Tissue-Specific Carcinogenicity Prediction Using Multi-Task Learning on Attention-based Graph Neural Networks

Tissue-Specific Carcinogenicity Prediction Using Multi-Task Learning on Attention-based Graph Neural Networks | 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 Tissue-Specific Carcinogenicity Prediction Using Multi-Task Learning on Attention-based Graph Neural Networks Yunju Song, Myeongjin Kim, Sunyong Yoo This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6518252/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 Cancer is caused by the uncontrolled growth and division of abnormal cells. In industrialized societies, chemical exposure is one of the leading causes of cancer. Indeed, since certain compounds can induce cancer by damaging genes or affecting cellular metabolism, studying carcinogens is essential. However, previous studies have not considered that compounds may promote different tissue-specific carcinogenicity. Therefore, this study developed a multi-task learning framework to predict tissue-specific carcinogenicity in the liver, lung, stomach, and breast tissues. This framework consisted of a shared layer to extract common features and task-specific layers to perform task-specific predictions. The shared layer contains a graph attention network (GAT) layer to make atom representations that reflect the importance of neighboring atoms and parallel fully connected layers designed for each task combination. These shared representations are then passed to task-specific layers to predict tissue-specific carcinogenicity. This entire training process was conducted through stepwise learning, whereby the model was trained in the first step using partially labeled data for tissues, and the initial weights were determined during this process. The second step trained the model using fully labeled data for all tissues, allowing the model to perform the final training for carcinogenicity prediction. The results demonstrated that the proposed multi-task model achieved superior performance overall. The best performance was observed in the stomach task (AUROC: 0.825; AUPR: 0.867), outperforming single-task models (AUROC: 0.800; AUPR: 0.840) and previous studies (AUROC: 0.743–0.791; AUPR 0.788–0.827). We further analyzed molecules with high predicted carcinogenicity in each tissue and identified critical substructures for the prediction using the attention mechanism. This research can contribute to predicting the tissue-specific carcinogenicity of candidate chemicals in the early stages of drug development, thereby reducing research costs and time. cancer carcinogenicity prediction multi-task learning graph attention network attention mechanism tissue specificity Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Scientific contribution This study presents a framework for predicting tissue-specific carcinogenicity and addresses the limitations of previous research, which treated carcinogenicity independently of tissue type. This framework uses multi-task learning with GAT and stepwise learning to extract shared molecular representations and perform tissue-specific predictions. This model outperforms existing approaches and provides interpretability via attention-based substructure analysis. Introduction Cancer is a disease caused by the uncontrolled proliferation of cells and represents one of the leading causes of death worldwide, causing millions of deaths annually [ 1 ]. Cancer can be caused by environmental factors, with exposure to carcinogens being a major risk [ 2 ]. Humans in industrialized societies are continuously exposed to various chemicals; some are classified as carcinogens [ 3 ]. These carcinogenic chemicals have the potential to promote cancer through DNA damage or by disrupting cellular metabolic processes [ 4 ]. Subsequently, the analysis of carcinogens is becoming increasingly important. Carcinogen evaluations have traditionally utilized in vitro and in vivo experimental assays [ 5 – 7 ]. In vitro assays have the advantages of not requiring animal experiments and reducing resources and time [ 8 , 9 ]. However, these in vitro assays cannot account for the physiological effects of compounds on tissues, as these assays are limited to isolated cellular environments [ 9 ]. In contrast, while in vivo assays can assess systemic effects, these are resource- and time-intensive. Moreover, the carcinogenicity of only a small proportion of chemical compounds has been tested using in vitro and in vivo methods [ 9 , 10 ]. Thus, evaluating numerous chemicals using these experimental approaches is time-consuming and costly. To address these issues, in silico methods represent an effective time and cost alternative to predict the carcinogenicity of many compounds [ 11 ]. Therefore, these in silico methods can be effectively applied to select substances for analysis before conducting in vitro or in vivo experiments. Previous studies using in silico methods have utilized machine learning techniques such as support vector machines, random forest, artificial neural networks, and graph-based models [ 12 – 19 ]. Within these methods, it is important to select model structures that effectively capture the properties of compounds for accurate carcinogenicity prediction. Most compounds have complex structures, and various interactions occur between the atoms in the molecular structure [ 20 , 21 ]. These interactions are critical elements that determine the physicochemical properties, biological activity, toxicity, and pharmacological effects of the compounds. To capture these relationships, graph-based methods such as graph convolutional networks (GCNs) and graph attention networks (GATs) were introduced to represent molecules as graphs, where atoms are nodes and bonds are edges [ 22 – 26 ]. These graph-based methods allow the direct encoding of molecular structures without requiring complex transformations, enabling efficient learning of chemical properties. However, GCNs aggregate information from neighboring nodes with the same weight to update the representation of each node, which may constrain the ability of this network to capture the critical atoms in molecular structures [ 26 – 28 ]. To overcome this limitation, a GAT assigns attention weights to relationships between nodes during information aggregation from neighbors, giving higher weights to important atoms [ 28 – 30 ]. Therefore, carcinogenicity prediction using GAT methods has been actively studied recently; however, most of these studies are limited, focusing on general carcinogenicity prediction without considering tissue specificity [ 14 , 16 ]. This has the problem of not specifically predicting whether a particular compound can cause cancer in a certain tissue or organ, leading to underestimation or overestimation of the true risk of the compound. The carcinogenic effects of chemicals are different in each tissue and are closely related to the tissue specificity of the cancer [ 31 ]. Several studies have demonstrated that tumor suppressor genes and oncogenes exhibit tissue-specific expression, indicating that certain compounds may induce cancer in some tissues but not others [ 32 , 33 ]. To address these issues, multi-task learning is needed to determine the carcinogenicity of compounds in multiple tissues simultaneously. This approach is effective because it has been shown to improve the generalization performance of models by learning various tasks simultaneously. Therefore, previous studies have applied multi-task learning to biological tasks such as predicting properties of compounds (e.g., side effects, toxicity, protein targets) and predicting aquatic toxicity or lethal doses in different species [ 34 – 37 ]. Based on previous biological approaches, we constructed a multi-task learning framework for tissue-specific carcinogenicity prediction using GAT that reflects the molecular structure and properties. This model uses three-task combination data in the first step and data from all tasks in the second. This approach allows us to simultaneously capture common features across tasks and tissue-specific features to more efficiently use the information and improve prediction performance compared to building individual binary classification models [ 38 , 39 ]. Materials and Methods 2.1 Data collection To construct the integrated database, we collected data from four sources, including the Carcinogenic Potency Database (CPDB), the Chemical Carcinogenesis Research Information System (CCRIS), the United States Environmental Protection Agency (EPA) Integrated Risk Information System (IRIS), and the Toxin and Toxin Target Database (T3DB). Each database contains carcinogenicity evaluation data for various chemicals and provides comprehensive information on the relevant experiments. The CPDB was developed at the National Library of Medicine and the National Institutes of Health [ 40 ]. The CPDB contains the results of carcinogenicity bioassays for 1547 substances. This database provides information on the following variables, including species, sex, route of administration, target organ, and carcinogenicity status, and includes experimental results published in scientific papers and experimental results reported in technical reports by the National Cancer Institute (NCI)/National Toxicology Program (NTP). The CCRIS was developed at the NCI and comprises separate databases on carcinogenicity, mutagenicity, tumor promotion, and tumor suppression bioassay data for chemical substances [ 41 ]. This study uses the carcinogenicity bioassays database, which provides information on 1767 substances, including details on the species, sex, route of administration, tumor site, and carcinogenicity status. The United States Environmental Protection Agency (EPA) maintains the IRIS, which includes human health risk assessment information for 738 chemicals commonly found in the environment [ 42 ]. The database provides detailed information on the route of administration, tumor site and type, toxicity values, and human carcinogenicity status of the chemical substances under investigation. The T3DB is a toxin and toxin–target database developed by the Wishart Research Group [ 43 ]. The T3DB contains data on 3678 toxins and provides extensive information, including toxicity values, human carcinogenicity classification, and chemical characteristics of tissues and toxins. This study only used the all-toxin data from the T3DB. 2.2 Data labeling The CPDB contains information on simplified molecular input line entry system (SMILES) data of compounds, target tissues, toxic dose 50% (TD50), and carcinogenicity experimental data. For the carcinogenicity experimental results, we removed data labeled “0”, which does not provide clear evidence of carcinogenicity, and data labeled “e”, which indicates uncertain evidence of carcinogenicity; data labeled “c”, “p”, “a”, and “+” were classified as “carcinogenic positive”, while data labeled “-” were classified as “carcinogenic negative” (Table 1 ). Table 1 Classification of carcinogenicity labels based on CPDB. CPDB data CPDB results Descriptions Labels Literature-based data + This indicates that the author in the literature evaluated the tissue–tumor combination as being induced by the test agent. Positive - The author explicitly indicated that the test agent did not induce tumors at the specific site. Negative 0 The author provided no opinion or an ambiguous opinion on the carcinogenicity of the test agent. Remove NCI/NTP data c The test agent was evaluated as carcinogenic in the NCI/NTP Technical Report. Positive p There is some evidence of carcinogenicity. Positive a The tumors are associated with carcinogenicity, or the evidence is suggestive. Positive e There is equivocal evidence of carcinogenicity. Remove 0 NCI/NTP did not present an evaluation for this tissue–tumor combination, or the experiment was evaluated as inadequate. Remove - For NCI/NTP experiments without a “c”, “p”, “a”, or “e” opinion, one site is denoted a “-” opinion unless the experiment was inadequate. Negative The CCRIS contains information on SMILES of compounds, tumor site information, and carcinogenicity experimental results. The results of the carcinogenicity experiments are presented in a string format, such as “negative”, indicating no carcinogenicity in the test, “positive” indicating carcinogenicity, and “equivocal”, indicating an ambiguous result. To ensure consistency within the dataset, we excluded data representing a certain percentage of tumors or unclear results. Only data classified as “positive” or “negative” were used. The “positive” data were labeled as “carcinogenic positive”, and the “negative” data as “carcinogenic negative”. The IRIS database comprises numerous pieces of information, including the CAS number, assessment type (cancer/non-cancer), critical effect tumor type (toxicity type caused by compound exposure), and toxicity value. This study initially mapped the SMILES data provided by PubChem based on the CAS number. Meanwhile, based on assessment type data, compounds were labeled as “carcinogenic positive” when classified as “cancer” and “carcinogenic negative” when classified as “non-cancer”. Furthermore, data were extracted on the tissues in which carcinogenicity was reported through the critical effect tumor type field. The T3DB comprises CAS numbers, toxicity data, carcinogenicity data, and data regarding the affected tissue. As with the IRIS database, the SMILES data provided by PubChem were initially mapped based on the CAS number. The carcinogenicity evaluation was conducted using the carcinogenicity data, which adheres to the carcinogenic classification criteria defined by the International Agency for Research on Cancer (IARC) [ 44 ]. Accordingly, data corresponding to IARC Group 1, 2A, and 2B, which indicate a potential carcinogenic risk, were labeled as “carcinogenic positive”. Conversely, data corresponding to Group 3 or showing no carcinogenic evidence were labeled “carcinogenic negative” (Table 2 ). Table 2 Carcinogenicity labeling of T3DB data based on IARC classification. IARC classification Descriptions Label Group 1 Carcinogenic to humans Positive Group 2A Probably carcinogenic to humans Positive Group 2B Possibly carcinogenic to humans Positive Group 3 Unclassifiable carcinogenic status to humans Negative 2.3 Data integration and preprocessing To integrate the data from the CPDB, CCRIS, IRIS, and T3DB, any data with unclear or redundant tissue names (e.g., upper digestive tract, respiratory organs, etc.) were removed, and mixtures of multiple compounds, single elements, and ionic compounds that cannot form graph structures were also excluded. Then, compounds with one or more toxic dose 50% (TD50) values or experimentally confirmed as carcinogenic were labeled carcinogenic [ 19 , 45 ]. Finally, we chose four tissues in which compounds can cause cancer and that represent the most frequent cancers in the statistics of the Global Cancer Observatory: liver cancer, lung cancer, stomach cancer, and breast cancer [ 46 ]. A database comprising 343 compounds was constructed, with each compound annotated with binary carcinogenicity labels across four distinct tissue types (Table 3 ). Table 3 Tissue-specific distribution of carcinogenicity database. Liver Lung Stomach Breast Positive 177 184 186 174 Negative 166 159 157 169 Total 343 343 343 343 2.4 Molecular graph feature vector generation We used the RDkit library to extract the structure and chemical properties of compounds from SMILES information, which is a string representation of a molecular structure [ 47 ]. The RDKit library provides functions for molecular structure analysis, which extracts atom and bond information from the SMILES string. The extracted data are converted into a graph structure with each atom in the molecule as a node and the bonds between atoms as edges. Each atom node is represented by a 75-dimensional feature vector based on 10 properties, which reflects the chemical and structural information (Fig. 1 ). These properties include the atom symbol, atom degree, formal charge, explicit valence, implicit valence, hybridization type, the total number of bonded hydrogen atoms, radicals, chirality, and aromaticity. A description is provided for each of the 10 atom properties, example values, and the number of encoding dimensions (Table 4 ). The atom symbol is represented using one-hot encoding, resulting in 37 dimensions. The atom degree is represented by six dimensions, the formal charge by three dimensions, the explicit and implicit valence by seven dimensions each, the hybridization type by three dimensions, the number of bonded hydrogen atoms by five dimensions, the radical by three dimensions, the chirality by three dimensions, and the aromaticity by one dimension. Additionally, the adjacency matrix is a two-dimensional array of connections between atoms and atoms in a graph. The adjacency matrix is constructed as a matrix of size \(\:N\times\:N\) , where \(\:N\) is the number of nodes, and provides information about the connections between nodes. Table 4 Atom feature vectors and dimensions for molecular modeling. Atom representation Descriptions Features Dimensions Atom symbol The symbol of the element represented using one-hot encoding C,N,Br,F… 37 Atom degree The number of neighboring atoms directly bonded to the atom 0,1,2,3,4,5 6 Formal charge The formal charge value of the atom -1,0,1 3 Explicit valence The number of electrons the atom has available for forming bonds with other atoms 0,1,2,3,4,5,6 7 Implicit valence The number of additional bonds the atom can potentially form 0,1,2,3,4,5,6 7 Hybridization type The hybridized orbital state of the atom sp, sp2, sp3 3 Total Hs number The total number of hydrogen atoms bonded to the atom 0,1,2,3,4 5 Radical The number of radical electrons the atom possesses 0,1,2 3 Chirality Whether the atom is a chiral center and the arrangement of its substituents around that center R,S,T/F 3 Aromatic Whether the atom has aromatic characteristics T/F 1 Total 75 2.5 Model overview We proposed a graph-based multi-task model that can address different tissue-specific tasks. The input for the model is provided in the form of molecular graphs, where a 75-dimensional feature vector represents each atom. This feature vector is processed through the GAT layer, where the features of each atom are updated through interactions with its neighbors. The GAT is applied in the first layer using the attention mechanism to learn how important each neighboring node is to the target node. Next, global attention pooling is applied after updating the features of each node through the GAT layer. This pooling mechanism learns weights for the nodes and aggregates their features to produce a 150 × 1 vector representing the entire graph [ 48 ]. The model is trained using three-task combinations to predict the four tasks (liver, lung, stomach, and breast). Independent, fully connected layers are constructed for each task combination, and the pooling vector is passed through these layers. The vector input to the layer corresponds to tissue-specific cancer information related to the molecule. After passing through the layers for each task combination, the vectors pass through the task-specific layer to generate the carcinogenicity prediction for each task. 2.6 Training multi-task model The proposed stepwise multi-task learning consists of two steps, where each step is designed to maximize the interaction between tasks during the learning process. In the first step of learning, three of the four tissue carcinogenicity prediction tasks are combined, resulting in four combinations: liver–lung–stomach, liver–lung–breast, liver–stomach–breast, and lung–stomach–breast. To train each combination, we used a corresponding dataset that contains only the binary carcinogenicity labels for the selected three tissues (Table 5 ). Each combination layer operates independently, focusing on learning features specific to its task combination. This step allows the model to learn important features shared in the three-task combinations while obtaining the optimal weights for each combination to be used as initial weights in the next step. In the second step, the model is trained using the initial weights learned in the first step. Using these initial weights enables faster and more reliable learning when the model is finally trained on data containing labels for all four tasks. Through this stepwise learning process, the model sufficiently reflects the task interactions in the first step from three-task combinations and then the second step with the entire dataset. Table 5 Three-tissue combinations carcinogenicity database for first-step learning. The number of three-tissue combination data Liver–lung–stomach 10 Liver–lung–breast 12 Liver–stomach–breast 5 Lung–stomach–breast 6 Total 33 During the training, hyperparameters such as learning rate, weight decay, and layer dimension were optimized through grid searches. Moreover, K -fold cross-validation was used to evaluate the generalization performance of the model. This method divides the entire dataset into K folds and performs cross-validation on each fold. This study used 20 folds, ensuring the trained model demonstrates consistent performance across the whole dataset. Additionally, to address the issue of sample imbalance within the dataset, the focal loss was used as the loss function [ 49 ]. Focal loss is a variation of binary cross entropy that applies additional weighting for class imbalance to help the model focus more on difficult predictions. Focal loss is defined as shown in formula (1), where \(\:{P}_{t}\) represents the probability of the class predicted by the model, \(\:{\alpha\:}_{t}\) denotes the class-specific weight, and \(\:\gamma\:\) is the focusing parameter, which is used to assign greater weight to difficult examples. We set the value of \(\:{\alpha\:}_{t}\) to 0.25 and \(\:\gamma\:\) to 4 to reduce the imbalance between carcinogenic and non-carcinogenic samples. $$\:FL\left({P}_{t}\right)=-{\alpha\:}_{t}{\left(1-{p}_{t}\right)}^{\gamma\:}log\left({p}_{t}\right)$$ 1 To accurately measure prediction probabilities for the binary classification tasks, task-specific loss was calculated and then summed across tasks to compute the overall loss. Through this process, the parameters of the model were updated to minimize the loss by considering the losses from all tasks. 2.7 Calculating attention weights The global attention pooling method generates a single representative feature vector for the entire graph. This method computes and normalizes a learnable importance score for each node embedding. The graph representation vector is then produced as a weighted sum of the transformed features. This approach makes it possible to identify important nodes in the overall graph, as shown in formula (2). In this formula, \(\:{r}_{i}\) represents the global attention pooling vector in the graph \(\:i\) , where \(\:{N}_{i}\:\) is the number of nodes in the graph, and \(\:{x}_{n}\:\) is the node feature vector for the node \(\:n\) . The function \(\:{h}_{gate}\left({x}_{n}\right)\) implemented as a simple multi-layer perceptron computes the importance of each node that produces a scalar. These scalars are normalized by \(\:softmax\) to yield attention weights \(\:{a}_{n}\) which sum to one. Attention weights \(\:{a}_{n}\) reflects how the critical node \(\:n\) is to the overall representation. The function \(\:{h}_{{\Theta\:}}(\bullet\:)\) then transforms the features of each node \(\:{x}_{n}\) producing \(\:{h}_{{\Theta\:}}\left({x}_{n}\right)\) . By multiplying the \(\:{a}_{n}\) with the transformed features \(\:{h}_{{\Theta\:}}\left({x}_{n}\right)\) , the importance of each node is reflected in its feature representation. Finally, the information from all nodes is aggregated by summing the weighted feature representations to produce the global representation \(\:{r}_{i}\:\) of graph \(\:i\) . Instead of mean pooling or max pooling, this approach provides a pooling mechanism using learnable importance to generate a graph representation. $$\:attention\:weights{=a}_{n}=\:softmax\left({h}_{gate}\left({x}_{n}\right)\right)$$ $$\:{r}_{i}=\sum\:_{n=1}^{{N}_{i}}{a}_{n}⨀{h}_{{\Theta\:}}\left({x}_{n}\right)$$ 2 2.8 Model evaluation metrics The evaluation of this model was conducted for each task using performance metrics such as the area under the receiver operating characteristic curve (AUROC) and the area under the precision–recall curve (AUPR). These allow for evaluating how effectively the model generalizes for each task. Formula (3) was used to calculate the performance metrics, where TP, FP, TN, and FN represent true positives, false positives, true negatives, and false negatives, respectively. \(\:TPR=\frac{TP}{TP+FN}\) \(\:FPR=\frac{FP}{FP+TN}\) \(\:Precision=\frac{TP}{TP+FP}\) \(\:Recall=\frac{TP}{TP+FN}\) (3) The AUROC is a metric that represents the probability of the model, correctly classifying positive classes as positive and negative classes as negative, and it is calculated using the true positive rate (TPR) and false positive rate (FPR). The TPR is the percentage of true positives that are correctly predicted to be positive, and the FPR is the percentage of true negatives that are incorrectly predicted to be positive. The ROC curve is a graph that presents the FPR on the x-axis and the TPR on the y-axis, and the AUROC is the area under this curve. The AUROC value ranges from 0 to 1, with values closer to 1 indicating better discrimination between positive and negative classes. The AUPR is a useful metric for evaluating model performance on imbalanced datasets, calculated using precision and recall. Precision refers to the proportion of true positives among the samples predicted as positive, while recall refers to the proportion of actual positives that were correctly predicted. The PR curve is a graph with recall presented on the x-axis and precision on the y-axis; the AUPR represents the area under this curve. The AUPR value ranges from 0 to 1, with values closer to 1 indicating that the model effectively predicts positive samples even when there is a larger amount of negative data. Results 3.1 Performance evaluation of carcinogenicity prediction We compared the evaluation results of our multi-task, single-task, and the current best-performing carcinogenicity prediction single-task models, CarcGC [ 15 ] and DCAMCP [ 16 ], to assess their performances. The CarcGC model predicts carcinogenicity using only GCN. In contrast, the DCAMCP model combines graph representation with molecular fingerprints and employs a capsule network to predict carcinogenicity. In evaluating the four specified tissue types, our multi-task model consistently outperformed the competing models regarding the AUROC and AUPR. Specifically, the multi-task approach in liver tissue achieved an AUROC of 0.757 ± 0.100 and an AUPR of 0.767 ± 0.113. This performance surpasses the single-task model (AUROC 0.706 ± 0.095; AUPR 0.742 ± 0.093), CarcGC model (AUROC 0.700 ± 0.152; AUPR 0.738 ± 0.142), and DCAMCP model (AUROC 0.680 ± 0.124; AUPR 0.722 ± 0.112). A similar pattern was observed in the lung tissue, with the multi-task model achieving an AUROC of 0.762 ± 0.099 and an AUPR of 0.788 ± 0.112. This performance exceeded that of the single-task model (AUROC 0.736 ± 0.106; AUPR 0.756 ± 0.090), CarcGC model (AUROC 0.709 ± 0.136; AUPR 0.738 ± 0.120), and DCAMCP model (AUROC 0.688 ± 0.103; AUPR 0.742 ± 0.092). These results demonstrate that the performance of the multi-task model was significantly improved for the liver and lung tissues compared to the single-task models. Among these four tasks, the highest performance was observed in stomach tissue, with an AUROC of 0.825 ± 0.103 and an AUPR of 0.867 ± 0.097. This result surpassed the performance of the single-task model (AUROC 0.800 ± 0.092; AUPR 0.840 ± 0.073), CarcGC model (AUROC 0.743 ± 0. 122; AUPR 0.788 ± 0.104), and DCAMCP model (AUROC 0.791 ± 0.119; AUPR 0.827 ± 0.114). The multi-task model exhibited the highest performance in breast tissue, achieving an AUROC of 0.801 ± 0.116 and an AUPR of 0.808 ± 0.117. This performance surpassed that of the single-task model (AUROC 0.768 ± 0.164; AUPR 0.783 ± 0.153), as well as the CarcGC model (AUROC 0.769 ± 0.124; AUPR 0.782 ± 0.131) and DCAMCP model (AUROC 0.761 ± 0.107; AUPR 0.801 ± 0.095). Collectively, these results demonstrate that the multi-task model achieves superior predictive reliability across various tissues and emphasizes the effectiveness of a multi-task learning framework in improving tissue-specific prediction performance (Table 6 ). Table 6 Performance evaluation of multi-task, single-task, CarcGC, and DCAMCP models. Tissue Model AUROC AUPR Liver Multi-task (Ours) 0.757 ± 0.100 0.767 ± 0.113 Single-task 0.706 ± 0.095 0.742 ± 0.093 CarcGC 0.700 ± 0.152 0.738 ± 0.142 DCAMCP 0.680 ± 0.124 0.722 ± 0.112 Lung Multi-task (Ours) 0.762 ± 0.099 0.788 ± 0.112 Single-task 0.736 ± 0.106 0.756 ± 0.090 CarcGC 0.709 ± 0.136 0.738 ± 0.120 DCAMCP 0.688 ± 0.103 0.742 ± 0.092 Stomach Multi-task (Ours) 0.825 ± 0.103 0.867 ± 0.097 Single-task 0.800 ± 0.092 0.840 ± 0.073 CarcGC 0.743 ± 0.122 0.788 ± 0.104 DCAMCP 0.791 ± 0.119 0.827 ± 0.114 Breast Multi-task (Ours) 0.801 ± 0.116 0.808 ± 0.117 Single-task 0.768 ± 0.164 0.783 ± 0.153 CarcGC 0.769 ± 0.124 0.782 ± 0.131 DCAMCP 0.761 ± 0.107 0.801 ± 0.095 3.2 Analysis of gradient cosine similarity To analyze the reason our multi-task model outperformed other models, we examined the gradient cosine similarity between task pairs of the shared layer parameters. A higher gradient cosine similarity indicates that the directions in which tasks update the shared layer parameters are similar [ 50 ]. The comparison of the average cosine similarity of shared layer gradients collected during the 20-fold cross-validation revealed that the gradient similarity in the breast task with other tasks was close to zero compared to other tasks, indicating that its parameter updates were largely independent. This suggests that the features required for the breast task differed slightly from those of the different tasks, limiting the benefits of shared features. Consequently, the AUPR performance difference between the multi-task and DCAMCP models for the breast task was minimal compared to other tasks. Nevertheless, the overall gradients cosine similarity showed a positive correlation among all task pairs in multi-task learning. This suggests that multi-task learning was advantageous due to synergistic gradient updates across tasks, which explains how the multi-task learning model consistently outperformed single-task models. 3.3 Substructure analysis of carcinogenic molecules To analyze the substructures crucial for predicting carcinogenic compounds, we investigated the molecular substructures with high attention weights among the carcinogen compounds that achieved high prediction scores in the test set. The molecular substructures of compounds with high prediction scores were highlighted visually, and evidence-based analysis using existing scientific literature was conducted to confirm whether these substructures are related to carcinogenicity. Carcinogenic substances that exhibit toxicity across all tissues include 2,7-dichlorodibenzo-p-dioxin and benzyl butyl phthalate. First, 2,7-dichlorodibenzo-p-dioxin is a type of dioxin, a class of organic compounds containing chlorine and generally known for its toxicity. In this study, an attention weights analysis highlighted the chlorine-containing ring structures. Previous studies have demonstrated that the toxicity of dioxins varies depending on the number and position of the chlorine atoms bonded to the ring structure [ 51 ]. Furthermore, the highlighted substructure, chlorobenzene, is known to form covalent bonds with DNA, RNA, and proteins. Chlorobenzene has also been confirmed to interact with DNA via microsomes in vitro experiments (Fig. 5 a) [ 52 ]. Second, the structure containing the two highlighted benzene rings is monobenzyl phthalate (MBzP), a metabolite of benzyl butyl phthalate. MBzP is classified as a non-genotoxic carcinogen, while an experiment on MBzP noted a significant decrease in global DNA methylation levels. Additionally, MBzP altered the promoter methylation levels of tumor suppressor genes ( P16 and TP53 ) and proto-oncogenes ( BCL2 and CCND1 ). As a result, the P16 and TP53 expressions decreased, whereas the expression of BCL2 and CCND1 increased. This indicates a reduction in tumor suppression and the potential to promote cell proliferation and cancer formation (Fig. 5 b) [ 53 ]. Compounds that exhibit carcinogenicity in the liver, lung, and stomach tissues include 1,1-dichloroethane, which contains an aliphatic halide. The aliphatic halide has been identified as a key substructure in predicting the carcinogenicity of 1,1-dichloroethane. This substructure has been shown to have the potential to induce mutations and promote carcinogenesis through covalent bonds with DNA after being activated into a reactive intermediate (Fig. 5 c) [ 54 ]. In the liver, lung, and breast tissues, the carcinogenic compound anti-dibenzo(a,l)pyrene-11,12-dihydrodiol-13,14-epoxide contains a polycyclic aromatic hydrocarbons (PAHs) substructure, which can be metabolized into reactive substances such as diolepoxides and radical PAH cations. These reactive metabolites interact with DNA to form DNA adducts, leading to DNA replication errors and the loss of purine bases, ultimately transforming affected genes into oncogenes. Additionally, PAH metabolites, such as quinones, generate reactive oxygen species (ROS), further contributing to DNA damage (Fig. 5 d) [ 55 – 59 ]. Another compound, 2-amino-5-(5-nitro-2-furyl)-1,3,4-oxadiazole, exhibits carcinogenic effects in the lung, stomach, and breast tissues. This compound contains a highlighted furan substructure known to possess potent carcinogenic properties. Nitrofuran also includes this furan substructure and has been identified as a carcinogenic compound. As a result of the carcinogenic and mutagenic potential, these compounds are banned in food-producing animals in the United States, the EU, Australia, and several other countries (Fig. 5 e) [ 60 , 61 ]. Lastly, the carcinogenic compound epichlorohydrin is highlighted the epoxide substructure in the stomach. Various epoxides and epoxide-forming chemicals have been proven to be carcinogenic. Epoxides act as strong alkylating agents and can form covalent bonds with DNA, potentially causing mutations and carcinogenesis. Epoxides can also react with biological nucleophiles in vivo (e.g., protein residues, DNA/RNA bases), leading to DNA damage (Fig. 5 f) [ 62 , 63 ]. Discussion Compared to previous methods that treat tasks independently, this proposed multi-task model learns more abundant representations by reflecting shared features across task combinations. Overall, the multi-task model demonstrated outstanding performance in all tasks, making it a powerful tool for handling various tissue-specific tasks. These results show the potential for the multi-task model to be widely applied to numerous types of biological data, demonstrating the utility of multi-task learning. Furthermore, this study analyzed the specific substructures of carcinogenic compounds that the model prioritized for prediction using attention weights, which represent the learned importance of each node in constructing the final graph representation. As a result, structures reflecting different carcinogenic mechanisms, such as chlorobenzene, aliphatic halide, PAHs, epoxide, furan, and MBzP, were identified as key substructures for carcinogenicity prediction. This indicates that the model captures various carcinogenic mechanisms, including DNA damage and epigenetic regulation. The model developed in this study demonstrates high predictive performance and explainability in all tissues, but there are some limitations and improvements to consider. The first issue is the restricted availability of databases, as most databases only record the general carcinogenicity of a compound. Even when using different databases, numerous duplicate compounds exist, which results in a lack of information on tissue-specific carcinogenicity. This limitation in training data may reduce the ability to provide tissue-specific predictions for new compounds. The second is a limitation in multi-task learning, which is generally considered effective for learning with limited data; however, a previous study has shown that multi-task learning requires a dataset large enough to train the minimum number of features [ 64 ]. However, due to the lack of carcinogen datasets, there is not enough data on compounds that cause carcinogenicity by tissue, making it difficult to make predictions for additional tissues. Therefore, we plan to conduct further research once a sufficient dataset is built, as the continued accumulation of experimental data is expected to enable more accurate and detailed tissue-specific carcinogenicity analyses. The last limitation is that the model proposed in this study does not consider factors important for carcinogenicity and toxicity assessment, such as route of administration, dose, and species specificity of compounds. This is due to the lack of structured databases and differences in toxicity evaluation criteria across databases. Therefore, future studies may need to establish structured databases or standardize those existing ones. Conclusion The continuous increase in cancer patients due to exposure to various carcinogenic compounds has highlighted the need for carcinogenicity prediction research. The multi-task learning model proposed in this study improved performance by reflecting the common features among task combinations during the learning process. This enables more accurate tissue-specific carcinogenicity predictions by effectively learning molecular structures and interactions. Moreover, analyzing the substructures of carcinogenic molecules allowed the identification of tissue-specific carcinogenic mechanisms and carcinogenic patterns that could be common to multiple tissues. These findings may contribute to identifying carcinogenic patterns of concern in populations exposed to carcinogens and provide data for more accurate carcinogenicity predictions. In conclusion, this study can be an important tool to reduce time and costs in evaluating and regulating carcinogenic substances, and it is expected to significantly contribute to developing personalized medicine research and cancer prevention strategies. Declarations Availability of data and materials The datasets generated and/or analyzed during the current study are available in the GitHub repository: https://github.com/bmil-jnu/TS-Carcinogenicity. Competing interests The authors declare the existence of no competing interests. Funding This research was supported by a grant (RS-2024-00332003, RS-2025-02215961) from Ministry of Food and Drug Safety, National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2023-00217317) and Innovative Human Resource Development for Local Intellectualization program through the Institute of Information & Communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT)(IITP-2024-RS-2022-00156287). Author’s contributions Y.S. and S.Y. contributed to the research methodology, model architecture design, result analysis, and writing of the original draft. Y.S. performed data curation, model implementation and training, and visualization. M.K. and S.Y. were responsible for project administration, manuscript review, and editing. All authors reviewed and approved the final manuscript. Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. References Jokhadze N, Das A, Dizon DS (2024). Global cancer statistics: A healthy population relies on population health. CA: A Cancer Journal for Clinicians, 74(3). Reif AE (1981). The causes of cancer: While some cancers are genetically fated to appear, most have now been traced to environmental factors. American Scientist, 69(4):437-447. Belpomme D, Irigaray P, Hardell L, Clapp R, Montagnier L, Epstein S, Sasco AJ (2007). The multitude and diversity of environmental carcinogens. Environmental research, 105(3):414-429. Blackadar CB (2016). Historical review of the causes of cancer. World journal of clinical oncology, 7(1):54. Bourcier T, McGovern T, Stavitskaya L, Kruhlak N, Jacobson-Kram D (2015). Improving prediction of carcinogenicity to reduce, refine, and replace the use of experimental animals. Journal of the American Association for Laboratory Animal Science, 54(2):163-169. Felter SP, Bhat VS, Botham PA, Bussard DA, Casey W, Hayes AW, Hilton GM, Magurany KA, Sauer UG, Ohanian EV (2021). Assessing chemical carcinogenicity: hazard identification, classification, and risk assessment. Insight from a toxicology forum state-of-the-science workshop. Critical Reviews in Toxicology, 51(8):653-694. Benigni R (2012). Alternatives to the carcinogenicity bioassay for toxicity prediction: are we there yet? Expert Opinion on Drug Metabolism & Toxicology, 8(4):407-417. Katt ME, Placone AL, Wong AD, Xu ZS, Searson PC (2016). In vitro tumor models: advantages, disadvantages, variables, and selecting the right platform. Frontiers in bioengineering and biotechnology, 4:12. Saeidnia S, Manayi A, Abdollahi M (2015). From in vitro experiments to in vivo and clinical studies; pros and cons. Current drug discovery technologies, 12(4):218-224. Fröhlich E, Salar-Behzadi S (2014). Toxicological assessment of inhaled nanoparticles: role of in vivo, ex vivo, in vitro, and in silico studies. International journal of molecular sciences, 15(3):4795-4822. Cheng F, Li W, Liu G, Tang Y (2013). In silico ADMET prediction: recent advances, current challenges and future trends. Current topics in medicinal chemistry, 13(11):1273-1289. Zhang L, Ai H, Chen W, Yin Z, Hu H, Zhu J, Zhao J, Zhao Q, Liu H (2017). CarcinoPred-EL: Novel models for predicting the carcinogenicity of chemicals using molecular fingerprints and ensemble learning methods. Scientific reports, 7(1):2118. Wang Y-W, Huang L, Jiang S-W, Li K, Zou J, Yang S-Y (2020). CapsCarcino: A novel sparse data deep learning tool for predicting carcinogens. Food and Chemical Toxicology, 135:110921. Fradkin P, Young A, Atanackovic L, Frey B, Lee LJ, Wang B (2022). A graph neural network approach for molecule carcinogenicity prediction. Bioinformatics, 38(Supplement_1):i84-i91. Zhang H, Yi H, Hao Y, Zhao L, Pan W, Xue Q, Liu X, Fu J, Zhang A (2024). Deciphering exogenous chemical carcinogenicity through interpretable deep learning: A novel approach for evaluating atmospheric pollutant hazards. Journal of Hazardous Materials, 465:133092. Chen Z, Zhang L, Sun J, Meng R, Yin S, Zhao Q (2023). DCAMCP: A deep learning model based on capsule network and attention mechanism for molecular carcinogenicity prediction. Journal of cellular and molecular medicine, 27(20):3117-3126. Tanabe K, Kurita T, Nishida K, Lučić B, Amić D, Suzuki T (2013). Improvement of carcinogenicity prediction performances based on sensitivity analysis in variable selection of SVM models. SAR and QSAR in Environmental Research, 24(7):565-580. Li T, Tong W, Roberts R, Liu Z, Thakkar S (2021). DeepCarc: Deep learning-powered carcinogenicity prediction using model-level representation. Frontiers in artificial intelligence, 4:757780. Moorthy NHN, Kumar S, Poongavanam V (2017). Classification of carcinogenic and mutagenic properties using machine learning method. Computational Toxicology, 3:33-43. Wu Y, Wang G (2018). Machine learning based toxicity prediction: from chemical structural description to transcriptome analysis. International journal of molecular sciences, 19(8):2358. Yang K, Swanson K, Jin W, Coley C, Eiden P, Gao H, Guzman-Perez A, Hopper T, Kelley B, Mathea M (2019). Analyzing learned molecular representations for property prediction. Journal of chemical information and modeling, 59(8):3370-3388. Liu K, Sun X, Jia L, Ma J, Xing H, Wu J, Gao H, Sun Y, Boulnois F, Fan J (2019). Chemi-net: a molecular graph convolutional network for accurate drug property prediction. International journal of molecular sciences, 20(14):3389. Zitnik M, Agrawal M, Leskovec J (2018). Modeling polypharmacy side effects with graph convolutional networks. Bioinformatics, 34(13):i457-i466. Sun M, Zhao S, Gilvary C, Elemento O, Zhou J, Wang F (2019). Graph convolutional networks for computational drug development and discovery. Briefings in Bioinformatics, 21(3):919-935.10.1093/bib/bbz042. Sacha M, Błaz M, Byrski P, Dabrowski-Tumanski P, Chrominski M, Loska R, Włodarczyk-Pruszynski P, Jastrzebski S (2021). Molecule edit graph attention network: modeling chemical reactions as sequences of graph edits. Journal of Chemical Information and Modeling, 61(7):3273-3284. Le T, Noé F, Clevert D-A (2022). Equivariant graph attention networks for molecular property prediction. arXiv preprint arXiv:2202.09891. Zhang S, Tong H, Xu J, Maciejewski R (2019). Graph convolutional networks: a comprehensive review. Computational Social Networks, 6(1):1-23. Vrahatis AG, Lazaros K, Kotsiantis S (2024). Graph attention networks: a comprehensive review of methods and applications. Future Internet, 16(9):318. Veličković P, Cucurull G, Casanova A, Romero A, Lio P, Bengio Y (2017). Graph attention networks. arXiv preprint arXiv:1710.10903. Lavecchia A (2024). Advancing drug discovery with deep attention neural networks. Drug Discovery Today:104067. Arya A, Arya S, Arya M (2011). Chemical carcinogen and cancer risk: an overview. J Chem Pharm Res, 3(5):621-631. Haigis KM, Cichowski K, Elledge SJ (2019). Tissue-specificity in cancer: The rule, not the exception. Science, 363(6432):1150-1151. Schneider G, Schmidt-Supprian M, Rad R, Saur D (2017). Tissue-specific tumorigenesis: context matters. Nature Reviews Cancer, 17(4):239-253. Jain S, Siramshetty VB, Alves VM, Muratov EN, Kleinstreuer N, Tropsha A, Nicklaus MC, Simeonov A, Zakharov AV (2021). Large-scale modeling of multispecies acute toxicity end points using consensus of multitask deep learning methods. Journal of chemical information and modeling, 61(2):653-663. Xu M, Yang H, Liu G, Tang Y, Li W (2022). In silico prediction of chemical aquatic toxicity by multiple machine learning and deep learning approaches. Journal of Applied Toxicology, 42(11):1766-1776. Wang Y, Wang B, Jiang J, Guo J, Lai J, Lian X-Y, Wu J (2021). Multitask CapsNet: an imbalanced data deep learning method for predicting toxicants. ACS omega, 6(40):26545-26555. Lim S, Lee YO, editors. Predicting chemical properties using self-attention multi-task learning based on SMILES representation. 2020 25th International Conference on Pattern Recognition (ICPR); 2021: IEEE. Zhang Y, Yang Q (2021). A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34(12):5586-5609. Caruana R (1997). Multitask learning. Machine learning, 28:41-75. Carcinogenic Potency Database. Accessed 09 Aug 2024. Chemical Carcinogenesis Research Information System (CCRIS). Accessed 09 Aug 2024. EPA Integrated Risk Information System (IRIS). Accessed 09 Aug 2024. Lim E, Pon A, Djoumbou Y, Knox C, Shrivastava S, Guo AC, Neveu V, Wishart DS (2010). T3DB: a comprehensively annotated database of common toxins and their targets. Nucleic acids research, 38(suppl_1):D781-D786. Accessed 09 Aug 2024. Organization WH (2019). International agency for research on cancer. Zhong M, Nie X, Yan A, Yuan Q (2013). Carcinogenicity prediction of noncongeneric chemicals by a support vector machine. Chemical Research in Toxicology, 26(5):741-749. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A (2024). Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians, 74(3):229-263. Landrum G (2013). Rdkit documentation. Release, 1(1-79):4. Li Y, Tarlow D, Brockschmidt M, Zemel R (2015). Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493. Ross T-Y, Dollár G, editors. Focal loss for dense object detection. proceedings of the IEEE conference on computer vision and pattern recognition; 2017. Yu T, Kumar S, Gupta A, Levine S, Hausman K, Finn C (2020). Gradient surgery for multi-task learning. Advances in neural information processing systems, 33:5824-5836. Li Q, Cui Y, Wang Z, Li Y, Yang H (2024). Toxicity assessment of dioxins and their transformation by-products from inferred degradation pathways. Science of The Total Environment, 937:173416. GRILLI S, ARFELLINI G, COLACCI A, MAZZULLO M, PRODI G (1985). In vivo and in vitro covalent binding of chlorobenzene to nucleic acids. Japanese Journal of Cancer Research GANN, 76(8):745-751. Sicińska P, Jabłońska E, Bukowska B, Balcerczyk A, Reszka E (2022). The selected epigenetic effects of phthalates: DBP, BBP and their metabolites: MBP, MBzP on human peripheral blood mononuclear cells (In Vitro). Toxicology in Vitro, 82:105369. DiRenzo A, Gandolfi A, Sipes I (1982). Microsomal bioactivation and covalent binding of aliphatic halides to DNA. Toxicology Letters, 11(3-4):243-252. Xue W, Warshawsky D (2005). Metabolic activation of polycyclic and heterocyclic aromatic hydrocarbons and DNA damage: a review. Toxicology and applied pharmacology, 206(1):73-93. Shimada T, Fujii‐Kuriyama Y (2004). Metabolic activation of polycyclic aromatic hydrocarbons to carcinogens by cytochromes P450 1A1 and1B1. Cancer science, 95(1):1-6. Baird WM, Hooven LA, Mahadevan B (2005). Carcinogenic polycyclic aromatic hydrocarbon‐DNA adducts and mechanism of action. Environmental and molecular mutagenesis, 45(2‐3):106-114. Henkler F, Stolpmann K, Luch A (2012). Exposure to polycyclic aromatic hydrocarbons: bulky DNA adducts and cellular responses. Molecular, Clinical and Environmental Toxicology: Volume 3: Environmental Toxicology:107-131. Dipple A. Polycyclic aromatic hydrocarbon carcinogenesis: an introduction. ACS Publications; 1985. McCalla D (1983). Mutagenicity of nitrofuran derivatives. Environmental mutagenesis, 5(5):745-765. Singh A, Smith DJ (2025). Disposition of Orally Administered [14C]-Nitrofurazone in Broiler Chickens. Journal of Agricultural and Food Chemistry. Melnick RL (2002). Carcinogenicity and mechanistic insights on the behavior of epoxides and epoxide‐forming chemicals. Annals of the New York Academy of Sciences, 982(1):177-189. Kostal J, Voutchkova-Kostal A, Weeks B, Zimmerman JB, Anastas PT (2012). A free energy approach to the prediction of olefin and epoxide mutagenicity and carcinogenicity. Chemical research in toxicology, 25(12):2780-2787. Standley T, Zamir A, Chen D, Guibas L, Malik J, Savarese S, editors. Which tasks should be learned together in multi-task learning? International conference on machine learning; 2020: PMLR. 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-6518252","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":450044937,"identity":"e056dd34-1819-4d1f-811a-fe22d32ab4fd","order_by":0,"name":"Yunju Song","email":"","orcid":"","institution":"Chonnam National University","correspondingAuthor":false,"prefix":"","firstName":"Yunju","middleName":"","lastName":"Song","suffix":""},{"id":450044938,"identity":"23efda1e-73bc-42fd-8718-07d64443dae5","order_by":1,"name":"Myeongjin Kim","email":"","orcid":"","institution":"Chonnam National University","correspondingAuthor":false,"prefix":"","firstName":"Myeongjin","middleName":"","lastName":"Kim","suffix":""},{"id":450044940,"identity":"8980d35a-5702-4dd0-9ba2-754c553e0b2b","order_by":2,"name":"Sunyong Yoo","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAuUlEQVRIiWNgGAWjYNCCCjYo4wDRWs6QrIWxDcYiRov8jBzDBx/n8eUZHGB++IHhzD3CWgxu5BgbztzGVmxwgM1YguFGMRFaJHLMpHm3sSVuOMBgxsDwIYEohwG1zAFpYf9GnBaGGyAtDSAtPEBbbhChxeDMs2LDGcfYEmce5imWSDhDjMPakzc++FBzLLHvePvGDx+OEeMwgQwDIHmMgYEZSBGjgYGB//gDIFlDlNpRMApGwSgYoQAAh6U6VyM6dMkAAAAASUVORK5CYII=","orcid":"","institution":"Chonnam National University","correspondingAuthor":true,"prefix":"","firstName":"Sunyong","middleName":"","lastName":"Yoo","suffix":""}],"badges":[],"createdAt":"2025-04-24 07:38:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6518252/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6518252/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":81990385,"identity":"a3d4fe54-1991-4248-8092-5c0aadbb9405","added_by":"auto","created_at":"2025-05-05 16:30:50","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":417720,"visible":true,"origin":"","legend":"\u003cp\u003eExample of a created molecular feature vector and adjacency matrix using ethanol. A 75-dimensional feature vector was generated for each atom in the molecule, including two carbons and one oxygen. The generated feature vectors were combined into a single vector of size 75 × \u003cem\u003eN\u003c/em\u003e, where \u003cem\u003eN\u003c/em\u003e is the number of atoms. Additionally, an adjacency matrix was created to represent the connectivity information where a value of 1 indicates that two atoms are connected, and a value of 0 indicates that they are not connected.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6518252/v1/6725da38d805ef88cee39997.png"},{"id":81990388,"identity":"955f49b8-ae7d-4251-912b-694d7f72d4e6","added_by":"auto","created_at":"2025-05-05 16:30:50","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":529297,"visible":true,"origin":"","legend":"\u003cp\u003eGraph-based multi-task model structure for compound carcinogenicity prediction. The adjacency matrix and feature vectors of the compounds are used as input for the model and passed through the GAT layer of the shared layer (blue box). After multi-head attention with two heads, the feature vectors are transformed to 150 × N and reduced to 150 × 1 vector by global attention pooling. The vector is input to the layer corresponding to the tissue information related to the carcinogenicity of the compound (red box), e.g., if a compound has carcinogenicity information about the liver, lung, and stomach, it is input to that layer. This information is passed through the task-specific layers to generate tissue-specific carcinogenicity predictions (yellow box).\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6518252/v1/167c7d66a9fae78d014cf21f.png"},{"id":81990391,"identity":"4184a009-9a82-43f9-8ce1-d2041e135b86","added_by":"auto","created_at":"2025-05-05 16:30:50","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":166825,"visible":true,"origin":"","legend":"\u003cp\u003eThe stepwise learning process for the proposed model. In the first step, the model weights are updated using only the data from three-task combinations. The updated weights are then used as the initial weights for the second step. The light blue box and light-yellow box represent the layer where weights have not been updated, while the dark blue box and dark yellow box indicate layers where the learned weights have been applied.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6518252/v1/be5648f700db0b8346361e96.png"},{"id":81990386,"identity":"e32e3d74-31be-45d4-be10-254f9ba2cd86","added_by":"auto","created_at":"2025-05-05 16:30:50","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":48809,"visible":true,"origin":"","legend":"\u003cp\u003eGradient cosine similarity observed through the heatmap. Task-specific gradients were extracted from the shared layer, and pairwise gradient cosine similarity analysis was conducted. This heatmap indicates that higher values are represented in red, while lower values are shown in blue. Since all gradient cosine similarities exhibit positive correlations, values of zero and negative correlations were not displayed. The heatmap reveals that the similarity between liver and stomach is the highest, whereas the similarity between lung and breast is the lowest.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6518252/v1/42c5176b80a30a66f1343e78.png"},{"id":81991199,"identity":"60bbb32e-0ae6-4b88-b665-ea1de24b73e9","added_by":"auto","created_at":"2025-05-05 16:38:50","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":451299,"visible":true,"origin":"","legend":"\u003cp\u003eTissue-specific carcinogen substructures with attention weights. (a) and (b) Compounds that exhibit carcinogenicity across all tissues. (c) Compounds that exhibit carcinogenicity in the liver, lung, and stomach, (d) in the liver, lung, and breast, (e) in the lung, stomach, and breast, and (f) in the breast. The red-highlighted regions indicate areas in the molecule with high attention weights. The yellow circles depict the potential substructures associated with carcinogenicity.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6518252/v1/d586c524fb3dcaa67786ca26.png"},{"id":92697432,"identity":"031e3528-7787-453c-bb21-2c941cd49fb8","added_by":"auto","created_at":"2025-10-03 07:17:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2560406,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6518252/v1/610aedc9-baa8-46e7-a7ba-e32988eececd.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Tissue-Specific Carcinogenicity Prediction Using Multi-Task Learning on Attention-based Graph Neural Networks","fulltext":[{"header":"Scientific contribution","content":"\u003cp\u003eThis study presents a framework for predicting tissue-specific carcinogenicity and addresses the limitations of previous research, which treated carcinogenicity independently of tissue type. This framework uses multi-task learning with GAT and stepwise learning to extract shared molecular representations and perform tissue-specific predictions. This model outperforms existing approaches and provides interpretability via attention-based substructure analysis.\u003c/p\u003e\n"},{"header":"Introduction","content":"\u003cp\u003eCancer is a disease caused by the uncontrolled proliferation of cells and represents one of the leading causes of death worldwide, causing millions of deaths annually [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Cancer can be caused by environmental factors, with exposure to carcinogens being a major risk [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Humans in industrialized societies are continuously exposed to various chemicals; some are classified as carcinogens [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. These carcinogenic chemicals have the potential to promote cancer through DNA damage or by disrupting cellular metabolic processes [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Subsequently, the analysis of carcinogens is becoming increasingly important. Carcinogen evaluations have traditionally utilized \u003cem\u003ein vitro\u003c/em\u003e and \u003cem\u003ein vivo\u003c/em\u003e experimental assays [\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. \u003cem\u003eIn vitro\u003c/em\u003e assays have the advantages of not requiring animal experiments and reducing resources and time [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. However, these \u003cem\u003ein vitro\u003c/em\u003e assays cannot account for the physiological effects of compounds on tissues, as these assays are limited to isolated cellular environments [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. In contrast, while \u003cem\u003ein vivo\u003c/em\u003e assays can assess systemic effects, these are resource- and time-intensive. Moreover, the carcinogenicity of only a small proportion of chemical compounds has been tested using \u003cem\u003ein vitro\u003c/em\u003e and \u003cem\u003ein vivo\u003c/em\u003e methods [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Thus, evaluating numerous chemicals using these experimental approaches is time-consuming and costly. To address these issues, \u003cem\u003ein silico\u003c/em\u003e methods represent an effective time and cost alternative to predict the carcinogenicity of many compounds [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Therefore, these \u003cem\u003ein silico\u003c/em\u003e methods can be effectively applied to select substances for analysis before conducting \u003cem\u003ein vitro\u003c/em\u003e or \u003cem\u003ein vivo\u003c/em\u003e experiments.\u003c/p\u003e \u003cp\u003ePrevious studies using \u003cem\u003ein silico\u003c/em\u003e methods have utilized machine learning techniques such as support vector machines, random forest, artificial neural networks, and graph-based models [\u003cspan additionalcitationids=\"CR13 CR14 CR15 CR16 CR17 CR18\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Within these methods, it is important to select model structures that effectively capture the properties of compounds for accurate carcinogenicity prediction. Most compounds have complex structures, and various interactions occur between the atoms in the molecular structure [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. These interactions are critical elements that determine the physicochemical properties, biological activity, toxicity, and pharmacological effects of the compounds. To capture these relationships, graph-based methods such as graph convolutional networks (GCNs) and graph attention networks (GATs) were introduced to represent molecules as graphs, where atoms are nodes and bonds are edges [\u003cspan additionalcitationids=\"CR23 CR24 CR25\" citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. These graph-based methods allow the direct encoding of molecular structures without requiring complex transformations, enabling efficient learning of chemical properties. However, GCNs aggregate information from neighboring nodes with the same weight to update the representation of each node, which may constrain the ability of this network to capture the critical atoms in molecular structures [\u003cspan additionalcitationids=\"CR27\" citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. To overcome this limitation, a GAT assigns attention weights to relationships between nodes during information aggregation from neighbors, giving higher weights to important atoms [\u003cspan additionalcitationids=\"CR29\" citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Therefore, carcinogenicity prediction using GAT methods has been actively studied recently; however, most of these studies are limited, focusing on general carcinogenicity prediction without considering tissue specificity [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. This has the problem of not specifically predicting whether a particular compound can cause cancer in a certain tissue or organ, leading to underestimation or overestimation of the true risk of the compound. The carcinogenic effects of chemicals are different in each tissue and are closely related to the tissue specificity of the cancer [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSeveral studies have demonstrated that tumor suppressor genes and oncogenes exhibit tissue-specific expression, indicating that certain compounds may induce cancer in some tissues but not others [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. To address these issues, multi-task learning is needed to determine the carcinogenicity of compounds in multiple tissues simultaneously. This approach is effective because it has been shown to improve the generalization performance of models by learning various tasks simultaneously. Therefore, previous studies have applied multi-task learning to biological tasks such as predicting properties of compounds (e.g., side effects, toxicity, protein targets) and predicting aquatic toxicity or lethal doses in different species [\u003cspan additionalcitationids=\"CR35 CR36\" citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBased on previous biological approaches, we constructed a multi-task learning framework for tissue-specific carcinogenicity prediction using GAT that reflects the molecular structure and properties. This model uses three-task combination data in the first step and data from all tasks in the second. This approach allows us to simultaneously capture common features across tasks and tissue-specific features to more efficiently use the information and improve prediction performance compared to building individual binary classification models [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e].\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Data collection\u003c/h2\u003e \u003cp\u003eTo construct the integrated database, we collected data from four sources, including the Carcinogenic Potency Database (CPDB), the Chemical Carcinogenesis Research Information System (CCRIS), the United States Environmental Protection Agency (EPA) Integrated Risk Information System (IRIS), and the Toxin and Toxin Target Database (T3DB). Each database contains carcinogenicity evaluation data for various chemicals and provides comprehensive information on the relevant experiments.\u003c/p\u003e \u003cp\u003eThe CPDB was developed at the National Library of Medicine and the National Institutes of Health [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]. The CPDB contains the results of carcinogenicity bioassays for 1547 substances. This database provides information on the following variables, including species, sex, route of administration, target organ, and carcinogenicity status, and includes experimental results published in scientific papers and experimental results reported in technical reports by the National Cancer Institute (NCI)/National Toxicology Program (NTP).\u003c/p\u003e \u003cp\u003eThe CCRIS was developed at the NCI and comprises separate databases on carcinogenicity, mutagenicity, tumor promotion, and tumor suppression bioassay data for chemical substances [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e]. This study uses the carcinogenicity bioassays database, which provides information on 1767 substances, including details on the species, sex, route of administration, tumor site, and carcinogenicity status.\u003c/p\u003e \u003cp\u003eThe United States Environmental Protection Agency (EPA) maintains the IRIS, which includes human health risk assessment information for 738 chemicals commonly found in the environment [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. The database provides detailed information on the route of administration, tumor site and type, toxicity values, and human carcinogenicity status of the chemical substances under investigation.\u003c/p\u003e \u003cp\u003eThe T3DB is a toxin and toxin\u0026ndash;target database developed by the Wishart Research Group [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e]. The T3DB contains data on 3678 toxins and provides extensive information, including toxicity values, human carcinogenicity classification, and chemical characteristics of tissues and toxins. This study only used the all-toxin data from the T3DB.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e2.2 Data labeling\u003c/h3\u003e\n\u003cp\u003eThe CPDB contains information on simplified molecular input line entry system (SMILES) data of compounds, target tissues, toxic dose 50% (TD50), and carcinogenicity experimental data. For the carcinogenicity experimental results, we removed data labeled \u0026ldquo;0\u0026rdquo;, which does not provide clear evidence of carcinogenicity, and data labeled \u0026ldquo;e\u0026rdquo;, which indicates uncertain evidence of carcinogenicity; data labeled \u0026ldquo;c\u0026rdquo;, \u0026ldquo;p\u0026rdquo;, \u0026ldquo;a\u0026rdquo;, and \u0026ldquo;+\u0026rdquo; were classified as \u0026ldquo;carcinogenic positive\u0026rdquo;, while data labeled \u0026ldquo;-\u0026rdquo; were classified as \u0026ldquo;carcinogenic negative\u0026rdquo; (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\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\u003eClassification of carcinogenicity labels based on CPDB.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCPDB data\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCPDB results\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDescriptions\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLabels\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eLiterature-based data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eThis indicates that the author in the literature evaluated the tissue\u0026ndash;tumor combination as being induced by the test agent.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eThe author explicitly indicated that the test agent did not induce tumors at the specific site.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eThe author provided no opinion or an ambiguous opinion on the carcinogenicity of the test agent.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRemove\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eNCI/NTP data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ec\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eThe test agent was evaluated as carcinogenic in the NCI/NTP Technical Report.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ep\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eThere is some evidence of carcinogenicity.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ea\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eThe tumors are associated with carcinogenicity, or the evidence is suggestive.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ee\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eThere is equivocal evidence of carcinogenicity.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRemove\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNCI/NTP did not present an evaluation for this tissue\u0026ndash;tumor combination, or the experiment was evaluated as inadequate.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRemove\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFor NCI/NTP experiments without a \u0026ldquo;c\u0026rdquo;, \u0026ldquo;p\u0026rdquo;, \u0026ldquo;a\u0026rdquo;, or \u0026ldquo;e\u0026rdquo; opinion, one site is denoted a \u0026ldquo;-\u0026rdquo; opinion unless the experiment was inadequate.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNegative\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 CCRIS contains information on SMILES of compounds, tumor site information, and carcinogenicity experimental results. The results of the carcinogenicity experiments are presented in a string format, such as \u0026ldquo;negative\u0026rdquo;, indicating no carcinogenicity in the test, \u0026ldquo;positive\u0026rdquo; indicating carcinogenicity, and \u0026ldquo;equivocal\u0026rdquo;, indicating an ambiguous result. To ensure consistency within the dataset, we excluded data representing a certain percentage of tumors or unclear results. Only data classified as \u0026ldquo;positive\u0026rdquo; or \u0026ldquo;negative\u0026rdquo; were used. The \u0026ldquo;positive\u0026rdquo; data were labeled as \u0026ldquo;carcinogenic positive\u0026rdquo;, and the \u0026ldquo;negative\u0026rdquo; data as \u0026ldquo;carcinogenic negative\u0026rdquo;.\u003c/p\u003e \u003cp\u003eThe IRIS database comprises numerous pieces of information, including the CAS number, assessment type (cancer/non-cancer), critical effect tumor type (toxicity type caused by compound exposure), and toxicity value. This study initially mapped the SMILES data provided by PubChem based on the CAS number. Meanwhile, based on assessment type data, compounds were labeled as \u0026ldquo;carcinogenic positive\u0026rdquo; when classified as \u0026ldquo;cancer\u0026rdquo; and \u0026ldquo;carcinogenic negative\u0026rdquo; when classified as \u0026ldquo;non-cancer\u0026rdquo;. Furthermore, data were extracted on the tissues in which carcinogenicity was reported through the critical effect tumor type field.\u003c/p\u003e \u003cp\u003eThe T3DB comprises CAS numbers, toxicity data, carcinogenicity data, and data regarding the affected tissue. As with the IRIS database, the SMILES data provided by PubChem were initially mapped based on the CAS number. The carcinogenicity evaluation was conducted using the carcinogenicity data, which adheres to the carcinogenic classification criteria defined by the International Agency for Research on Cancer (IARC) [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. Accordingly, data corresponding to IARC Group 1, 2A, and 2B, which indicate a potential carcinogenic risk, were labeled as \u0026ldquo;carcinogenic positive\u0026rdquo;. Conversely, data corresponding to Group 3 or showing no carcinogenic evidence were labeled \u0026ldquo;carcinogenic negative\u0026rdquo; (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\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\u003eCarcinogenicity labeling of T3DB data based on IARC classification.\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=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIARC classification\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescriptions\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLabel\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroup 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCarcinogenic to humans\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroup 2A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eProbably carcinogenic to humans\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroup 2B\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePossibly carcinogenic to humans\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGroup 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUnclassifiable carcinogenic status to humans\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003e2.3 Data integration and preprocessing\u003c/h3\u003e\n\u003cp\u003eTo integrate the data from the CPDB, CCRIS, IRIS, and T3DB, any data with unclear or redundant tissue names (e.g., upper digestive tract, respiratory organs, etc.) were removed, and mixtures of multiple compounds, single elements, and ionic compounds that cannot form graph structures were also excluded. Then, compounds with one or more toxic dose 50% (TD50) values or experimentally confirmed as carcinogenic were labeled carcinogenic [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. Finally, we chose four tissues in which compounds can cause cancer and that represent the most frequent cancers in the statistics of the Global Cancer Observatory: liver cancer, lung cancer, stomach cancer, and breast cancer [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. A database comprising 343 compounds was constructed, with each compound annotated with binary carcinogenicity labels across four distinct tissue types (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\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\u003eTissue-specific distribution of carcinogenicity database.\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=\"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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLiver\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLung\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eStomach\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eBreast\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e177\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e184\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e186\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e174\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e166\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e159\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e157\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e169\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e343\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e343\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e343\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e343\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003e2.4 Molecular graph feature vector generation\u003c/h3\u003e\n\u003cp\u003eWe used the RDkit library to extract the structure and chemical properties of compounds from SMILES information, which is a string representation of a molecular structure [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. The RDKit library provides functions for molecular structure analysis, which extracts atom and bond information from the SMILES string. The extracted data are converted into a graph structure with each atom in the molecule as a node and the bonds between atoms as edges. Each atom node is represented by a 75-dimensional feature vector based on 10 properties, which reflects the chemical and structural information (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThese properties include the atom symbol, atom degree, formal charge, explicit valence, implicit valence, hybridization type, the total number of bonded hydrogen atoms, radicals, chirality, and aromaticity. A description is provided for each of the 10 atom properties, example values, and the number of encoding dimensions (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). The atom symbol is represented using one-hot encoding, resulting in 37 dimensions. The atom degree is represented by six dimensions, the formal charge by three dimensions, the explicit and implicit valence by seven dimensions each, the hybridization type by three dimensions, the number of bonded hydrogen atoms by five dimensions, the radical by three dimensions, the chirality by three dimensions, and the aromaticity by one dimension. Additionally, the adjacency matrix is a two-dimensional array of connections between atoms and atoms in a graph. The adjacency matrix is constructed as a matrix of size \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:N\\times\\:N\\)\u003c/span\u003e\u003c/span\u003e, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:N\\)\u003c/span\u003e\u003c/span\u003e is the number of nodes, and provides information about the connections between nodes.\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\u003eAtom feature vectors and dimensions for molecular modeling.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAtom representation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescriptions\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFeatures\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDimensions\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAtom symbol\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe symbol of the element represented using one-hot encoding\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eC,N,Br,F\u0026hellip;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e37\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAtom degree\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe number of neighboring atoms directly bonded to the atom\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0,1,2,3,4,5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFormal charge\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe formal charge value of the atom\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-1,0,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExplicit valence\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe number of electrons the atom has available for forming bonds with other atoms\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0,1,2,3,4,5,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eImplicit valence\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe number of additional bonds the atom can potentially form\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0,1,2,3,4,5,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHybridization type\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe hybridized orbital state of the atom\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003esp, sp2, sp3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal Hs number\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe total number of hydrogen atoms bonded to the atom\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0,1,2,3,4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRadical\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe number of radical electrons the atom possesses\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0,1,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChirality\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWhether the atom is a chiral center and the arrangement of its substituents around that center\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR,S,T/F\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAromatic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWhether the atom has aromatic characteristics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eT/F\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003e75\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\n\u003ch3\u003e2.5 Model overview\u003c/h3\u003e\n\u003cp\u003eWe proposed a graph-based multi-task model that can address different tissue-specific tasks. The input for the model is provided in the form of molecular graphs, where a 75-dimensional feature vector represents each atom. This feature vector is processed through the GAT layer, where the features of each atom are updated through interactions with its neighbors. The GAT is applied in the first layer using the attention mechanism to learn how important each neighboring node is to the target node. Next, global attention pooling is applied after updating the features of each node through the GAT layer. This pooling mechanism learns weights for the nodes and aggregates their features to produce a 150 \u0026times; 1 vector representing the entire graph [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. The model is trained using three-task combinations to predict the four tasks (liver, lung, stomach, and breast). Independent, fully connected layers are constructed for each task combination, and the pooling vector is passed through these layers. The vector input to the layer corresponds to tissue-specific cancer information related to the molecule. After passing through the layers for each task combination, the vectors pass through the task-specific layer to generate the carcinogenicity prediction for each task.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Training multi-task model\u003c/h2\u003e \u003cp\u003eThe proposed stepwise multi-task learning consists of two steps, where each step is designed to maximize the interaction between tasks during the learning process. In the first step of learning, three of the four tissue carcinogenicity prediction tasks are combined, resulting in four combinations: liver\u0026ndash;lung\u0026ndash;stomach, liver\u0026ndash;lung\u0026ndash;breast, liver\u0026ndash;stomach\u0026ndash;breast, and lung\u0026ndash;stomach\u0026ndash;breast. To train each combination, we used a corresponding dataset that contains only the binary carcinogenicity labels for the selected three tissues (Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). Each combination layer operates independently, focusing on learning features specific to its task combination. This step allows the model to learn important features shared in the three-task combinations while obtaining the optimal weights for each combination to be used as initial weights in the next step. In the second step, the model is trained using the initial weights learned in the first step. Using these initial weights enables faster and more reliable learning when the model is finally trained on data containing labels for all four tasks. Through this stepwise learning process, the model sufficiently reflects the task interactions in the first step from three-task combinations and then the second step with the entire dataset.\u003c/p\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\u003eThree-tissue combinations carcinogenicity database for first-step learning.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe number of three-tissue combination data\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLiver\u0026ndash;lung\u0026ndash;stomach\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLiver\u0026ndash;lung\u0026ndash;breast\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLiver\u0026ndash;stomach\u0026ndash;breast\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLung\u0026ndash;stomach\u0026ndash;breast\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e33\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\u003eDuring the training, hyperparameters such as learning rate, weight decay, and layer dimension were optimized through grid searches. Moreover, \u003cem\u003eK\u003c/em\u003e-fold cross-validation was used to evaluate the generalization performance of the model. This method divides the entire dataset into \u003cem\u003eK\u003c/em\u003e folds and performs cross-validation on each fold. This study used 20 folds, ensuring the trained model demonstrates consistent performance across the whole dataset. Additionally, to address the issue of sample imbalance within the dataset, the focal loss was used as the loss function [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. Focal loss is a variation of binary cross entropy that applies additional weighting for class imbalance to help the model focus more on difficult predictions. Focal loss is defined as shown in formula (1), where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{P}_{t}\\)\u003c/span\u003e\u003c/span\u003e represents the probability of the class predicted by the model, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\alpha\\:}_{t}\\)\u003c/span\u003e\u003c/span\u003e denotes the class-specific weight, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\gamma\\:\\)\u003c/span\u003e\u003c/span\u003e is the focusing parameter, which is used to assign greater weight to difficult examples. We set the value of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\alpha\\:}_{t}\\)\u003c/span\u003e\u003c/span\u003e to 0.25 and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\gamma\\:\\)\u003c/span\u003e\u003c/span\u003e to 4 to reduce the imbalance between carcinogenic and non-carcinogenic samples.\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:FL\\left({P}_{t}\\right)=-{\\alpha\\:}_{t}{\\left(1-{p}_{t}\\right)}^{\\gamma\\:}log\\left({p}_{t}\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTo accurately measure prediction probabilities for the binary classification tasks, task-specific loss was calculated and then summed across tasks to compute the overall loss. Through this process, the parameters of the model were updated to minimize the loss by considering the losses from all tasks.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e2.7 Calculating attention weights\u003c/h3\u003e\n\u003cp\u003eThe global attention pooling method generates a single representative feature vector for the entire graph. This method computes and normalizes a learnable importance score for each node embedding. The graph representation vector is then produced as a weighted sum of the transformed features. This approach makes it possible to identify important nodes in the overall graph, as shown in formula (2). In this formula, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{r}_{i}\\)\u003c/span\u003e\u003c/span\u003e represents the global attention pooling vector in the graph \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e, where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{N}_{i}\\:\\)\u003c/span\u003e\u003c/span\u003eis the number of nodes in the graph, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{n}\\:\\)\u003c/span\u003e\u003c/span\u003eis the node feature vector for the node \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:n\\)\u003c/span\u003e\u003c/span\u003e. The function \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{h}_{gate}\\left({x}_{n}\\right)\\)\u003c/span\u003e\u003c/span\u003e implemented as a simple multi-layer perceptron computes the importance of each node that produces a scalar. These scalars are normalized by \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:softmax\\)\u003c/span\u003e\u003c/span\u003e to yield attention weights \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{a}_{n}\\)\u003c/span\u003e\u003c/span\u003e which sum to one. Attention weights \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{a}_{n}\\)\u003c/span\u003e\u003c/span\u003e reflects how the critical node \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:n\\)\u003c/span\u003e\u003c/span\u003e is to the overall representation. The function \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{h}_{{\\Theta\\:}}(\\bullet\\:)\\)\u003c/span\u003e\u003c/span\u003e then transforms the features of each node \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{n}\\)\u003c/span\u003e\u003c/span\u003e producing \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{h}_{{\\Theta\\:}}\\left({x}_{n}\\right)\\)\u003c/span\u003e\u003c/span\u003e. By multiplying the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{a}_{n}\\)\u003c/span\u003e\u003c/span\u003e with the transformed features \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{h}_{{\\Theta\\:}}\\left({x}_{n}\\right)\\)\u003c/span\u003e\u003c/span\u003e, the importance of each node is reflected in its feature representation. Finally, the information from all nodes is aggregated by summing the weighted feature representations to produce the global representation \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{r}_{i}\\:\\)\u003c/span\u003e\u003c/span\u003eof graph \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e. Instead of mean pooling or max pooling, this approach provides a pooling mechanism using learnable importance to generate a graph representation.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:attention\\:weights{=a}_{n}=\\:softmax\\left({h}_{gate}\\left({x}_{n}\\right)\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:{r}_{i}=\\sum\\:_{n=1}^{{N}_{i}}{a}_{n}⨀{h}_{{\\Theta\\:}}\\left({x}_{n}\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\n\u003ch3\u003e2.8 Model evaluation metrics\u003c/h3\u003e\n\u003cp\u003eThe evaluation of this model was conducted for each task using performance metrics such as the area under the receiver operating characteristic curve (AUROC) and the area under the precision\u0026ndash;recall curve (AUPR). These allow for evaluating how effectively the model generalizes for each task. Formula (3) was used to calculate the performance metrics, where TP, FP, TN, and FN represent true positives, false positives, true negatives, and false negatives, respectively.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"2\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:TPR=\\frac{TP}{TP+FN}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:FPR=\\frac{FP}{FP+TN}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Precision=\\frac{TP}{TP+FP}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Recall=\\frac{TP}{TP+FN}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e(3)\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 AUROC is a metric that represents the probability of the model, correctly classifying positive classes as positive and negative classes as negative, and it is calculated using the true positive rate (TPR) and false positive rate (FPR). The TPR is the percentage of true positives that are correctly predicted to be positive, and the FPR is the percentage of true negatives that are incorrectly predicted to be positive. The ROC curve is a graph that presents the FPR on the x-axis and the TPR on the y-axis, and the AUROC is the area under this curve. The AUROC value ranges from 0 to 1, with values closer to 1 indicating better discrimination between positive and negative classes. The AUPR is a useful metric for evaluating model performance on imbalanced datasets, calculated using precision and recall. Precision refers to the proportion of true positives among the samples predicted as positive, while recall refers to the proportion of actual positives that were correctly predicted. The PR curve is a graph with recall presented on the x-axis and precision on the y-axis; the AUPR represents the area under this curve. The AUPR value ranges from 0 to 1, with values closer to 1 indicating that the model effectively predicts positive samples even when there is a larger amount of negative data.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Performance evaluation of carcinogenicity prediction\u003c/h2\u003e \u003cp\u003eWe compared the evaluation results of our multi-task, single-task, and the current best-performing carcinogenicity prediction single-task models, CarcGC [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] and DCAMCP [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], to assess their performances. The CarcGC model predicts carcinogenicity using only GCN. In contrast, the DCAMCP model combines graph representation with molecular fingerprints and employs a capsule network to predict carcinogenicity.\u003c/p\u003e \u003cp\u003eIn evaluating the four specified tissue types, our multi-task model consistently outperformed the competing models regarding the AUROC and AUPR. Specifically, the multi-task approach in liver tissue achieved an AUROC of 0.757\u0026thinsp;\u0026plusmn;\u0026thinsp;0.100 and an AUPR of 0.767\u0026thinsp;\u0026plusmn;\u0026thinsp;0.113. This performance surpasses the single-task model (AUROC 0.706\u0026thinsp;\u0026plusmn;\u0026thinsp;0.095; AUPR 0.742\u0026thinsp;\u0026plusmn;\u0026thinsp;0.093), CarcGC model (AUROC 0.700\u0026thinsp;\u0026plusmn;\u0026thinsp;0.152; AUPR 0.738\u0026thinsp;\u0026plusmn;\u0026thinsp;0.142), and DCAMCP model (AUROC 0.680\u0026thinsp;\u0026plusmn;\u0026thinsp;0.124; AUPR 0.722\u0026thinsp;\u0026plusmn;\u0026thinsp;0.112). A similar pattern was observed in the lung tissue, with the multi-task model achieving an AUROC of 0.762\u0026thinsp;\u0026plusmn;\u0026thinsp;0.099 and an AUPR of 0.788\u0026thinsp;\u0026plusmn;\u0026thinsp;0.112. This performance exceeded that of the single-task model (AUROC 0.736\u0026thinsp;\u0026plusmn;\u0026thinsp;0.106; AUPR 0.756\u0026thinsp;\u0026plusmn;\u0026thinsp;0.090), CarcGC model (AUROC 0.709\u0026thinsp;\u0026plusmn;\u0026thinsp;0.136; AUPR 0.738\u0026thinsp;\u0026plusmn;\u0026thinsp;0.120), and DCAMCP model (AUROC 0.688\u0026thinsp;\u0026plusmn;\u0026thinsp;0.103; AUPR 0.742\u0026thinsp;\u0026plusmn;\u0026thinsp;0.092). These results demonstrate that the performance of the multi-task model was significantly improved for the liver and lung tissues compared to the single-task models. Among these four tasks, the highest performance was observed in stomach tissue, with an AUROC of 0.825\u0026thinsp;\u0026plusmn;\u0026thinsp;0.103 and an AUPR of 0.867\u0026thinsp;\u0026plusmn;\u0026thinsp;0.097. This result surpassed the performance of the single-task model (AUROC 0.800\u0026thinsp;\u0026plusmn;\u0026thinsp;0.092; AUPR 0.840\u0026thinsp;\u0026plusmn;\u0026thinsp;0.073), CarcGC model (AUROC 0.743\u0026thinsp;\u0026plusmn;\u0026thinsp;0. 122; AUPR 0.788\u0026thinsp;\u0026plusmn;\u0026thinsp;0.104), and DCAMCP model (AUROC 0.791\u0026thinsp;\u0026plusmn;\u0026thinsp;0.119; AUPR 0.827\u0026thinsp;\u0026plusmn;\u0026thinsp;0.114). The multi-task model exhibited the highest performance in breast tissue, achieving an AUROC of 0.801\u0026thinsp;\u0026plusmn;\u0026thinsp;0.116 and an AUPR of 0.808\u0026thinsp;\u0026plusmn;\u0026thinsp;0.117. This performance surpassed that of the single-task model (AUROC 0.768\u0026thinsp;\u0026plusmn;\u0026thinsp;0.164; AUPR 0.783\u0026thinsp;\u0026plusmn;\u0026thinsp;0.153), as well as the CarcGC model (AUROC 0.769\u0026thinsp;\u0026plusmn;\u0026thinsp;0.124; AUPR 0.782\u0026thinsp;\u0026plusmn;\u0026thinsp;0.131) and DCAMCP model (AUROC 0.761\u0026thinsp;\u0026plusmn;\u0026thinsp;0.107; AUPR 0.801\u0026thinsp;\u0026plusmn;\u0026thinsp;0.095). Collectively, these results demonstrate that the multi-task model achieves superior predictive reliability across various tissues and emphasizes the effectiveness of a multi-task learning framework in improving tissue-specific prediction performance (Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance evaluation of multi-task, single-task, CarcGC, and DCAMCP models.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTissue\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAUROC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAUPR\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e\u003cb\u003eLiver\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMulti-task (Ours)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.757\u0026thinsp;\u0026plusmn;\u0026thinsp;0.100\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.767\u0026thinsp;\u0026plusmn;\u0026thinsp;0.113\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSingle-task\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.706\u0026thinsp;\u0026plusmn;\u0026thinsp;0.095\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.742\u0026thinsp;\u0026plusmn;\u0026thinsp;0.093\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCarcGC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.700\u0026thinsp;\u0026plusmn;\u0026thinsp;0.152\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.738\u0026thinsp;\u0026plusmn;\u0026thinsp;0.142\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDCAMCP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.680\u0026thinsp;\u0026plusmn;\u0026thinsp;0.124\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.722\u0026thinsp;\u0026plusmn;\u0026thinsp;0.112\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e\u003cb\u003eLung\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMulti-task (Ours)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.762\u0026thinsp;\u0026plusmn;\u0026thinsp;0.099\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.788\u0026thinsp;\u0026plusmn;\u0026thinsp;0.112\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSingle-task\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.736\u0026thinsp;\u0026plusmn;\u0026thinsp;0.106\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.756\u0026thinsp;\u0026plusmn;\u0026thinsp;0.090\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCarcGC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.709\u0026thinsp;\u0026plusmn;\u0026thinsp;0.136\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.738\u0026thinsp;\u0026plusmn;\u0026thinsp;0.120\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDCAMCP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.688\u0026thinsp;\u0026plusmn;\u0026thinsp;0.103\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.742\u0026thinsp;\u0026plusmn;\u0026thinsp;0.092\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e\u003cb\u003eStomach\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMulti-task (Ours)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.825\u0026thinsp;\u0026plusmn;\u0026thinsp;0.103\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.867\u0026thinsp;\u0026plusmn;\u0026thinsp;0.097\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSingle-task\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.800\u0026thinsp;\u0026plusmn;\u0026thinsp;0.092\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.840\u0026thinsp;\u0026plusmn;\u0026thinsp;0.073\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCarcGC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.743\u0026thinsp;\u0026plusmn;\u0026thinsp;0.122\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.788\u0026thinsp;\u0026plusmn;\u0026thinsp;0.104\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDCAMCP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.791\u0026thinsp;\u0026plusmn;\u0026thinsp;0.119\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.827\u0026thinsp;\u0026plusmn;\u0026thinsp;0.114\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e\u003cb\u003eBreast\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMulti-task (Ours)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e0.801\u0026thinsp;\u0026plusmn;\u0026thinsp;0.116\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e0.808\u0026thinsp;\u0026plusmn;\u0026thinsp;0.117\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSingle-task\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.768\u0026thinsp;\u0026plusmn;\u0026thinsp;0.164\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.783\u0026thinsp;\u0026plusmn;\u0026thinsp;0.153\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCarcGC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.769\u0026thinsp;\u0026plusmn;\u0026thinsp;0.124\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.782\u0026thinsp;\u0026plusmn;\u0026thinsp;0.131\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDCAMCP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.761\u0026thinsp;\u0026plusmn;\u0026thinsp;0.107\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.801\u0026thinsp;\u0026plusmn;\u0026thinsp;0.095\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\u003e3.2 Analysis of gradient cosine similarity\u003c/h2\u003e \u003cp\u003eTo analyze the reason our multi-task model outperformed other models, we examined the gradient cosine similarity between task pairs of the shared layer parameters. A higher gradient cosine similarity indicates that the directions in which tasks update the shared layer parameters are similar [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e]. The comparison of the average cosine similarity of shared layer gradients collected during the 20-fold cross-validation revealed that the gradient similarity in the breast task with other tasks was close to zero compared to other tasks, indicating that its parameter updates were largely independent. This suggests that the features required for the breast task differed slightly from those of the different tasks, limiting the benefits of shared features. Consequently, the AUPR performance difference between the multi-task and DCAMCP models for the breast task was minimal compared to other tasks. Nevertheless, the overall gradients cosine similarity showed a positive correlation among all task pairs in multi-task learning. This suggests that multi-task learning was advantageous due to synergistic gradient updates across tasks, which explains how the multi-task learning model consistently outperformed single-task models.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Substructure analysis of carcinogenic molecules\u003c/h2\u003e \u003cp\u003eTo analyze the substructures crucial for predicting carcinogenic compounds, we investigated the molecular substructures with high attention weights among the carcinogen compounds that achieved high prediction scores in the test set. The molecular substructures of compounds with high prediction scores were highlighted visually, and evidence-based analysis using existing scientific literature was conducted to confirm whether these substructures are related to carcinogenicity. Carcinogenic substances that exhibit toxicity across all tissues include 2,7-dichlorodibenzo-p-dioxin and benzyl butyl phthalate. First, 2,7-dichlorodibenzo-p-dioxin is a type of dioxin, a class of organic compounds containing chlorine and generally known for its toxicity. In this study, an attention weights analysis highlighted the chlorine-containing ring structures. Previous studies have demonstrated that the toxicity of dioxins varies depending on the number and position of the chlorine atoms bonded to the ring structure [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]. Furthermore, the highlighted substructure, chlorobenzene, is known to form covalent bonds with DNA, RNA, and proteins. Chlorobenzene has also been confirmed to interact with DNA via microsomes \u003cem\u003ein vitro\u003c/em\u003e experiments (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea) [\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e]. Second, the structure containing the two highlighted benzene rings is monobenzyl phthalate (MBzP), a metabolite of benzyl butyl phthalate. MBzP is classified as a non-genotoxic carcinogen, while an experiment on MBzP noted a significant decrease in global DNA methylation levels. Additionally, MBzP altered the promoter methylation levels of tumor suppressor genes (\u003cem\u003eP16\u003c/em\u003e and \u003cem\u003eTP53\u003c/em\u003e) and proto-oncogenes (\u003cem\u003eBCL2\u003c/em\u003e and \u003cem\u003eCCND1\u003c/em\u003e). As a result, the \u003cem\u003eP16\u003c/em\u003e and \u003cem\u003eTP53\u003c/em\u003e expressions decreased, whereas the expression of \u003cem\u003eBCL2\u003c/em\u003e and \u003cem\u003eCCND1\u003c/em\u003e increased. This indicates a reduction in tumor suppression and the potential to promote cell proliferation and cancer formation (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb) [\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCompounds that exhibit carcinogenicity in the liver, lung, and stomach tissues include 1,1-dichloroethane, which contains an aliphatic halide. The aliphatic halide has been identified as a key substructure in predicting the carcinogenicity of 1,1-dichloroethane. This substructure has been shown to have the potential to induce mutations and promote carcinogenesis through covalent bonds with DNA after being activated into a reactive intermediate (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec) [\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e]. In the liver, lung, and breast tissues, the carcinogenic compound anti-dibenzo(a,l)pyrene-11,12-dihydrodiol-13,14-epoxide contains a polycyclic aromatic hydrocarbons (PAHs) substructure, which can be metabolized into reactive substances such as diolepoxides and radical PAH cations. These reactive metabolites interact with DNA to form DNA adducts, leading to DNA replication errors and the loss of purine bases, ultimately transforming affected genes into oncogenes. Additionally, PAH metabolites, such as quinones, generate reactive oxygen species (ROS), further contributing to DNA damage (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ed) [\u003cspan additionalcitationids=\"CR56 CR57 CR58\" citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e]. Another compound, 2-amino-5-(5-nitro-2-furyl)-1,3,4-oxadiazole, exhibits carcinogenic effects in the lung, stomach, and breast tissues. This compound contains a highlighted furan substructure known to possess potent carcinogenic properties. Nitrofuran also includes this furan substructure and has been identified as a carcinogenic compound. As a result of the carcinogenic and mutagenic potential, these compounds are banned in food-producing animals in the United States, the EU, Australia, and several other countries (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ee) [\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e, \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e]. Lastly, the carcinogenic compound epichlorohydrin is highlighted the epoxide substructure in the stomach. Various epoxides and epoxide-forming chemicals have been proven to be carcinogenic. Epoxides act as strong alkylating agents and can form covalent bonds with DNA, potentially causing mutations and carcinogenesis. Epoxides can also react with biological nucleophiles \u003cem\u003ein vivo\u003c/em\u003e (e.g., protein residues, DNA/RNA bases), leading to DNA damage (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ef) [\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e, \u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eCompared to previous methods that treat tasks independently, this proposed multi-task model learns more abundant representations by reflecting shared features across task combinations. Overall, the multi-task model demonstrated outstanding performance in all tasks, making it a powerful tool for handling various tissue-specific tasks. These results show the potential for the multi-task model to be widely applied to numerous types of biological data, demonstrating the utility of multi-task learning. Furthermore, this study analyzed the specific substructures of carcinogenic compounds that the model prioritized for prediction using attention weights, which represent the learned importance of each node in constructing the final graph representation. As a result, structures reflecting different carcinogenic mechanisms, such as chlorobenzene, aliphatic halide, PAHs, epoxide, furan, and MBzP, were identified as key substructures for carcinogenicity prediction. This indicates that the model captures various carcinogenic mechanisms, including DNA damage and epigenetic regulation.\u003c/p\u003e \u003cp\u003eThe model developed in this study demonstrates high predictive performance and explainability in all tissues, but there are some limitations and improvements to consider. The first issue is the restricted availability of databases, as most databases only record the general carcinogenicity of a compound. Even when using different databases, numerous duplicate compounds exist, which results in a lack of information on tissue-specific carcinogenicity. This limitation in training data may reduce the ability to provide tissue-specific predictions for new compounds. The second is a limitation in multi-task learning, which is generally considered effective for learning with limited data; however, a previous study has shown that multi-task learning requires a dataset large enough to train the minimum number of features [\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e]. However, due to the lack of carcinogen datasets, there is not enough data on compounds that cause carcinogenicity by tissue, making it difficult to make predictions for additional tissues. Therefore, we plan to conduct further research once a sufficient dataset is built, as the continued accumulation of experimental data is expected to enable more accurate and detailed tissue-specific carcinogenicity analyses. The last limitation is that the model proposed in this study does not consider factors important for carcinogenicity and toxicity assessment, such as route of administration, dose, and species specificity of compounds. This is due to the lack of structured databases and differences in toxicity evaluation criteria across databases. Therefore, future studies may need to establish structured databases or standardize those existing ones.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThe continuous increase in cancer patients due to exposure to various carcinogenic compounds has highlighted the need for carcinogenicity prediction research. The multi-task learning model proposed in this study improved performance by reflecting the common features among task combinations during the learning process. This enables more accurate tissue-specific carcinogenicity predictions by effectively learning molecular structures and interactions. Moreover, analyzing the substructures of carcinogenic molecules allowed the identification of tissue-specific carcinogenic mechanisms and carcinogenic patterns that could be common to multiple tissues. These findings may contribute to identifying carcinogenic patterns of concern in populations exposed to carcinogens and provide data for more accurate carcinogenicity predictions. In conclusion, this study can be an important tool to reduce time and costs in evaluating and regulating carcinogenic substances, and it is expected to significantly contribute to developing personalized medicine research and cancer prevention strategies.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated and/or analyzed during the current study are available in the GitHub repository: https://github.com/bmil-jnu/TS-Carcinogenicity.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare the existence of no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was supported by a grant (RS-2024-00332003, RS-2025-02215961) from Ministry of Food and Drug Safety, National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2023-00217317) and Innovative Human Resource Development for Local Intellectualization program through the Institute of Information \u0026amp; Communications Technology Planning \u0026amp; Evaluation(IITP) grant funded by the Korea government(MSIT)(IITP-2024-RS-2022-00156287).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor\u0026rsquo;s contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eY.S. and S.Y. contributed to the research methodology, model architecture design, result analysis, and writing of the original draft. Y.S. performed data curation, model implementation and training, and visualization. M.K. and S.Y. were responsible for project administration, manuscript review, and editing. All authors reviewed and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eJokhadze N, Das A, Dizon DS (2024). Global cancer statistics: A healthy population relies on population health. CA: A Cancer Journal for Clinicians, 74(3).\u003c/li\u003e\n\u003cli\u003eReif AE (1981). The causes of cancer: While some cancers are genetically fated to appear, most have now been traced to environmental factors. American Scientist, 69(4):437-447.\u003c/li\u003e\n\u003cli\u003eBelpomme D, Irigaray P, Hardell L, Clapp R, Montagnier L, Epstein S, Sasco AJ (2007). The multitude and diversity of environmental carcinogens. Environmental research, 105(3):414-429.\u003c/li\u003e\n\u003cli\u003eBlackadar CB (2016). Historical review of the causes of cancer. World journal of clinical oncology, 7(1):54.\u003c/li\u003e\n\u003cli\u003eBourcier T, McGovern T, Stavitskaya L, Kruhlak N, Jacobson-Kram D (2015). Improving prediction of carcinogenicity to reduce, refine, and replace the use of experimental animals. Journal of the American Association for Laboratory Animal Science, 54(2):163-169.\u003c/li\u003e\n\u003cli\u003eFelter SP, Bhat VS, Botham PA, Bussard DA, Casey W, Hayes AW, Hilton GM, Magurany KA, Sauer UG, Ohanian EV (2021). Assessing chemical carcinogenicity: hazard identification, classification, and risk assessment. Insight from a toxicology forum state-of-the-science workshop. Critical Reviews in Toxicology, 51(8):653-694.\u003c/li\u003e\n\u003cli\u003eBenigni R (2012). Alternatives to the carcinogenicity bioassay for toxicity prediction: are we there yet? Expert Opinion on Drug Metabolism \u0026amp; Toxicology, 8(4):407-417.\u003c/li\u003e\n\u003cli\u003eKatt ME, Placone AL, Wong AD, Xu ZS, Searson PC (2016). In vitro tumor models: advantages, disadvantages, variables, and selecting the right platform. Frontiers in bioengineering and biotechnology, 4:12.\u003c/li\u003e\n\u003cli\u003eSaeidnia S, Manayi A, Abdollahi M (2015). From in vitro experiments to in vivo and clinical studies; pros and cons. Current drug discovery technologies, 12(4):218-224.\u003c/li\u003e\n\u003cli\u003eFr\u0026ouml;hlich E, Salar-Behzadi S (2014). Toxicological assessment of inhaled nanoparticles: role of in vivo, ex vivo, in vitro, and in silico studies. International journal of molecular sciences, 15(3):4795-4822.\u003c/li\u003e\n\u003cli\u003eCheng F, Li W, Liu G, Tang Y (2013). In silico ADMET prediction: recent advances, current challenges and future trends. Current topics in medicinal chemistry, 13(11):1273-1289.\u003c/li\u003e\n\u003cli\u003eZhang L, Ai H, Chen W, Yin Z, Hu H, Zhu J, Zhao J, Zhao Q, Liu H (2017). CarcinoPred-EL: Novel models for predicting the carcinogenicity of chemicals using molecular fingerprints and ensemble learning methods. Scientific reports, 7(1):2118.\u003c/li\u003e\n\u003cli\u003eWang Y-W, Huang L, Jiang S-W, Li K, Zou J, Yang S-Y (2020). CapsCarcino: A novel sparse data deep learning tool for predicting carcinogens. Food and Chemical Toxicology, 135:110921.\u003c/li\u003e\n\u003cli\u003eFradkin P, Young A, Atanackovic L, Frey B, Lee LJ, Wang B (2022). A graph neural network approach for molecule carcinogenicity prediction. Bioinformatics, 38(Supplement_1):i84-i91.\u003c/li\u003e\n\u003cli\u003eZhang H, Yi H, Hao Y, Zhao L, Pan W, Xue Q, Liu X, Fu J, Zhang A (2024). Deciphering exogenous chemical carcinogenicity through interpretable deep learning: A novel approach for evaluating atmospheric pollutant hazards. Journal of Hazardous Materials, 465:133092.\u003c/li\u003e\n\u003cli\u003eChen Z, Zhang L, Sun J, Meng R, Yin S, Zhao Q (2023). DCAMCP: A deep learning model based on capsule network and attention mechanism for molecular carcinogenicity prediction. Journal of cellular and molecular medicine, 27(20):3117-3126.\u003c/li\u003e\n\u003cli\u003eTanabe K, Kurita T, Nishida K, Lučić B, Amić D, Suzuki T (2013). Improvement of carcinogenicity prediction performances based on sensitivity analysis in variable selection of SVM models. SAR and QSAR in Environmental Research, 24(7):565-580.\u003c/li\u003e\n\u003cli\u003eLi T, Tong W, Roberts R, Liu Z, Thakkar S (2021). DeepCarc: Deep learning-powered carcinogenicity prediction using model-level representation. Frontiers in artificial intelligence, 4:757780.\u003c/li\u003e\n\u003cli\u003eMoorthy NHN, Kumar S, Poongavanam V (2017). Classification of carcinogenic and mutagenic properties using machine learning method. Computational Toxicology, 3:33-43.\u003c/li\u003e\n\u003cli\u003eWu Y, Wang G (2018). Machine learning based toxicity prediction: from chemical structural description to transcriptome analysis. International journal of molecular sciences, 19(8):2358.\u003c/li\u003e\n\u003cli\u003eYang K, Swanson K, Jin W, Coley C, Eiden P, Gao H, Guzman-Perez A, Hopper T, Kelley B, Mathea M (2019). Analyzing learned molecular representations for property prediction. Journal of chemical information and modeling, 59(8):3370-3388.\u003c/li\u003e\n\u003cli\u003eLiu K, Sun X, Jia L, Ma J, Xing H, Wu J, Gao H, Sun Y, Boulnois F, Fan J (2019). Chemi-net: a molecular graph convolutional network for accurate drug property prediction. International journal of molecular sciences, 20(14):3389.\u003c/li\u003e\n\u003cli\u003eZitnik M, Agrawal M, Leskovec J (2018). Modeling polypharmacy side effects with graph convolutional networks. Bioinformatics, 34(13):i457-i466.\u003c/li\u003e\n\u003cli\u003eSun M, Zhao S, Gilvary C, Elemento O, Zhou J, Wang F (2019). Graph convolutional networks for computational drug development and discovery. Briefings in Bioinformatics, 21(3):919-935.10.1093/bib/bbz042.\u003c/li\u003e\n\u003cli\u003eSacha M, Błaz M, Byrski P, Dabrowski-Tumanski P, Chrominski M, Loska R, Włodarczyk-Pruszynski P, Jastrzebski S (2021). Molecule edit graph attention network: modeling chemical reactions as sequences of graph edits. Journal of Chemical Information and Modeling, 61(7):3273-3284.\u003c/li\u003e\n\u003cli\u003eLe T, No\u0026eacute; F, Clevert D-A (2022). Equivariant graph attention networks for molecular property prediction. arXiv preprint arXiv:2202.09891.\u003c/li\u003e\n\u003cli\u003eZhang S, Tong H, Xu J, Maciejewski R (2019). Graph convolutional networks: a comprehensive review. Computational Social Networks, 6(1):1-23.\u003c/li\u003e\n\u003cli\u003eVrahatis AG, Lazaros K, Kotsiantis S (2024). Graph attention networks: a comprehensive review of methods and applications. Future Internet, 16(9):318.\u003c/li\u003e\n\u003cli\u003eVeličković P, Cucurull G, Casanova A, Romero A, Lio P, Bengio Y (2017). Graph attention networks. arXiv preprint arXiv:1710.10903.\u003c/li\u003e\n\u003cli\u003eLavecchia A (2024). Advancing drug discovery with deep attention neural networks. Drug Discovery Today:104067.\u003c/li\u003e\n\u003cli\u003eArya A, Arya S, Arya M (2011). Chemical carcinogen and cancer risk: an overview. J Chem Pharm Res, 3(5):621-631.\u003c/li\u003e\n\u003cli\u003eHaigis KM, Cichowski K, Elledge SJ (2019). Tissue-specificity in cancer: The rule, not the exception. Science, 363(6432):1150-1151.\u003c/li\u003e\n\u003cli\u003eSchneider G, Schmidt-Supprian M, Rad R, Saur D (2017). Tissue-specific tumorigenesis: context matters. Nature Reviews Cancer, 17(4):239-253.\u003c/li\u003e\n\u003cli\u003eJain S, Siramshetty VB, Alves VM, Muratov EN, Kleinstreuer N, Tropsha A, Nicklaus MC, Simeonov A, Zakharov AV (2021). Large-scale modeling of multispecies acute toxicity end points using consensus of multitask deep learning methods. Journal of chemical information and modeling, 61(2):653-663.\u003c/li\u003e\n\u003cli\u003eXu M, Yang H, Liu G, Tang Y, Li W (2022). In silico prediction of chemical aquatic toxicity by multiple machine learning and deep learning approaches. Journal of Applied Toxicology, 42(11):1766-1776.\u003c/li\u003e\n\u003cli\u003eWang Y, Wang B, Jiang J, Guo J, Lai J, Lian X-Y, Wu J (2021). Multitask CapsNet: an imbalanced data deep learning method for predicting toxicants. ACS omega, 6(40):26545-26555.\u003c/li\u003e\n\u003cli\u003eLim S, Lee YO, editors. Predicting chemical properties using self-attention multi-task learning based on SMILES representation. 2020 25th International Conference on Pattern Recognition (ICPR); 2021: IEEE.\u003c/li\u003e\n\u003cli\u003eZhang Y, Yang Q (2021). A survey on multi-task learning. IEEE transactions on knowledge and data engineering, 34(12):5586-5609.\u003c/li\u003e\n\u003cli\u003eCaruana R (1997). Multitask learning. Machine learning, 28:41-75.\u003c/li\u003e\n\u003cli\u003eCarcinogenic Potency Database. Accessed 09 Aug 2024.\u003c/li\u003e\n\u003cli\u003eChemical Carcinogenesis Research Information System (CCRIS). Accessed 09 Aug 2024.\u003c/li\u003e\n\u003cli\u003eEPA Integrated Risk Information System (IRIS). Accessed 09 Aug 2024.\u003c/li\u003e\n\u003cli\u003eLim E, Pon A, Djoumbou Y, Knox C, Shrivastava S, Guo AC, Neveu V, Wishart DS (2010). T3DB: a comprehensively annotated database of common toxins and their targets. Nucleic acids research, 38(suppl_1):D781-D786. Accessed 09 Aug 2024.\u003c/li\u003e\n\u003cli\u003eOrganization WH (2019). International agency for research on cancer.\u003c/li\u003e\n\u003cli\u003eZhong M, Nie X, Yan A, Yuan Q (2013). Carcinogenicity prediction of noncongeneric chemicals by a support vector machine. Chemical Research in Toxicology, 26(5):741-749.\u003c/li\u003e\n\u003cli\u003eBray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A (2024). Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians, 74(3):229-263.\u003c/li\u003e\n\u003cli\u003eLandrum G (2013). Rdkit documentation. Release, 1(1-79):4.\u003c/li\u003e\n\u003cli\u003eLi Y, Tarlow D, Brockschmidt M, Zemel R (2015). Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493.\u003c/li\u003e\n\u003cli\u003eRoss T-Y, Doll\u0026aacute;r G, editors. Focal loss for dense object detection. proceedings of the IEEE conference on computer vision and pattern recognition; 2017.\u003c/li\u003e\n\u003cli\u003eYu T, Kumar S, Gupta A, Levine S, Hausman K, Finn C (2020). Gradient surgery for multi-task learning. Advances in neural information processing systems, 33:5824-5836.\u003c/li\u003e\n\u003cli\u003eLi Q, Cui Y, Wang Z, Li Y, Yang H (2024). Toxicity assessment of dioxins and their transformation by-products from inferred degradation pathways. Science of The Total Environment, 937:173416.\u003c/li\u003e\n\u003cli\u003eGRILLI S, ARFELLINI G, COLACCI A, MAZZULLO M, PRODI G (1985). In vivo and in vitro covalent binding of chlorobenzene to nucleic acids. Japanese Journal of Cancer Research GANN, 76(8):745-751.\u003c/li\u003e\n\u003cli\u003eSicińska P, Jabłońska E, Bukowska B, Balcerczyk A, Reszka E (2022). The selected epigenetic effects of phthalates: DBP, BBP and their metabolites: MBP, MBzP on human peripheral blood mononuclear cells (In Vitro). Toxicology in Vitro, 82:105369.\u003c/li\u003e\n\u003cli\u003eDiRenzo A, Gandolfi A, Sipes I (1982). Microsomal bioactivation and covalent binding of aliphatic halides to DNA. Toxicology Letters, 11(3-4):243-252.\u003c/li\u003e\n\u003cli\u003eXue W, Warshawsky D (2005). Metabolic activation of polycyclic and heterocyclic aromatic hydrocarbons and DNA damage: a review. Toxicology and applied pharmacology, 206(1):73-93.\u003c/li\u003e\n\u003cli\u003eShimada T, Fujii‐Kuriyama Y (2004). Metabolic activation of polycyclic aromatic hydrocarbons to carcinogens by cytochromes P450 1A1 and1B1. Cancer science, 95(1):1-6.\u003c/li\u003e\n\u003cli\u003eBaird WM, Hooven LA, Mahadevan B (2005). Carcinogenic polycyclic aromatic hydrocarbon‐DNA adducts and mechanism of action. Environmental and molecular mutagenesis, 45(2‐3):106-114.\u003c/li\u003e\n\u003cli\u003eHenkler F, Stolpmann K, Luch A (2012). Exposure to polycyclic aromatic hydrocarbons: bulky DNA adducts and cellular responses. Molecular, Clinical and Environmental Toxicology: Volume 3: Environmental Toxicology:107-131.\u003c/li\u003e\n\u003cli\u003eDipple A. Polycyclic aromatic hydrocarbon carcinogenesis: an introduction. ACS Publications; 1985.\u003c/li\u003e\n\u003cli\u003eMcCalla D (1983). Mutagenicity of nitrofuran derivatives. Environmental mutagenesis, 5(5):745-765.\u003c/li\u003e\n\u003cli\u003eSingh A, Smith DJ (2025). Disposition of Orally Administered [14C]-Nitrofurazone in Broiler Chickens. Journal of Agricultural and Food Chemistry.\u003c/li\u003e\n\u003cli\u003eMelnick RL (2002). Carcinogenicity and mechanistic insights on the behavior of epoxides and epoxide‐forming chemicals. Annals of the New York Academy of Sciences, 982(1):177-189.\u003c/li\u003e\n\u003cli\u003eKostal J, Voutchkova-Kostal A, Weeks B, Zimmerman JB, Anastas PT (2012). A free energy approach to the prediction of olefin and epoxide mutagenicity and carcinogenicity. Chemical research in toxicology, 25(12):2780-2787.\u003c/li\u003e\n\u003cli\u003eStandley T, Zamir A, Chen D, Guibas L, Malik J, Savarese S, editors. Which tasks should be learned together in multi-task learning? International conference on machine learning; 2020: PMLR.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"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":"cancer, carcinogenicity prediction, multi-task learning, graph attention network, attention mechanism, tissue specificity","lastPublishedDoi":"10.21203/rs.3.rs-6518252/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6518252/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eCancer is caused by the uncontrolled growth and division of abnormal cells. In industrialized societies, chemical exposure is one of the leading causes of cancer. Indeed, since certain compounds can induce cancer by damaging genes or affecting cellular metabolism, studying carcinogens is essential. However, previous studies have not considered that compounds may promote different tissue-specific carcinogenicity. Therefore, this study developed a multi-task learning framework to predict tissue-specific carcinogenicity in the liver, lung, stomach, and breast tissues. This framework consisted of a shared layer to extract common features and task-specific layers to perform task-specific predictions. The shared layer contains a graph attention network (GAT) layer to make atom representations that reflect the importance of neighboring atoms and parallel fully connected layers designed for each task combination. These shared representations are then passed to task-specific layers to predict tissue-specific carcinogenicity. This entire training process was conducted through stepwise learning, whereby the model was trained in the first step using partially labeled data for tissues, and the initial weights were determined during this process. The second step trained the model using fully labeled data for all tissues, allowing the model to perform the final training for carcinogenicity prediction. The results demonstrated that the proposed multi-task model achieved superior performance overall. The best performance was observed in the stomach task (AUROC: 0.825; AUPR: 0.867), outperforming single-task models (AUROC: 0.800; AUPR: 0.840) and previous studies (AUROC: 0.743–0.791; AUPR 0.788–0.827). We further analyzed molecules with high predicted carcinogenicity in each tissue and identified critical substructures for the prediction using the attention mechanism. This research can contribute to predicting the tissue-specific carcinogenicity of candidate chemicals in the early stages of drug development, thereby reducing research costs and time.\u003c/p\u003e","manuscriptTitle":"Tissue-Specific Carcinogenicity Prediction Using Multi-Task Learning on Attention-based Graph Neural Networks","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-05 16:30:46","doi":"10.21203/rs.3.rs-6518252/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":"33625310-1131-466d-ae0d-d63637614ecf","owner":[],"postedDate":"May 5th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-10-03T07:09:05+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-05 16:30:46","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6518252","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6518252","identity":"rs-6518252","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}