculenet.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e datasets-1.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis work was supported in part by the National Science and Technology Major Project of China under grant 2022ZD0119501, the National Natural Science Foundation of China under grant 52374221, the Natural Science Foundation of Shandong Province of China under grant ZR2023MF097 and ZR2024QF107, the Major Basic Research Project of the Shandong Provincial Natural Science Foundation under grant ZR2024ZD22.Qingdao West Coast New Area Science and Technology Project (Science and Technology Tackling Key Problems with Open Bidding) under grant 2022-5.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eFaming Lu proposed the core methodology and technical pipeline, conducted the main experimental design, model development, and result analysis, and drafted the initial manuscript. Yishan Zhu contributed to model implementation, experimental validation, and data analysis, and participated in revising and improving the manuscript. Zedong Lin provided overall supervision, participated in research discussions, and offered critical academic insights. Yunxia Bao, as one of the corresponding authors, guided the research direction, supervised the experimental design, and provided important revisions to the manuscript structure and academic presentation. Wanpeng Huang contributed to the analysis and discussion of experimental results and assisted in manuscript polishing and revision. All authors reviewed and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe molecular SMILES strings and corresponding datasets are available https:// moleculenet. org/ datasets-1.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eYou Y, Chen T, Sui Y et al. Graph Contrastive Learning with Augmentations[J]. arXiv preprint arXiv:2006.08242, 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHou Z, Wu Q, Zhang Z et al. GraphMAE: Self-Supervised Masked Graph Autoencoders[J]. arXiv preprint arXiv:2205.10803, 2022.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang S, Jiang D, Gao S et al. MGSSL: Self-Supervised Learning of Molecular Graphs with Domain Knowledge Enhanced Motif-Based Augmentation[J]. arXiv preprint arXiv:2106.13049, 2021.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Y, Wang J, Cao Z, Farimani AB. Molecular contrastive learning of representations via graph neural networks. Nat Mach Intell. 2022;4:279\u0026ndash;87.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu Z, et al. Moleculenet: a benchmark for molecular machine learning. Chem Sci. 2018;9:513\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu W, Liu Y, Yang Z. (2020). MolPre: Pretraining Molecular Graph Representations for Predicting Molecular Properties. arXiv preprint arXiv:2009.10013.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang X, Xu Y, Jiang C et al. MoleMCL: a multi-level contrastive learning framework for molecular pre-training.[J]. Bioinformatics (Oxford, England),2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBROUWERS E, VAN DER SCHAAF H. Improving Graph Attention Networks with Multi-hop Attention Aggregation[J]. IEEE Trans Neural Networks Learn Syst. 2021;32(5):2046\u0026ndash;55.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXU Z, XIE W. Temporal GraphSAGE: A Graph Neural Network for Dynamic Graphs[J]. IEEE Trans Neural Networks Learn Syst. 2022;33(9):4255\u0026ndash;66.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMA T, YOU Z, LIPTON Z, et al. Multi-View Graph Autoencoders for Multimodal Integration and Retrieval[J]. IEEE Trans Pattern Anal Mach Intell. 2021;43(7):2226\u0026ndash;37.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi Chenghong Z, Xiaofei. Dual-channel graph random convolutional network for semi-supervised node classification[J]. J Chin Comput Syst 2023,44(08):1656\u0026ndash;64.\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.20009/j.cnki.21-1106/TP.2021-0890\u003c/span\u003e\u003cspan address=\"10.20009/j.cnki.21-1106/TP.2021-0890\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVELIČKOVIĆ P, FEDUS W, HAMILTON W, L et al. Deep Graph Infomax[C]//International Conference on Learning Representations (ICLR). New Orleans, LA: ICLR, 2019.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang T, Wang Z, ,Yu W et al. Mix-Key: graph mixup with key structures for molecular property prediction. [J] Briefings Bioinf, 2024, 25 (3).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKipf NT, Welling M. Semi-Supervised Classification with Graph Convolutional Networks.[J]. CoRR,2016,abs/1609.02907.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu K, Hu W, Leskovec J, Jegelka S. How powerful are graph neural networks? In International Conference onLearning Representations (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCORTES C. Support-vector networks[J]. Mach Learn. 1995;20(3):273\u0026ndash;97.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBREIMAN L. Random forests[J]. Mach Learn. 2001;45(1):5\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLI C, WANG H, LI K et al. Joint optimized adversarial objective for graph representation learning[J]. Proceedings of the International Conference on Learning Representations (ICLR), 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLI C, WANG H, ZHU W et al. JOAOv2: Joint optimized adversarial objective for graph representation learning with improved performance[J]. Proceedings of the 37th International Conference on Machine Learning (ICML), 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGilmer J, Schoenholz SS, Riley PF, Vinyals O, Dahl G. E.Neural message passing for quantum chemistry. In Proc. 34th International Conference on Machine Learning 70 (eds Doina, P. \u0026amp;Teh, Y. W.) 1263\u0026ndash;1272 (PMLR, 2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVELIČKOVIĆ P, CUCURULL G, CASANOVA A et al. Graph attention networks[J]. Proceedings of the International Conference on Learning Representations (ICLR), 2018.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLIU S, WANG H, LIU W et al. AttentiveFP: attentive message passing for molecular property prediction[J]. Proceedings of the International Conference on Learning Representations (ICLR), 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOca\u0026ntilde;a AP, Rodriguez M J,Oca\u0026ntilde;a MMJ et al. An AI-Powered Methodology for Atomic-Scale Analysis of Heterogenized Correlated Single-Atom Catalysts.[J]. Small methods,2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJang YS, Jang UJ, Yoo YG et al. Nano-Interconnected 1D/2D Boron Nitride Hybrid Networks: Unlocking Superior Thermal Conductivity in Electrically Insulating Thermal Interface Nanocomposites Based on Hybrid Thermal Percolation Model.[J]. Small methods,2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWen J, Ji Y, Hu Y et al. Electrolyzing Outside the Box: Non-Traditional Approaches to Electrochemical Water Splitting.[J]. Small methods,2025.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Molecular Property Prediction, Self-Supervised Learning, Molecular Contrastive Learning, Dual-Channel Feature Separation Network, Intrinsic Features, Auxiliary Features","lastPublishedDoi":"10.21203/rs.3.rs-8846431/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8846431/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eContrastive learning, a prominent self-supervised technique, has shown great potential in molecular property prediction by reducing reliance on labeled data. However, existing methods face challenges such as disrupting molecular structures through aggressive graph augmentations and feature interference from extracting all features within a single network. To address these issues, this paper propose DCFS-CL (Dual-Channel Feature Separation with Contrastive Learning). First, a property-preserving augmentation strategy modifies non-essential structures while retaining key scaffolds. Then, a dual-channel network is introduced to separately extract inherent and auxiliary features, enhancing interpretability and task adaptability. As a result, DCFS-CL achieves outstanding accuracy and robustness across seven biological classification datasets and demonstrates high predictive performance on an electrochemical regression dataset. This framework offers strong generalization, making it well-suited for molecular screening and property prediction in data-scarce scenarios.\u003c/p\u003e","manuscriptTitle":"Molecular Property Prediction Based on Dual-Channel Feature Separation Network and Contrastive Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-20 01:33:37","doi":"10.21203/rs.3.rs-8846431/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"bab7484e-3038-434b-9a48-5f88f250f54f","owner":[],"postedDate":"February 20th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-24T15:40:21+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-20 01:33:37","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8846431","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8846431","identity":"rs-8846431","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}