PROTAC-Databank: The Present Largest Integrated Resource of PROTACs, Enabling the Enhanced DeepPROTACs 2.0 for Degradation Prediction

PROTAC-Databank: The Present Largest Integrated Resource of PROTACs, Enabling the Enhanced DeepPROTACs 2.0 for Degradation Prediction | 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 Article PROTAC-Databank: The Present Largest Integrated Resource of PROTACs, Enabling the Enhanced DeepPROTACs 2.0 for Degradation Prediction Fang Bai, Siyuan Tian, Yilin Tang, Fenglei Li, Zhaoxuan Li This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5021266/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 Proteolysis-targeting chimera(PROTAC), is an emerging therapeutic modality with the potential to treat disease-driven proteins that emerged in the new century . Since the first PROTAC drug, ARV-110, entered Phase II clinical trials in 2019, research in this field has surged. The growing body of data has created opportunities for developing artificial intelligence (AI) methods in PROTAC design. In this study, we present the largest PROTAC database to date, named as PROTAC-Databank. Each molecule in the database is labeled with manually reviewed and uniformly formatted degradation efficiency data and protein-ligand complex structure information, providing a valuable resource for AI-driven PROTAC modeling or design. Additionally, we have enhanced the prediction accuracy of the DeepPROTACs, a widely used tool for predicting PROTAC degradation efficiency developed by our group two years ago. The update version, DeepPROTACs 2.0, which is developed based on a complete new deep learning framework. Compared with the previous version, DeepPROTACs 2.0 shows a significant improvement in prediction accuracy, increasing from 77.15% to 83.45%. The introduction of PROTAC-Databank and DeepPROTACs 2.0 offers optimized tools for PROTAC design, streamlining the screening process and reducing both time and costs. Biological sciences/Computational biology and bioinformatics/Databases Biological sciences/Computational biology and bioinformatics/Computational models Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction In recent decades, various cancer therapy strategies have been developed inhibit the activity of oncogene proteins. Small molecular inhibitors have been used in cancer treatment by blocking the active sites of target proteins, thereby disabling oncogene proteins. While these inhibitors have shown promise, their further development limited by issues such as the accumulation of blocked target proteins, drug resistance, and the undruggable nature of some protein structure. 1 Antibodies, though highly specific to their target proteins, suffer from low membrane permeability, which limits their effectiveness inside cancer cells. 2 Another approach involves the use of siRNAs to directly inhibit gene expression, but off-target effects and challenges in drug delivery have hindered their broader application. 3 Therefore, new molecule modalities are required in the current drug discovery landscape. 4 Proteolysis targeting chimeras (PROTACs) represent a promising strategy designed to target and accelerate the degradation of specific proteins. 4 , 5 PROTACs are a class of small molecules with unique structures that utilizes the ubiquitin-protease system (UPS) to degrade target proteins. 5 , 6 , 7 , 8 A typical PROTAC molecule comprises three components: a target protein ligand (also known as the warhead), a linker, and an E3 ubiquitin ligase recruiting ligand (E3 ligand). These three parts are covalently boned to form the complete molecule. 9 The ligands at both ends of the PROTAC bind to the target proteins and the E3 ligase respectively, bringing the target protein into closely proximity with the E3 ubiquitin complex. 7 This closely proximity facilitates the transfer of ubiquitin from the E2 enzyme to a lysine residue on the surface of the target protein. Subsequently, the polyubiquitinated target protein is recognized by 26S proteasome and degraded into peptide fragments. 10 In essence, PROTACs hijack the UPS to specifically degrade target proteins, thereby reducing intracellular levels of these proteins. With the advantages of small molecules, PROTACs offer easier drug delivery compared to antibodies and siRNA. Unlike traditional small molecule inhibitors, which can lead to the accumulation of inactivated protein within cells, PROTACs completely eliminate target proteins, preventing such accumulation. 9 , 10 Moreover, recently, PROTACs have also been designed to treat diseases beyond cancer, such as autoimmune disorders, 11 virus infection 12 . PROTACs are well-suited for targeting “undruggable” targets, because they don’t require strong binding affinities between themselves and target proteins, making them highly versatile. Since Craig M. Crews developed the first PROTAC in 2001, the field has seen tremendous development. 4 , 13 In recent years, several PROTAC candidate drugs has entered clinical trials, including ARV-110, an androgen receptor (AR) target PROTAC for treating metastatic castration-resistant prostate cancer. 14 , 15 The success of PROTAC research and development has significantly driven the growth of related literature. Since 2018, the volume of literature on PROTACs has grown rapidly, with newly developed PROTACs being reported at an unprecedented rate. 5 , 16 As of November 2023, PubMed lists 2,007 PROTAC-related publications, with1,880 of them published after 2018. The explosion of data presents opportunities to apply artificial intelligence (AI) in designing methods development for PROTACs. The advancements in high-performance computing and multi-omics data have facilitated the practical application of AI application in any field lies in the availability of high-quality datasets. 17 , 18 , 19 Fortunately, numerous high-quality biological information databases, such as, ChEMBL 20 , DrugBank 21 , PDB 22 , and PubChem 23 , allow AI techniques to be applied to multiple aspects of drug discovery, including protein structure prediction, hit identification 24 , drug target discovery 25 , bioavailability prediction 26 , and de novo drug design 27 etc. Currently, AI methods have begun to be employed in the rational design of PROTACs. 28 , 29 To facilitate rational PROTAC design, it is essential to predict their properties accurately, necessitating the establishment of a database specifically for PROTACs to support AI models. PROTAC-DB 30 , 31 is one such database, its updated version PROTAC-DB 2.0 currently archives 3,270 PROTACs. 31 It includes biological activity data such as DC 50 (half minimum degradation concentration)data, cellular activity data, western blotting data, and binding affinity. Additionally, the database contains ternary structural data for target-protac-E3 ligase is part of the database. However, PROTAC-DB provides limited ternary structure for target-protac-E3ligase complexes, with only 18 crystal structures and 664 predicted structures available, leaving the binding patterns of warhead-target and E3 ligand-E3 ligase complexes missing for 80% of PROTACs. Moreover, the format of PROTAC-DB is not directly usable in AI models due to inconsistencies in PROTAC degradation standards. Our literature review also revealed that a significant number of PROTACs are not included in PROTAC-DB. In this paper, we present a newly established PROTAC database containing target-warhead and E3 ligase-E3 ligand complexes for each PROTAC molecule. Our data were sourced from both PRORAC-DB and extensive literature research, with degradation labels assigned to each target-PROTAC-E3 Ligase pair. And degradation labels were tagged to each Target-PROTAC-E3ligase pairs. The data can be directly used as dataset for AI models such as DeepPROTACs 32 for PROTAC degradation efficiency prediction. Our database provided the pocket sites information of target protein and E3 ligase. Besides, the property comparation between traditional small molecular drugs and PROTAC are also displayed in this paper. Such analysis may help further PROTAC design and modification. Besides the construction of database, we also developed a second version of DeepPROTACs by designing a new deep learning framework. 32 . DeepPROTACs was previously developed by our group. It can predicted the PROTAC degradation efficiency using input files containing the PROTAC molecule, and target-warhead complex structure, and E3 ligase-ligand complex structure. Building on the work of DeepPROTACs 1.0, we refined the architecture of the model by using Graph Attention Networks (GANs) and improved accuracy rate using the original data set. We also incorporate new datasets from the PROTAC-Databank, further enhancing prediction efficiency. With updates to both the model architecture and dataset, we have released DeepPROTACs 2.0 in this manuscript. A webserver is also available for free use via the link ( https://bailab.siais.shanghaitech.edu.cn/services/deepprotac-db ). Method Data source of PROTAC-Databank The raw data of our database were sourced from PROTAC-DB 30 , 31 and other literature-reported PROTACs. We obtained information such as PROTAC SMILES 33 , warhead (target ligand) SMILES, E3 ligand SMILES, Target protein type, E3 ligase type from PROTAC-DB 2.0 31 . We then we queried “PROTAC” in PubMed to collect literature not included in PROTAC-DB 2.0. Information such as PROTAC SMILES, warhead SMILES, E3 ligand SMILES, target protein, E3 ligase, DC 50 , D max , etc, were extracted from literatures and extended in database. (Fig. 1 .A) Data processing of PROTAC-Databank Protein Structure Collection . After collecting PROTAC raw data, we generated complex binding structures between warhead and targets, as well as complex models between E3 ligase and E3 ligand by molecular docking simulations. (Fig. 1 .A) We first listed the types of target proteins and E3 ligases from the raw data, then searched the PDB database for experimentally determined complex structures corresponding to these proteins. 22 For each protein structure, we identified the pocket site that binds with the warhead or E3 ligand, selecting the lowest resolution structure containing models with the same or similar ligands as the PROTAC warhead(or E3 ligand). The complex structure of warheads with target proteins or ligands with E3 ligases were generated by molecule docking simulations, confirming the pocket sites and reference ligand binding poses for each protein-ligand pair. Although most of protein structures for these PROTAC models were found in PDB, some were not available as experimental structures. For these proteins, we used predicted structures by AlphaFold2 34, 35 (Protein types using AlphaFold2 predicted models: GSK3A, TYR). Once the models were ready, ligands were clustered into different groups based on their target protein or E3 ligase. We also converted SMILES and E3 ligand SMILES into 3D structures in PDB format using the RDKit 36 package in Python. Molecule Docking . To obtain the binding complex of protein-ligand, molecular docking simulations were performed. Firstly, Protein Preparation Workflow module of Schrodinger 2023 37 was used to process the protein models from the above obtained databases, including filling the missing side chain, optimizing H-bond assignment, minimizing structure in OPLS_4 force field 38 and deleting water in models. To produce the low-energy 3D conformations of collected ligands, we process molecule PDB files by LigPrep module of Schrodinger 2023. 39 The reference ligands for docking small molecules into the pocket were selected from downloaded PDB complex models. Ligands for the same protein were aligned to the reference pose of the corresponding reference ligand using Schrodinger 2023's Ligand Alignment module. Subsequently, the aligned small molecules were combined with prepared protein models individually, generating protein-ligand complex models for each pair of target-warhead and E3 ligase-E3 ligand. Schrodinger 2023's Prime 40 , 41 , 42 minimize were employed to optimize structure and minimize energy of models. Through this process, we generated both target-warhead complex models and E3 ligase-E3 ligand models. (Fig. 1 .A) Classification of degradation efficiency labels: A unified standard for PROTAC degradation 43 efficiency was established for each target-PROTAC-E3 ligase pair in the database. DC 50 (half maximal degradation concentration) is one of the measurements for evaluating the degradation efficiency. For the “Degradation efficacy” label, pairs with DC 50 < 10nM are labelled “excellent”, 10nM ≤ DC 50 < 100nM are labelled “good”, 100nM ≤ DC 50 < 500nM are labelled “moderate”, 500nM ≤ DC 50 < 1000nM are labelled “not good”, DC 50 ≥ 1000nM are labelled “poor”. D max (maximum degradation rate ) in another factor that is induced to measure the PROTAC degradation level.Two criteria have been established for labeling PROTAC degradation effectiveness. For “Label 1”, “Criterion 1” classifies pairs with a DC 50 ≤ 100 nM and a D max ≥ 80 as "Good," while pairs that do not meet these thresholds are categorized as “Poor.” For “Label 2”, “Criterion 2” deems pairs with a DC 50 ≤ 1000 nM and a D max ≥ 70 as "Good," with all other configurations considered "Poor". When the DC 50 data was missing for some PROTAC pairs, western blot analysis graph of degradation experiments in original literature would be viewed to estimate the approximate DC 50 and D max values to label the degradation efficiency. (Fig. 1 .A) Data analysis of PROTAC-Databank The RDkit package was used to calculate molecular properties of PROTACs in PROTAC-Databank and drug molecules in DrugBank 21 . The following properties were computed using RDkit: linker heavy atom numbers of linkers, rotatable bonds of linkers of PROTACs; H-bond donor counts, H-bond acceptor counts, partition coefficient (logP), topological polar surface area (TPSA), molecular weight, heavy atom counts, rotatable atom counts of both PROTACs molecule in PROTAC-Databank, and drug molecules in DrugBank. Origin 2020 44 was used to analysis and visualize those statistical results. Updating from DeepPROTACs to DeepPROTACs 2.0 Previously, the DeepPROTACs model was trained on a dataset of 2,832 labeled PROTACs. Despite its success in predicting PROTAC degradation rates, the dataset size was relatively small for deep learning applications. To address this, we expanded the dataset to include 4,142 labeled PROTACs, which were used to train the updated DeepPROTACs 2.0 model. In the data processing pipeline, based on the original DeepPROTACs approach, we extracted the binding pockets by selecting residues within 5Å around the binding ligand using PyMol software. The extracted structures were then converted to Mol2files 45 , 46 using OpenBabel 47 . These files contain information about the types, coordinates, and bonding of the atoms, allowing for the simple reconstruction of molecular structures. Since atoms and their bonds in molecules can naturally be represented as nodes and edges in graphs, we used a graph data structure to represent binding pockets and ligands. Graph representations were built based on the Mol2 files using an adjacency matrix. In these graphs, atoms were represented as nodes, with atomic types encoded in a dictionary (C, N, O, S, and others as 0, 1, 2, 3, 4 respectively). Covalent bonds were represented as edges, with adjacency matrix positions labeled 1 for bonded atoms and 0 otherwise. Given the relatively limited number of covalent bonds between the molecules, the adjacency matrix of the molecules is relatively sparse, and at the same time the amount of operations brought by the adjacency matrix is relatively high. For this reason, we adopted a simple structure of edge list to represent the bonding information between atoms. If there was a covalent bond between atom n and atom m, then two dot pairs [n,m] and [m,n] are added to the list. In the graphs of ligand, we also used atoms to represent nodes and bonds to represent edges. However, the number of atom types in the ligand increases and a larger dictionary is needed to represent this mapping relationship, with C, N, O, S, F, Cl, Br, I, P, and other atoms denoted by 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, respectively. In addition, the different types of bonds between atoms were also encoded as attribute of the edge nodes in the graph, there are single bond, double bond, triple bond, aromatic bond, and amide bond, each of which was encoded by 1, 2, 3, 4, 5. PROTACs' linkers were encoded in SMILES format and based on a table from the ZINC database 48 , with the most frequent 39 characters encoded from 1 to 39, and the remaining characters encoded as 40. In advancing the DeepPROTACs model, we transitioned from conventional sequence-base models to Graph Convolutional Networks (GCNs) 49 for feature extraction from graph representations of protein pockets. In GCNs, the features of each node are averaged from its neighboring nodes, implying equal contributions from all neighbors. However, the functionality and influence of a molecular group is shaped by its neighbors, with varying impacts on the compound’s properties, meaning that neighboring nodes should contribute differently to the feature computation. This revealed a limitation in GCNs, where treating all neighboring nodes as equally contributive fails to capture the differential impacts in molecular contexts. To overcome this, we introduced Graph Attention Networks (GATs) 50 for feature learning within graph representations of protein pockets. GATs uses an attention mechanism that allows individual nodes to dynamically assign different weights to their neighboring nodes during the messae-passing phase. This enables a node to focus on the most significant or relevant neighbors in tis structural or functional context. By incorporating this attention-based approach, GATs provide greater flexibility and precision in capturing the complex relationships between nodes, allowing for a more accurate representation of interactions within the protein pocket. Furthermore, recognizing the importance of atom-to-atom distances in potential non-bonded interactions, we refined our graph representation by including these distances as additional features for each atom. This enhancement not only enriches the graph representation but also optimizes the subsequent graph network convolution calculations, equipping our refined model to better understand and analyze the intricate details of the protein pocket structure and non-bonded molecular interactions. The Architecture of DeepPROTACs 2.0 The architecture of our network, illustrated in Figure.2, is developed entirely using the PyTorch 51 and PyTorch Geometric frameworks 52 . Initially, protein pockets are represented graphically through a series of mapping layers. An embedding layer then increases the dimensionality of the data to 64, preparing it for processing by the subsequent graph neural network (GNN) 53 . The core of the GNN comprises two GAT convolution layers, followed by a max pooling layer, maintaining an output dimensionality of 64. In parallel, the encoding of linker SMILES strings is processed through a separate embedding layer to ensure the linker representation matches the dimensionality of the protein pocket. This encoded linker data is then processed by a Bidirectional Long Short-Term Memory (Bi-LSTM) 54 layer and a fully connected layer, with the output dimensionality also set to 64. Finally, the outputs from both the linker and protein pocket/ligand modules are concatenated and fed into a two-layer Multi-Layer Perceptron (MLP) 55 to generate the final output. Throughout the network, the Rectified Linear Unit (ReLU) 56 function serves as the activation mechanism, enabling non-linear learning capabilities. Results Statistical Analysis of PROTAC-Databank Following the construction of the PROTAC-Databank, a comprehensive statistical analysis was performed. The database contains a total of 3,645 PROTACs, including 612 types of warheads, 145 types of E3 ligands, and 1,535 types of linkers (Fig. 3 ). Regarding degradation labels, 74.1% of the data entries include degradation labels, while the remaining 25.1% lack these labels due to missing DC 50 and D max values in the source data. For entries with degradation labels, protein-ligand complexes of both the Target-warhead and E3 ligase-E3 ligand pairs were constructed, resulting in 4,142 pairs of protein-ligand complex structures. The PROTAC-Databank includes 4,142 complete data sets, each comprising basic information, a PDB file of the Target-warhead complex, a PDB file of the E3 ligase-E3 ligand complex, a SMILES file of the PROTAC, three degradation labels, and the source literature. Some data sets also contain additional information, such as K D values (PROTAC to E3 ligase and PROTAC to Target), IC 50 values from specific experiments, and other relevant data. The PROTAC-Databank data is stored in a CSV file and is accessible online ( https://bailab.siais.shanghaitech.edu.cn/services/deepprotac-db ). The PROTAC IDs from PROTAC-DB 2.0 are consistent with those in the PROTAC-Databank, while entries not from PROTAC-DBs are renumbered with the prefix "b." The PROTAC-Databank exhibits greater diversity than the database PROTAC-DB. It contains 3,645 types of PROTACs, surpassing the 3,270 types found in PROTAC-DB. Furthermore, PROTAC-Databank includes 612 distinct warheads, significantly more than the 365 found in PROTAC-DB. However, both databases show limited diversity in E3 ligands and E3 ligases. PROTAC-Databank includes 145 types of E3 ligands, while PROTAC-DB has 82. The number of linker types is comparable between the two, with 1,535 types in PROTAC-Databk and 1,510 in PROTAC-DB, this also reflects that the previous design strategies for PROTAC molecules were primarily focused on empirically replacing a limited variety of linkers, with very limited innovation in the molecules themselves. Therefore, developing linker generation algorithms based on generative methods, such as the deep learning-based PROTAC linker generation method, i.e., DiffPROTACs 57 , developed by our research group, is particularly important for expanding the diversity of PROTAC linkers and advancing the field. A key distinction between the two databases is that PROTAC-Databank includes more structural information. It contains 4,142 pairs of target-warhead and E3 ligase-E3 ligand complexes, compared to 664 ternary complexes in PROTAC-DB. By incorporating protein-ligand complexes, PROTAC-Databank offers a new dimension for PROTAC design and the development of AI methods related to PROTACs. The database includes 14 types of E3 ligase proteins, with CRBN being the most frequently utilized, representing 65.8% of the PROTAC pairs targeting protein degradation. VHL is the second most common, used in 28.8% of the pairs. Over 90% of the E3 ligases employed by PROTACs in the database are either CRBN or VHL. The distribution of target proteins is more varied compared to E3 ligases (Fig. 4 A). The database contains 337 types of target proteins, with different mutants of a single wild-type protein considered as distinct targets. The Androgen Receptor (AR) is the most common target, representing 5% of all target proteins. Other frequently targeted proteins include the Estrogen Receptor (ER), BRD4, BTK, and CDK9. In the statistical results, 67.5% of the target proteins are individual proteins, each constituting less than 1.7% of the total targets (Fig. 4 A). Linker properties in PROTACs are a crucial factor in determining degradation efficiency. A cluster of PROTAC molecules with identical warheads and E3 ligands can exhibit varying degradation efficiencies due to differences in their linkers. Therefore, we also analyzed the types of linkers in the PROTAC-Databank (Fig. 4 A). Half of the linkers are difficult to classify into specific types due to their complex structures. However, 35.2% of the linkers contain PEG segments, 8.8% are alkanes, and others include piperazine, piperidine, and alkyne segments. The degradation labels in the database are categorized into five groups based on the DC 50 values obtained from PROTAC degradation experiments (Fig. 4 .B). To simplify the labels for AI model input, two classifications “Label 1” and “Label 2” were classified into "Good" or "Bad" categories based on both DC 50 and D max values. The criteria for these classifications differ: In “Criteria 1” for “Label 1”, PROTAC pairs with DC 50 ≤ 100 nM and Dmax ≥ 80 are labeled as "Good," while all other cases are labeled as "Bad". In “Criteria 2” for “Label 2”, pairs with DC 50 ≤ 1000 nM and D max ≥ 70 are labeled as "Good," with the remainder labeled as "Bad." According to the statistical analysis of "Label 1," the number of negative labels in the database exceeds positive labels by approximately 30% (Fig. 4 C). However, with the relaxed criteria in "Label 2," the numbers of "Good" and "Bad" labels are nearly equivalent (Fig. 4 C). Comparison of Properties Between PROTAC Molecules and Tradition Drug Molecules . Due to the unique mechanisms and structures of PROTACs, their properties differ significantly from traditional small-molecule drugs. A comparison was conducted between the molecular properties of PROTACs in the PROTAC-Bank and traditional drug molecules in DrugBank. Additionally, PROTACs were categorized into "Active" and "Inactive" groups based on the "Good" and "Bad" labels from the "Label 1" during the statistical analysis (Fig. 5 ). The analysis revealed significant differences in multiple drug properties. The molecular weight of PROTACs is consistently higher than that of traditional drugs, ranging from 600 to 1300 Da, with a peak at 800–900 Da, while most traditional drugs have a molecular weight below 500 Da. This difference in size leads to distinct variations in other properties as well. PROTAC molecules have more hydrogen bond donors and acceptors, as well as more rotatable bonds, compared to DrugBank molecules (Fig. 5 . A-G). Additionally, the topological polar surface area (TPSA) and logP were calculated to further analyze these differences. TPSA tends to be higher in PROTACs compared to traditional small-molecule drugs. The peak of TPSA distribution for PROTACs is around 200, while for DrugBank compounds, it is below 150. The logarithm of the partition coefficient between octanol and water (logP) indicates the differential solubility of a compound in these solvents. Given that many PROTAC linkers have alkane backbones, their lipophilicity is increased, resulting in a right-shifted logP distribution for PROTACs compared to traditional drug compounds. In summary, the properties of PROTAC molecules differ significantly from those of traditional small molecular drugs in DrugBank, suggesting potential challenges in designing druggable PROTACs and ensuring their effective delivery. Additionally, the distribution of linker heavy atom counts and linker rotatable bond counts in the PROTAC-Databank was analyzed (Fig. 5 . H, I). The heavy atom linker counts showed an overlapping distribution between active and inactive PROTACs, with linkers containing 5 to 9 atoms being the most common in both groups. However, there is a slight difference in the peaks of these distributions: linkers with 5 and 6 atoms are most prevalent in inactive molecules, while linkers with 7 and 8 atoms dominate in active molecules. For the number of rotatable bonds in linkers, inactive molecules tend to have one more rotatable bond on average compared to active molecules. The peak number of rotatable bonds for "active" molecules is 4 to 5, while for "inactive" molecules, it is 5 to 7. This suggests that PROTACs with better degradation efficiency may have relatively more rigid linkers compared to those with poorer efficiency. These statistical results provide insight into the current trends in PROTAC linker design. Over half of the reported PROTAC molecules have linkers with 5 to 9 heavy atoms and 3 to 7 rotatable bonds, regardless of whether the molecules are classified as "active" or "inactive." Prediction Efficiency of DeepPROTACs 2.0 Compared to Multiple Methods. The initial dataset used in the previous DeepPROTACs study comprised 2,832 labeled PROTACs from PROTAC-DB. These were categorized based on their DC ₅₀ and D max values, with PROTACs having DC ₅₀ below 100 nM and D max above 80% labeled as "good," while all others were labeled as "bad." This resulted in 988 "good" and 1,844 "bad" data points. Upon expanding the dataset, the total number of PROTACs increased to 4,140, with 1,474 "good" and 2,666 "bad" data points. In previous work using DeepPROTACs, the model achieved an accuracy of 77.15% and an AUROC of 0.8246 on the 2,832 data points with optimal parameter settings. After expanding the dataset, the same model yielded an improved accuracy of 81.82% and an AUROC of 0.8734 (Table 1 , with the ROC curve presented in Fig. 6 A). This significant enhancement in accuracy following the database expansion demonstrates the effectiveness of our data collection and processing. In our current work, we further optimized the original model by incorporating inter-atomic distances as a feature. Experimental results showed that the new model achieved an accuracy of 79.68% and an AUROC of 0.8493 on the original 2,832 data points. On the expanded dataset of 4,140 data points, the model achieved an accuracy of 83.45% and an AUROC of 0.9001. In both cases, the new model outperformed the previous version. Overall, our findings indicate that, compared to the original DeepPROTACs, our model's accuracy improved by over 8% and the AUROC by over 9%. Table 1 Performance comparison of DeepPROTACs models on different datasets. The table displays average accuracy and AUROC scores for DeepPROTACs 1.0 and DeepPROTACs 2.0 using the PROTAC-DB and PROTAC-Databank datasets. Model Dataset Average Accuracy AUROC DeepPROTACs 1.0 PROTAC-DB 77.15% 0.8246 DeepPROTACs 1.0 PROTAC-Databank 81.82% 0.8734 DeepPROTACs 2.0 PROTAC-DB 79.68% 0.8493 DeepPROTACs 2.0 PROTAC-Databank 83.45% 0.9001 When compared to traditional machine learning models such as Support Vector Machines (SVM) 58 and Random Forests (RF) 59 , which utilized MACCS and Morgan molecular fingerprints for PROTAC ligands and ligases 60 , 61 , our DeepPROTACs 2.0 approach achieved superior accuracy and AUROC (Table 2 , with the ROC curve presented in Fig. 6 B). Table 2 The average accuracy and AUROC scores for DeepPROTACs 2.0, SVM with MACCS and Morgan fingerprints, and RF with MACCS and Morgan fingerprints. Model Fingerprints Average Accuracy AUROC DeepPROTACs 2.0 - 83.45% 0.9001 SVM MACCS 72.54% 0.7333 Morgan 77.78% 0.7989 RF MACCS 78.50% 0.8550 Morgan 78.66% 0.8746 We also assessed the generalization ability of our model. We removed data for specific ligands from the training set and used the remaining data to train the model. After training, the removed data served as a test set to evaluate model predictions. The results showed that, after optimizing the dataset and model, the prediction accuracy for specific target proteins exceeded 90% (Table 3 ). This surpasses the performance reported in the original DeepPROTACs study, indicating that our improved model has enhanced generalization capabilities. To further validate our model's predictive capacity, we used an experimental database from DeepPROTACs, which includes data on 16 PROTACs. While DeepPROTACs accurately predicted 11 out of 16 PROTACs (68.75% accuracy), our current model successfully predicted 12 PROTACs, achieving a 75% accuracy rate. In the expanded dataset of 4,140 PROTACs, there were 1,474 "good" data points and 2,666 "bad" data points. In addition to the original data distribution, we employed both over-sampling and under-sampling techniques, adjusting the data ratios to 1,474:1,474 and 2,666:2,666, respectively. For each sampling method, we independently trained our model four times and conducted predictions, calculating the average accuracy and AUROC for each method. The results, as shown in the figure, indicate that models trained on oversampled data exhibit better performance in both accuracy and AUROC (Table 4 ). This aligns with conclusions drawn from previous DeepPROTACs work. We have also launched a model prediction service, allowing users to input data using the same PROTAC processing methods described in this article to obtain degradation rate predictions. The dataset used for training the model is accessible on our website ( https://bailab.siais.shanghaitech.edu.cn/services/deepprotac-db ), where users can search by PROTAC ID, Target, and E3 ligases. Table 3 Average accuracy and AUROC of DeepPROTACs 2.0 on four different targets(EZH2, eIF4E, STAT, FLT) Target Average Accuracy AUROC EZH2 90.32% 0.96131 eIF4E 92.70% 0.95635 STAT 92.70% 0.94048 FLT 91.07% 0.92653 Table 4 Average accuracy and AUROC of different sample methods Sampling Method Average Accuracy AUROC Under-sampling 78.15% 0.8524 Normal-sampling 83.45% 0.9001 Over-sampling 84.10% 0.9109 Discussion In this study, we developed PROTAC-bank, a comprehensive database of predicted ligand-protein complex structures for existing PROTAC molecules. The database aggregates and processes data from PROTAC-DB and relevant literature, constructing and refining both "warhead-target protein" and "E3 ligase-E3 ligand" structures for each PROTAC. These structures are labeled according to their corresponding PROTAC degradation levels. The statistical analysis of PROTAC-Bank, as presented above, highlights the distribution of target proteins and E3 ligases, and underscores the distinctions between PROTACs and traditional small molecule drugs. Moreover, we utilized PROTAC-Bank to enhance the prediction accuracy of DeepPROTACs by expanding the dataset. Additionally, we optimized the AI model underlying DeepPROTACs, leading to a significant improvement in accuracy from 77.15–83.45%. PROTAC-Bank is a unique resource that focuses on protein-ligand structures, providing datasets for target-warhead and E3 ligase-E3 ligand interactions. By linking PROTAC degradation efficiency with these ligand-protein structures, PROTAC-bank offers a significant advantage: the ability to incorporate structural information and ligand-protein interactions into existing training sets as new features. This facilitates the future development and refinement of AI models for designing and evaluating PROTAC molecules. Furthermore, the database serves as a valuable tool for optimizing and updating existing models and can contribute to theoretical research on PROTAC degradation mechanisms and the design of novel PROTACs. The database also reveals the chemical property differences between PROTACs and traditional small molecule drugs. The statistical analysis shows that most PROTACs have a molecular weight exceeding 500 Da, often failing to comply with Lipinski's rule of five. Moreover, when considering molecular properties such as logP, H-bond acceptor and donor counts, and the number of rotatable bonds, the majority of PROTAC molecules do not meet the criteria set by Lipinski’s rules, indicating their limited druggable potential. Addressing the administration of PROTACs is thus a crucial challenge that needs to be addressed. Future research directions should focus on refining PROTACs to optimize their ADMET properties and developing effective strategies for their delivery. The updated DeepPROTACs model has achieved a higher accuracy in predicting PROTAC degradation efficiency, making it a valuable tool for screening and evaluating designed PROTACs before molecule synthesis and biological experimentation. This can significantly reduce both the cost and time associated with PROTAC discovery. Despite the significant contributions of PROTAC-Bank to PROTAC research and AI model enhancement, there is still room for improvement. During data collection, some target proteins lacked experimental or predicted structures, and others were missing pocket site information, leading to their exclusion from PROTAC-Bank. Additionally, with the rapid increase in PROTAC-related literature over the past five years, maintaining an up-to-date PROTAC-bank requires continuous effort. Currently, data collection is done manually, but more efficient methods for updating PROTAC data are needed. A long-term goal is to create a platform that allows researchers to contribute their discovered PROTACs to the database. In conclusion, we have established the PROTAC ligand-protein structure database, PROTAC-Bank, and used it to update and retrain DeepPROTACs, achieving a higher accuracy rate. These advancements will greatly facilitate the design of novel PROTACs and support theoretical research in the field. Declarations Conflicts of interest/Competing interests The authors have no relevant financial or non-financial interests to disclose. Acknowledgement This work was supported by National Key R&D Program of China (Grant IDs: 2022YFC3400501 and 2022YFC3400500), Shanghai Science and Technology Development Fund (22ZR1441400), start-up package from ShanghaiTech University, and Shanghai Frontiers Science Center for Biomacromolecules and Precision Medicine at ShanghaiTech University. We are grateful for the support from HPC Platform of ShanghaiTech University. References Lu J , et al. Hijacking the E3 Ubiquitin Ligase Cereblon to Efficiently Target BRD4. Chemistry & biology 22 , 755-763 (2015). Lyu X , et al. The global landscape of approved antibody therapies. Antibody therapeutics 5 , 233-257 (2022). Alshaer W , et al. siRNA: Mechanism of action, challenges, and therapeutic approaches. European journal of pharmacology 905 , 174178 (2021). Sakamoto KM, Kim KB, Kumagai A, Mercurio F, Crews CM, Deshaies RJ. Protacs: chimeric molecules that target proteins to the Skp1-Cullin-F box complex for ubiquitination and degradation. Proceedings of the National Academy of Sciences of the United States of America 98 , 8554-8559 (2001). Zou Y, Ma D, Wang Y. The PROTAC technology in drug development. Cell biochemistry and function 37 , 21-30 (2019). He M , et al. PROTACs: great opportunities for academia and industry (an update from 2020 to 2021). Signal transduction and targeted therapy 7 , 181 (2022). Neklesa TK, Winkler JD, Crews CM. Targeted protein degradation by PROTACs. Pharmacology & therapeutics 174 , 138-144 (2017). Pettersson M, Crews CM. PROteolysis TArgeting Chimeras (PROTACs) — Past, present and future. Drug Discovery Today: Technologies 31 , 15-27 (2019). Toure M, Crews CM. Small-Molecule PROTACS: New Approaches to Protein Degradation. Angewandte Chemie (International ed in English) 55 , 1966-1973 (2016). Schneekloth AR, Pucheault M, Tae HS, Crews CM. Targeted intracellular protein degradation induced by a small molecule: En route to chemical proteomics. Bioorganic & Medicinal Chemistry Letters 18 , 5904-5908 (2008). Galla MS, Sharma N, Mishra P, Shankaraiah N. Recent insights of PROTAC developments in inflammation-mediated and autoimmune targets: a critical review. RSC medicinal chemistry 15 , 2585-2600 (2024). Alugubelli YR , et al. Discovery of First-in-Class PROTAC Degraders of SARS-CoV-2 Main Protease. bioRxiv , (2023). Li D, Yu D, Li Y, Yang R. A bibliometric analysis of PROTAC from 2001 to 2021. European journal of medicinal chemistry 244 , 114838 (2022). Neklesa T , et al. ARV-110: An oral androgen receptor PROTAC degrader for prostate cancer. Journal of Clinical Oncology 37 , 259-259 (2019). Gao X , et al. Phase 1/2 study of ARV-110, an androgen receptor (AR) PROTAC degrader, in metastatic castration-resistant prostate cancer (mCRPC). Journal of Clinical Oncology 40 , 17-17 (2022). Békés M, Langley DR, Crews CM. PROTAC targeted protein degraders: the past is prologue. Nature Reviews Drug Discovery 21 , 181-200 (2022). Chen W, Liu X, Zhang S, Chen S. Artificial intelligence for drug discovery: Resources, methods, and applications. Molecular therapy Nucleic acids 31 , 691-702 (2023). Fleming N. How artificial intelligence is changing drug discovery. Nature 557 , S55-s57 (2018). Frye L, Bhat S, Akinsanya K, Abel R. From computer-aided drug discovery to computer-driven drug discovery. Drug discovery today Technologies 39 , 111-117 (2021). Mendez D , et al. ChEMBL: towards direct deposition of bioassay data. Nucleic Acids Research 47 , D930-D940 (2019). Knox C , et al. DrugBank 6.0: the DrugBank Knowledgebase for 2024. NUCLEIC ACIDS RESEARCH , (2023). Berman HM , et al. The Protein Data Bank. Nucleic Acids Res 28 , 235-242 (2000). Kim S , et al. PubChem in 2021: new data content and improved web interfaces. Nucleic Acids Res 49 , D1388-d1395 (2021). Wei L , et al. Hit Identification Driven by Combining Artificial Intelligence and Computational Chemistry Methods: A PI5P4K-β Case Study. J Chem Inf Model 63 , 5341-5355 (2023). You Y , et al. Artificial intelligence in cancer target identification and drug discovery. Signal transduction and targeted therapy 7 , 156 (2022). Wei M , et al. HobPre: accurate prediction of human oral bioavailability for small molecules. Journal of Cheminformatics 14 , 1 (2022). Blaschke T , et al. REINVENT 2.0: An AI Tool for De Novo Drug Design. J Chem Inf Model 60 , 5918-5922 (2020). Imrie F, Bradley AR, van der Schaar M, Deane CM. Deep Generative Models for 3D Linker Design. Journal of Chemical Information and Modeling 60 , 1983-1995 (2020). Zheng S , et al. Accelerated rational PROTAC design via deep learning and molecular simulations. Nature Machine Intelligence 4 , 739-748 (2022). Weng G , et al. PROTAC-DB: an online database of PROTACs. Nucleic Acids Res 49 , D1381-d1387 (2021). Weng G , et al. PROTAC-DB 2.0: an updated database of PROTACs. Nucleic Acids Res 51 , D1367-d1372 (2023). Li F , et al. DeepPROTACs is a deep learning-based targeted degradation predictor for PROTACs. Nature Communications 13 , 7133 (2022). Weininger D. SMILES, a chemical language and information system. 1. introduction to methodology and encoding rules. 28 , 31–36 (1988). Varadi M , et al. AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Nucleic Acids Res 50 , D439-d444 (2022). Jumper J , et al. Highly accurate protein structure prediction with AlphaFold. Nature 596 , 583-589 (2021). G L. RDKit: Open-source cheminformatics., (2022). Sastry GM, Adzhigirey M, Day T, Annabhimoju R, Sherman W. Protein and ligand preparation: parameters, protocols, and influence on virtual screening enrichments. Journal of computer-aided molecular design 27 , 221-234 (2013). Lu C , et al. OPLS4: Improving Force Field Accuracy on Challenging Regimes of Chemical Space. Journal of Chemical Theory and Computation 17 , 4291-4300 (2021). Schrödinger Release 2023-1: LigPrep, Schrödinger, LLC, New York, NY, 2023. Jacobson MP , et al. A hierarchical approach to all-atom protein loop prediction. Proteins 55 , 351-367 (2004). Schrödinger Release 2023-1: Prime, Schrödinger, LLC, New York, NY, 2023. Jacobson MP, Friesner RA, Xiang Z, Honig B. On the Role of the Crystal Environment in Determining Protein Side-chain Conformations. J Mol Biol 320 , 597-608 (2002). Wang C, Zhang Y, Deng J, Liang B, Xing D. Developments of PROTACs technology in immune-related diseases. European journal of medicinal chemistry 249 , 115127 (2023). Origin, Version 2020. OriginLab Corporation, Northampton, MA, USA. Ash S, Cline MA, Homer RW, Hurst T, Smith GBJJoci, sciences c. SYBYL line notation (SLN): A versatile language for chemical structure representation. 37 , 71-79 (1997). Homer RW, Swanson J, Jilek RJ, Hurst T, Clark RDJJoci, modeling. SYBYL line notation (SLN): a single notation to represent chemical structures, queries, reactions, and virtual libraries. 48 , 2294-2307 (2008). O'Boyle NM, Banck M, James CA, Morley C, Vandermeersch T, Hutchison GR. Open Babel: An open chemical toolbox. Journal of Cheminformatics 3 , 33 (2011). Sterling T, Irwin JJ. ZINC 15 – Ligand Discovery for Everyone. Journal of chemical information and modeling 55 , 2324-2337 (2015). Kipf TN, Welling MJapa. Semi-supervised classification with graph convolutional networks. (2016). Velickovic P, Cucurull G, Casanova A, Romero A, Lio P, Bengio YJs. Graph attention networks. 1050 , 10-48550 (2017). Paszke A , et al. Pytorch: An imperative style, high-performance deep learning library. 32 , (2019). Fey M, Lenssen JEJapa. Fast graph representation learning with PyTorch Geometric. (2019). Sanchez-Lengeling B, Reif E, Pearce A, Wiltschko ABJD. A gentle introduction to graph neural networks. 6 , e33 (2021). Abbasi K, Razzaghi P, Poso A, Amanlou M, Ghasemi JB, Masoudi-Nejad AJB. DeepCDA: deep cross-domain compound–protein affinity prediction through LSTM and convolutional neural networks. 36 , 4633-4642 (2020). Kruse R, Mostaghim S, Borgelt C, Braune C, Steinbrecher M. Multi-layer Perceptrons. In: Computational Intelligence: A Methodological Introduction (eds Kruse R, Mostaghim S, Borgelt C, Braune C, Steinbrecher M). Springer International Publishing (2022). Nair V, Hinton GE. Rectified linear units improve restricted boltzmann machines. In: Proceedings of the 27th international conference on machine learning (ICML-10) ) (2010). Li F, Hu Q, Zhou Y, Yang H, Bai F. DiffPROTACs is a deep learning-based generator for proteolysis targeting chimeras. Briefings in bioinformatics 25 , (2024). Zhao C, Zhang H, Zhang X, Liu M, Hu Z, Fan BJT. Application of support vector machine (SVM) for prediction toxic activity of different data sets. 217 , 105-119 (2006). Cano G , et al. Automatic selection of molecular descriptors using random forest: Application to drug discovery. 72 , 151-159 (2017). Pattanaik L, Coley CWJC. Molecular representation: going long on fingerprints. 6 , 1204-1207 (2020). Cereto-Massagué A, Ojeda MJ, Valls C, Mulero M, Garcia-Vallvé S, Pujadas GJM. Molecular fingerprint similarity search in virtual screening. 71 , 58-63 (2015). Additional Declarations There is NO Competing Interest. 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-5021266","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":353624271,"identity":"c76ca4b0-0f6e-41f6-9903-48dd5b8a4aef","order_by":0,"name":"Fang Bai","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0003-1468-5568","institution":"ShanghaiTech University","correspondingAuthor":true,"prefix":"","firstName":"Fang","middleName":"","lastName":"Bai","suffix":""},{"id":353624272,"identity":"647dc84a-b425-403b-a248-9fbfd2b9f8a1","order_by":1,"name":"Siyuan Tian","email":"","orcid":"https://orcid.org/0009-0006-5528-6976","institution":"ShanghaiTech University","correspondingAuthor":false,"prefix":"","firstName":"Siyuan","middleName":"","lastName":"Tian","suffix":""},{"id":353624273,"identity":"27466d12-b703-4fe7-a1f1-0914b23d6b02","order_by":2,"name":"Yilin Tang","email":"","orcid":"https://orcid.org/0009-0000-4470-1886","institution":"ShanghaiTech University","correspondingAuthor":false,"prefix":"","firstName":"Yilin","middleName":"","lastName":"Tang","suffix":""},{"id":353624274,"identity":"f7a34a14-b9fc-4647-b69d-608352197dd8","order_by":3,"name":"Fenglei Li","email":"","orcid":"","institution":"Aalto University. Konemiehentie 2, Espoo, Finland.","correspondingAuthor":false,"prefix":"","firstName":"Fenglei","middleName":"","lastName":"Li","suffix":""},{"id":353624275,"identity":"c5025bc8-cfe0-46dd-9df5-1ebbce612922","order_by":4,"name":"Zhaoxuan Li","email":"","orcid":"","institution":"ShanghaiTech University","correspondingAuthor":false,"prefix":"","firstName":"Zhaoxuan","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2024-09-03 02:25:06","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5021266/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5021266/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":64582723,"identity":"9a94ea77-8e73-4192-840e-101530fb4528","added_by":"auto","created_at":"2024-09-16 06:51:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":5445286,"visible":true,"origin":"","legend":"\u003cp\u003eThe data collection process and data format for the PROTAC-Databank. (A) First,\u0026nbsp; Data in \u0026nbsp;PROTAC-Databank was collected both from the available database PROTAC-DB and literatures. The protein structures of target proteins and E3 ligases were sourced from the PDB and Alphafold database. The construction of the PROTAC-Databank involved two main processing steps. First, the warhead and E3 ligand were aligned or docked to the binding pockets of their respective target proteins and E3 ligases. This process resulted in the formation of protein-ligand complexes. Next, PROTAC degradation levels were classified based on \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e and \u003cem\u003eD\u003c/em\u003e\u003csub\u003emax\u003c/sub\u003e values, with detailed partition rules outlined in the figure. (B) The data structure of the PROTAC-Databank is illustrated in the figure. Each entry includes SMILES files for the PROTAC and linker molecules. Over 70% of the entries contain models of the E3 ligase-ligand complex model (in .PDB format), the target-warhead complex (in .PDB format), and three classification degradation labels according to the criteria, also it contains other molecular properties collected from PROTAC-DB and source literature. labels and properties are collected in a csv file on our database website.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5021266/v1/1ca9b075c27d395e1a9f9b9c.png"},{"id":64582718,"identity":"89ef956a-7f65-451e-81d8-3599bd32137f","added_by":"auto","created_at":"2024-09-16 06:51:45","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":435090,"visible":true,"origin":"","legend":"\u003cp\u003eThe detailed architecture of DeepPROTACs 2.0. The input data consists of information from the protein pocket and the PROTAC molecule. \u0026nbsp;Protein pockets are first processed into graphical representation through mapping layers, followed by an embedding layer that prepares the pocket data for processing by the Graph Neural Network (GNN). The GNN layer comprises two Graph Attention Network (GAT) convolution layers, followed by a max pooling layer. \u0026nbsp;Simultaneously, the encoding of linker SMILES strings is processed by a separate embedding layer. Finally, the outputs from both the linker and protein pocket/ligand modules are concatenated and fed into a two-layer Multi-Layer Perceptron (MLP) to generate the final output.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5021266/v1/a3eab1b6cf52c3e9dbf24edc.png"},{"id":64583085,"identity":"d75ecb6e-bf2f-4240-936d-0bbae0fe7e1c","added_by":"auto","created_at":"2024-09-16 06:59:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":21519,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of the quantity of various PROTAC components (including PROTACs, warheads, E3 ligands, linkers, and protein-ligand complexes) between PROTAC-Databank and PROTAC-DB.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5021266/v1/56874689d13908c4dea1f753.png"},{"id":64582722,"identity":"56855f96-18a1-4aa6-84ec-46bd3568add9","added_by":"auto","created_at":"2024-09-16 06:51:46","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":3512373,"visible":true,"origin":"","legend":"\u003cp\u003eDistributions of data in PROTAC-Databank (A) Type distribution of target proteins, E3 ligase, and linker structures in PROTAC-Databank. (B) Distribution Classification results under different criteria of degradation effacy in PROTAC-Databank. (C) Statisitics on the distribution of the degradation efficacy classifications under two different Criterias(Labe1s).\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5021266/v1/66218be30ee1a941321138c6.png"},{"id":64582724,"identity":"9ea41a54-6f22-48b3-bf5b-7ec3bdb628f5","added_by":"auto","created_at":"2024-09-16 06:51:50","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":3053972,"visible":true,"origin":"","legend":"\u003cp\u003eProperty distributions of inactive and active PROTACs in PROTAC-Databank compared with traditional small-molecule drugs from DrugBank. The distributions include (A) H-bond donor counts, (B) H-bond acceptor counts, (C) heavy atom counts, (D) molecular weight, (E) TPSA, (F) LogP, (G) number of rotatable bonds, (H) linker rotatable bonds, and (I) linker heavy atom counts.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-5021266/v1/f9044939e2fd285506a919fe.png"},{"id":64582726,"identity":"1f976052-a892-40eb-89a3-4a75df8e1ce4","added_by":"auto","created_at":"2024-09-16 06:51:51","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":77614,"visible":true,"origin":"","legend":"\u003cp\u003e(a)The ROC curves of different results of DeepPROTACs and DeepPROTACs 2.0 on PROTAC-DB and PROTAC-Databank.(b)ROC curves of DeepPROTACs 2.0, DeepPROTACs, SVM(Support Vector Machine) and RF(Random Forest) models.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-5021266/v1/a2b8c0cd06369a7c46e01df7.png"},{"id":69010923,"identity":"b4295c53-f06f-4273-8808-246d76ddea1a","added_by":"auto","created_at":"2024-11-14 13:49:03","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":14371633,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5021266/v1/90adac5f-532f-4cd0-9366-c5cb60051c67.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"PROTAC-Databank: The Present Largest Integrated Resource of PROTACs, Enabling the Enhanced DeepPROTACs 2.0 for Degradation Prediction","fulltext":[{"header":"Introduction","content":"\u003cp\u003eIn recent decades, various cancer therapy strategies have been developed inhibit the activity of oncogene proteins. Small molecular inhibitors have been used in cancer treatment by blocking the active sites of target proteins, thereby disabling oncogene proteins. While these inhibitors have shown promise, their further development limited by issues such as the accumulation of blocked target proteins, drug resistance, and the undruggable nature of some protein structure.\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e Antibodies, though highly specific to their target proteins, suffer from low membrane permeability, which limits their effectiveness inside cancer cells. \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e Another approach involves the use of siRNAs to directly inhibit gene expression, but off-target effects and challenges in drug delivery have hindered their broader application. \u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e Therefore, new molecule modalities are required in the current drug discovery landscape.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eProteolysis targeting chimeras (PROTACs) represent a promising strategy designed to target and accelerate the degradation of specific proteins.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e PROTACs are a class of small molecules with unique structures that utilizes the ubiquitin-protease system (UPS) to degrade target proteins.\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e A typical PROTAC molecule comprises three components: a target protein ligand (also known as the warhead), a linker, and an E3 ubiquitin ligase recruiting ligand (E3 ligand). These three parts are covalently boned to form the complete molecule.\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThe ligands at both ends of the PROTAC bind to the target proteins and the E3 ligase respectively, bringing the target protein into closely proximity with the E3 ubiquitin complex.\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e This closely proximity facilitates the transfer of ubiquitin from the E2 enzyme to a lysine residue on the surface of the target protein. Subsequently, the polyubiquitinated target protein is recognized by 26S proteasome and degraded into peptide fragments.\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e In essence, PROTACs hijack the UPS to specifically degrade target proteins, thereby reducing intracellular levels of these proteins.\u003c/p\u003e \u003cp\u003eWith the advantages of small molecules, PROTACs offer easier drug delivery compared to antibodies and siRNA. Unlike traditional small molecule inhibitors, which can lead to the accumulation of inactivated protein within cells, PROTACs completely eliminate target proteins, preventing such accumulation. \u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e Moreover, recently, PROTACs have also been designed to treat diseases beyond cancer, such as autoimmune disorders,\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e virus infection\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003ePROTACs are well-suited for targeting \u0026ldquo;undruggable\u0026rdquo; targets, because they don\u0026rsquo;t require strong binding affinities between themselves and target proteins, making them highly versatile. Since Craig M. Crews developed the first PROTAC in 2001, the field has seen tremendous development.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e In recent years, several PROTAC candidate drugs has entered clinical trials, including ARV-110, an androgen receptor (AR) target PROTAC for treating metastatic castration-resistant prostate cancer.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e The success of PROTAC research and development has significantly driven the growth of related literature. Since 2018, the volume of literature on PROTACs has grown rapidly, with newly developed PROTACs being reported at an unprecedented rate. \u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e As of November 2023, PubMed lists 2,007 PROTAC-related publications, with1,880 of them published after 2018. The explosion of data presents opportunities to apply artificial intelligence (AI) in designing methods development for PROTACs.\u003c/p\u003e \u003cp\u003eThe advancements in high-performance computing and multi-omics data have facilitated the practical application of AI application in any field lies in the availability of high-quality datasets. \u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e Fortunately, numerous high-quality biological information databases, such as, ChEMBL\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e, DrugBank\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e, PDB\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e, and PubChem\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e, allow AI techniques to be applied to multiple aspects of drug discovery, including protein structure prediction, hit identification\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e, drug target discovery\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, bioavailability prediction\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e, and \u003cem\u003ede novo\u003c/em\u003e drug design\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e etc.\u003c/p\u003e \u003cp\u003eCurrently, AI methods have begun to be employed in the rational design of PROTACs.\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e To facilitate rational PROTAC design, it is essential to predict their properties accurately, necessitating the establishment of a database specifically for PROTACs to support AI models. PROTAC-DB\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e is one such database, its updated version PROTAC-DB 2.0 currently archives 3,270 PROTACs. \u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e It includes biological activity data such as \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e (half minimum degradation concentration)data, cellular activity data, western blotting data, and binding affinity. Additionally, the database contains ternary structural data for target-protac-E3 ligase is part of the database. However, PROTAC-DB provides limited ternary structure for target-protac-E3ligase complexes, with only 18 crystal structures and 664 predicted structures available, leaving the binding patterns of warhead-target and E3 ligand-E3 ligase complexes missing for 80% of PROTACs. Moreover, the format of PROTAC-DB is not directly usable in AI models due to inconsistencies in PROTAC degradation standards. Our literature review also revealed that a significant number of PROTACs are not included in PROTAC-DB.\u003c/p\u003e \u003cp\u003eIn this paper, we present a newly established PROTAC database containing target-warhead and E3 ligase-E3 ligand complexes for each PROTAC molecule. Our data were sourced from both PRORAC-DB and extensive literature research, with degradation labels assigned to each target-PROTAC-E3 Ligase pair. And degradation labels were tagged to each Target-PROTAC-E3ligase pairs. The data can be directly used as dataset for AI models such as DeepPROTACs\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e for PROTAC degradation efficiency prediction. Our database provided the pocket sites information of target protein and E3 ligase. Besides, the property comparation between traditional small molecular drugs and PROTAC are also displayed in this paper. Such analysis may help further PROTAC design and modification.\u003c/p\u003e \u003cp\u003eBesides the construction of database, we also developed a second version of DeepPROTACs by designing a new deep learning framework.\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e. DeepPROTACs was previously developed by our group. It can predicted the PROTAC degradation efficiency using input files containing the PROTAC molecule, and target-warhead complex structure, and E3 ligase-ligand complex structure. Building on the work of DeepPROTACs 1.0, we refined the architecture of the model by using Graph Attention Networks (GANs) and improved accuracy rate using the original data set. We also incorporate new datasets from the PROTAC-Databank, further enhancing prediction efficiency. With updates to both the model architecture and dataset, we have released DeepPROTACs 2.0 in this manuscript. A webserver is also available for free use via the link\u003c/p\u003e \u003cp\u003e(\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://bailab.siais.shanghaitech.edu.cn/services/deepprotac-db\u003c/span\u003e\u003cspan address=\"https://bailab.siais.shanghaitech.edu.cn/services/deepprotac-db\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e"},{"header":"Method","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData source of PROTAC-Databank\u003c/h2\u003e \u003cp\u003eThe raw data of our database were sourced from PROTAC-DB\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e and other literature-reported PROTACs. We obtained information such as PROTAC SMILES\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e, warhead (target ligand) SMILES, E3 ligand SMILES, Target protein type, E3 ligase type from PROTAC-DB 2.0\u003csup\u003e31\u003c/sup\u003e. We then we queried \u0026ldquo;PROTAC\u0026rdquo; in PubMed to collect literature not included in PROTAC-DB 2.0. Information such as PROTAC SMILES, warhead SMILES, E3 ligand SMILES, target protein, E3 ligase, \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e, \u003cem\u003eD\u003c/em\u003e\u003csub\u003emax\u003c/sub\u003e, etc, were extracted from literatures and extended in database. (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.A)\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eData processing of PROTAC-Databank\u003c/h2\u003e \u003cp\u003e \u003cb\u003eProtein Structure Collection\u003c/b\u003e. After collecting PROTAC raw data, we generated complex binding structures between warhead and targets, as well as complex models between E3 ligase and E3 ligand by molecular docking simulations. (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.A) We first listed the types of target proteins and E3 ligases from the raw data, then searched the PDB database for experimentally determined complex structures corresponding to these proteins.\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e For each protein structure, we identified the pocket site that binds with the warhead or E3 ligand, selecting the lowest resolution structure containing models with the same or similar ligands as the PROTAC warhead(or E3 ligand). The complex structure of warheads with target proteins or ligands with E3 ligases were generated by molecule docking simulations, confirming the pocket sites and reference ligand binding poses for each protein-ligand pair.\u003c/p\u003e \u003cp\u003eAlthough most of protein structures for these PROTAC models were found in PDB, some were not available as experimental structures. For these proteins, we used predicted structures by AlphaFold2\u003csup\u003e34, 35\u003c/sup\u003e (Protein types using AlphaFold2 predicted models: GSK3A, TYR). Once the models were ready, ligands were clustered into different groups based on their target protein or E3 ligase. We also converted SMILES and E3 ligand SMILES into 3D structures in PDB format using the RDKit\u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e package in Python.\u003c/p\u003e \u003cp\u003e \u003cb\u003eMolecule Docking\u003c/b\u003e. To obtain the binding complex of protein-ligand, molecular docking simulations were performed. Firstly, Protein Preparation Workflow module of Schrodinger 2023\u003csup\u003e37\u003c/sup\u003e was used to process the protein models from the above obtained databases, including filling the missing side chain, optimizing H-bond assignment, minimizing structure in OPLS_4 force field\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e and deleting water in models. To produce the low-energy 3D conformations of collected ligands, we process molecule PDB files by LigPrep module of Schrodinger 2023.\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e The reference ligands for docking small molecules into the pocket were selected from downloaded PDB complex models. Ligands for the same protein were aligned to the reference pose of the corresponding reference ligand using Schrodinger 2023's Ligand Alignment module. Subsequently, the aligned small molecules were combined with prepared protein models individually, generating protein-ligand complex models for each pair of target-warhead and E3 ligase-E3 ligand. Schrodinger 2023's Prime\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e minimize were employed to optimize structure and minimize energy of models. Through this process, we generated both target-warhead complex models and E3 ligase-E3 ligand models. (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.A)\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eClassification of degradation efficiency labels:\u003c/h2\u003e \u003cp\u003eA unified standard for PROTAC degradation\u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e efficiency was established for each target-PROTAC-E3 ligase pair in the database. \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e(half maximal degradation concentration) is one of the measurements for evaluating the degradation efficiency. For the \u0026ldquo;Degradation efficacy\u0026rdquo; label, pairs with \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e\u0026thinsp;\u0026lt;\u0026thinsp;10nM are labelled \u0026ldquo;excellent\u0026rdquo;, 10nM\u0026thinsp;\u0026le;\u0026thinsp;\u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e\u0026thinsp;\u0026lt;\u0026thinsp;100nM are labelled \u0026ldquo;good\u0026rdquo;, 100nM\u0026thinsp;\u0026le;\u0026thinsp;\u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e\u0026thinsp;\u0026lt;\u0026thinsp;500nM are labelled \u0026ldquo;moderate\u0026rdquo;, 500nM\u0026thinsp;\u0026le;\u0026thinsp;\u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e\u0026thinsp;\u0026lt;\u0026thinsp;1000nM are labelled \u0026ldquo;not good\u0026rdquo;, \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e\u0026thinsp;\u0026ge;\u0026thinsp;1000nM are labelled \u0026ldquo;poor\u0026rdquo;. \u003cem\u003eD\u003c/em\u003e\u003csub\u003emax\u003c/sub\u003e(maximum degradation rate ) in another factor that is induced to measure the PROTAC degradation level.Two criteria have been established for labeling PROTAC degradation effectiveness. For \u0026ldquo;Label 1\u0026rdquo;, \u0026ldquo;Criterion 1\u0026rdquo; classifies pairs with a \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e\u0026thinsp;\u0026le;\u0026thinsp;100 nM and a \u003cem\u003eD\u003c/em\u003e\u003csub\u003emax\u003c/sub\u003e\u0026ge; 80 as \"Good,\" while pairs that do not meet these thresholds are categorized as \u0026ldquo;Poor.\u0026rdquo; For \u0026ldquo;Label 2\u0026rdquo;, \u0026ldquo;Criterion 2\u0026rdquo; deems pairs with a \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e\u0026thinsp;\u0026le;\u0026thinsp;1000 nM and a \u003cem\u003eD\u003c/em\u003e\u003csub\u003emax\u003c/sub\u003e\u0026ge; 70 as \"Good,\" with all other configurations considered \"Poor\". When the \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e data was missing for some PROTAC pairs, western blot analysis graph of degradation experiments in original literature would be viewed to estimate the approximate \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e and \u003cem\u003eD\u003c/em\u003e\u003csub\u003emax\u003c/sub\u003e values to label the degradation efficiency. (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.A)\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eData analysis of PROTAC-Databank\u003c/h2\u003e \u003cp\u003eThe RDkit package was used to calculate molecular properties of PROTACs in PROTAC-Databank and drug molecules in DrugBank\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. The following properties were computed using RDkit: linker heavy atom numbers of linkers, rotatable bonds of linkers of PROTACs; H-bond donor counts, H-bond acceptor counts, partition coefficient (logP), topological polar surface area (TPSA), molecular weight, heavy atom counts, rotatable atom counts of both PROTACs molecule in PROTAC-Databank, and drug molecules in DrugBank. Origin 2020 \u003csup\u003e44\u003c/sup\u003e was used to analysis and visualize those statistical results.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003eUpdating from DeepPROTACs to DeepPROTACs 2.0\u003c/h2\u003e \u003cp\u003ePreviously, the DeepPROTACs model was trained on a dataset of 2,832 labeled PROTACs. Despite its success in predicting PROTAC degradation rates, the dataset size was relatively small for deep learning applications. To address this, we expanded the dataset to include 4,142 labeled PROTACs, which were used to train the updated DeepPROTACs 2.0 model.\u003c/p\u003e \u003cp\u003eIn the data processing pipeline, based on the original DeepPROTACs approach, we extracted the binding pockets by selecting residues within 5\u0026Aring; around the binding ligand using PyMol software. The extracted structures were then converted to Mol2files\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e, \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e using OpenBabel\u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e. These files contain information about the types, coordinates, and bonding of the atoms, allowing for the simple reconstruction of molecular structures. Since atoms and their bonds in molecules can naturally be represented as nodes and edges in graphs, we used a graph data structure to represent binding pockets and ligands. Graph representations were built based on the Mol2 files using an adjacency matrix. In these graphs, atoms were represented as nodes, with atomic types encoded in a dictionary (C, N, O, S, and others as 0, 1, 2, 3, 4 respectively). Covalent bonds were represented as edges, with adjacency matrix positions labeled 1 for bonded atoms and 0 otherwise. Given the relatively limited number of covalent bonds between the molecules, the adjacency matrix of the molecules is relatively sparse, and at the same time the amount of operations brought by the adjacency matrix is relatively high. For this reason, we adopted a simple structure of edge list to represent the bonding information between atoms. If there was a covalent bond between atom n and atom m, then two dot pairs [n,m] and [m,n] are added to the list. In the graphs of ligand, we also used atoms to represent nodes and bonds to represent edges. However, the number of atom types in the ligand increases and a larger dictionary is needed to represent this mapping relationship, with C, N, O, S, F, Cl, Br, I, P, and other atoms denoted by 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, respectively. In addition, the different types of bonds between atoms were also encoded as attribute of the edge nodes in the graph, there are single bond, double bond, triple bond, aromatic bond, and amide bond, each of which was encoded by 1, 2, 3, 4, 5. PROTACs' linkers were encoded in SMILES format and based on a table from the ZINC database\u003csup\u003e\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e, with the most frequent 39 characters encoded from 1 to 39, and the remaining characters encoded as 40.\u003c/p\u003e \u003cp\u003eIn advancing the DeepPROTACs model, we transitioned from conventional sequence-base models to Graph Convolutional Networks (GCNs)\u003csup\u003e\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e for feature extraction from graph representations of protein pockets. In GCNs, the features of each node are averaged from its neighboring nodes, implying equal contributions from all neighbors. However, the functionality and influence of a molecular group is shaped by its neighbors, with varying impacts on the compound\u0026rsquo;s properties, meaning that neighboring nodes should contribute differently to the feature computation. This revealed a limitation in GCNs, where treating all neighboring nodes as equally contributive fails to capture the differential impacts in molecular contexts.\u003c/p\u003e \u003cp\u003eTo overcome this, we introduced Graph Attention Networks (GATs)\u003csup\u003e\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u003c/sup\u003e for feature learning within graph representations of protein pockets. GATs uses an attention mechanism that allows individual nodes to dynamically assign different weights to their neighboring nodes during the messae-passing phase. This enables a node to focus on the most significant or relevant neighbors in tis structural or functional context. By incorporating this attention-based approach, GATs provide greater flexibility and precision in capturing the complex relationships between nodes, allowing for a more accurate representation of interactions within the protein pocket.\u003c/p\u003e \u003cp\u003eFurthermore, recognizing the importance of atom-to-atom distances in potential non-bonded interactions, we refined our graph representation by including these distances as additional features for each atom. This enhancement not only enriches the graph representation but also optimizes the subsequent graph network convolution calculations, equipping our refined model to better understand and analyze the intricate details of the protein pocket structure and non-bonded molecular interactions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eThe Architecture of DeepPROTACs 2.0\u003c/h2\u003e \u003cp\u003eThe architecture of our network, illustrated in Figure.2, is developed entirely using the PyTorch\u003csup\u003e\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e\u003c/sup\u003e and PyTorch Geometric frameworks\u003csup\u003e\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e\u003c/sup\u003e. Initially, protein pockets are represented graphically through a series of mapping layers. An embedding layer then increases the dimensionality of the data to 64, preparing it for processing by the subsequent graph neural network (GNN)\u003csup\u003e\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e. The core of the GNN comprises two GAT convolution layers, followed by a max pooling layer, maintaining an output dimensionality of 64. In parallel, the encoding of linker SMILES strings is processed through a separate embedding layer to ensure the linker representation matches the dimensionality of the protein pocket. This encoded linker data is then processed by a Bidirectional Long Short-Term Memory (Bi-LSTM)\u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e layer and a fully connected layer, with the output dimensionality also set to 64. Finally, the outputs from both the linker and protein pocket/ligand modules are concatenated and fed into a two-layer Multi-Layer Perceptron (MLP)\u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e to generate the final output. Throughout the network, the Rectified Linear Unit (ReLU) \u003csup\u003e\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003efunction serves as the activation mechanism, enabling non-linear learning capabilities.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis of PROTAC-Databank\u003c/h2\u003e \u003cp\u003eFollowing the construction of the PROTAC-Databank, a comprehensive statistical analysis was performed. The database contains a total of 3,645 PROTACs, including 612 types of warheads, 145 types of E3 ligands, and 1,535 types of linkers (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Regarding degradation labels, 74.1% of the data entries include degradation labels, while the remaining 25.1% lack these labels due to missing \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e and \u003cem\u003eD\u003c/em\u003e\u003csub\u003emax\u003c/sub\u003e values in the source data. For entries with degradation labels, protein-ligand complexes of both the Target-warhead and E3 ligase-E3 ligand pairs were constructed, resulting in 4,142 pairs of protein-ligand complex structures. The PROTAC-Databank includes 4,142 complete data sets, each comprising basic information, a PDB file of the Target-warhead complex, a PDB file of the E3 ligase-E3 ligand complex, a SMILES file of the PROTAC, three degradation labels, and the source literature. Some data sets also contain additional information, such as \u003cem\u003eK\u003c/em\u003e\u003csub\u003eD\u003c/sub\u003e values (PROTAC to E3 ligase and PROTAC to Target), \u003cem\u003eIC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e values from specific experiments, and other relevant data. The PROTAC-Databank data is stored in a CSV file and is accessible online (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://bailab.siais.shanghaitech.edu.cn/services/deepprotac-db\u003c/span\u003e\u003cspan address=\"https://bailab.siais.shanghaitech.edu.cn/services/deepprotac-db\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). The PROTAC IDs from PROTAC-DB 2.0 are consistent with those in the PROTAC-Databank, while entries not from PROTAC-DBs are renumbered with the prefix \"b.\"\u003c/p\u003e \u003cp\u003eThe PROTAC-Databank exhibits greater diversity than the database PROTAC-DB. It contains 3,645 types of PROTACs, surpassing the 3,270 types found in PROTAC-DB. Furthermore, PROTAC-Databank includes 612 distinct warheads, significantly more than the 365 found in PROTAC-DB. However, both databases show limited diversity in E3 ligands and E3 ligases. PROTAC-Databank includes 145 types of E3 ligands, while PROTAC-DB has 82. The number of linker types is comparable between the two, with 1,535 types in PROTAC-Databk and 1,510 in PROTAC-DB, this also reflects that the previous design strategies for PROTAC molecules were primarily focused on empirically replacing a limited variety of linkers, with very limited innovation in the molecules themselves. Therefore, developing linker generation algorithms based on generative methods, such as the deep learning-based PROTAC linker generation method, i.e., DiffPROTACs\u003csup\u003e\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e\u003c/sup\u003e, developed by our research group, is particularly important for expanding the diversity of PROTAC linkers and advancing the field.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eA key distinction between the two databases is that PROTAC-Databank includes more structural information. It contains 4,142 pairs of target-warhead and E3 ligase-E3 ligand complexes, compared to 664 ternary complexes in PROTAC-DB. By incorporating protein-ligand complexes, PROTAC-Databank offers a new dimension for PROTAC design and the development of AI methods related to PROTACs.\u003c/p\u003e \u003cp\u003eThe database includes 14 types of E3 ligase proteins, with CRBN being the most frequently utilized, representing 65.8% of the PROTAC pairs targeting protein degradation. VHL is the second most common, used in 28.8% of the pairs. Over 90% of the E3 ligases employed by PROTACs in the database are either CRBN or VHL. The distribution of target proteins is more varied compared to E3 ligases (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA).\u003c/p\u003e \u003cp\u003eThe database contains 337 types of target proteins, with different mutants of a single wild-type protein considered as distinct targets. The Androgen Receptor (AR) is the most common target, representing 5% of all target proteins. Other frequently targeted proteins include the Estrogen Receptor (ER), BRD4, BTK, and CDK9. In the statistical results, 67.5% of the target proteins are individual proteins, each constituting less than 1.7% of the total targets (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA).\u003c/p\u003e \u003cp\u003eLinker properties in PROTACs are a crucial factor in determining degradation efficiency. A cluster of PROTAC molecules with identical warheads and E3 ligands can exhibit varying degradation efficiencies due to differences in their linkers. Therefore, we also analyzed the types of linkers in the PROTAC-Databank (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA). Half of the linkers are difficult to classify into specific types due to their complex structures. However, 35.2% of the linkers contain PEG segments, 8.8% are alkanes, and others include piperazine, piperidine, and alkyne segments.\u003c/p\u003e \u003cp\u003eThe degradation labels in the database are categorized into five groups based on the \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e values obtained from PROTAC degradation experiments (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.B). To simplify the labels for AI model input, two classifications \u0026ldquo;Label 1\u0026rdquo; and \u0026ldquo;Label 2\u0026rdquo; were classified into \"Good\" or \"Bad\" categories based on both \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e and \u003cem\u003eD\u003c/em\u003e\u003csub\u003emax\u003c/sub\u003e values. The criteria for these classifications differ: In \u0026ldquo;Criteria 1\u0026rdquo; for \u0026ldquo;Label 1\u0026rdquo;, PROTAC pairs with \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e\u0026thinsp;\u0026le;\u0026thinsp;100 nM and Dmax\u0026thinsp;\u0026ge;\u0026thinsp;80 are labeled as \"Good,\" while all other cases are labeled as \"Bad\". In \u0026ldquo;Criteria 2\u0026rdquo; for \u0026ldquo;Label 2\u0026rdquo;, pairs with \u003cem\u003eDC\u003c/em\u003e\u003csub\u003e50\u003c/sub\u003e\u0026thinsp;\u0026le;\u0026thinsp;1000 nM and \u003cem\u003eD\u003c/em\u003e\u003csub\u003emax\u003c/sub\u003e \u0026ge; 70 are labeled as \"Good,\" with the remainder labeled as \"Bad.\" According to the statistical analysis of \"Label 1,\" the number of negative labels in the database exceeds positive labels by approximately 30% (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eC). However, with the relaxed criteria in \"Label 2,\" the numbers of \"Good\" and \"Bad\" labels are nearly equivalent (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eC).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eComparison of Properties Between PROTAC Molecules and Tradition Drug Molecules\u003c/b\u003e.\u003c/p\u003e \u003cp\u003eDue to the unique mechanisms and structures of PROTACs, their properties differ significantly from traditional small-molecule drugs. A comparison was conducted between the molecular properties of PROTACs in the PROTAC-Bank and traditional drug molecules in DrugBank. Additionally, PROTACs were categorized into \"Active\" and \"Inactive\" groups based on the \"Good\" and \"Bad\" labels from the \"Label 1\" during the statistical analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). The analysis revealed significant differences in multiple drug properties.\u003c/p\u003e \u003cp\u003eThe molecular weight of PROTACs is consistently higher than that of traditional drugs, ranging from 600 to 1300 Da, with a peak at 800\u0026ndash;900 Da, while most traditional drugs have a molecular weight below 500 Da. This difference in size leads to distinct variations in other properties as well. PROTAC molecules have more hydrogen bond donors and acceptors, as well as more rotatable bonds, compared to DrugBank molecules (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. A-G). Additionally, the topological polar surface area (TPSA) and logP were calculated to further analyze these differences.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTPSA tends to be higher in PROTACs compared to traditional small-molecule drugs. The peak of TPSA distribution for PROTACs is around 200, while for DrugBank compounds, it is below 150. The logarithm of the partition coefficient between octanol and water (logP) indicates the differential solubility of a compound in these solvents. Given that many PROTAC linkers have alkane backbones, their lipophilicity is increased, resulting in a right-shifted logP distribution for PROTACs compared to traditional drug compounds.\u003c/p\u003e \u003cp\u003eIn summary, the properties of PROTAC molecules differ significantly from those of traditional small molecular drugs in DrugBank, suggesting potential challenges in designing druggable PROTACs and ensuring their effective delivery.\u003c/p\u003e \u003cp\u003eAdditionally, the distribution of linker heavy atom counts and linker rotatable bond counts in the PROTAC-Databank was analyzed (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. H, I). The heavy atom linker counts showed an overlapping distribution between active and inactive PROTACs, with linkers containing 5 to 9 atoms being the most common in both groups. However, there is a slight difference in the peaks of these distributions: linkers with 5 and 6 atoms are most prevalent in inactive molecules, while linkers with 7 and 8 atoms dominate in active molecules.\u003c/p\u003e \u003cp\u003eFor the number of rotatable bonds in linkers, inactive molecules tend to have one more rotatable bond on average compared to active molecules. The peak number of rotatable bonds for \"active\" molecules is 4 to 5, while for \"inactive\" molecules, it is 5 to 7. This suggests that PROTACs with better degradation efficiency may have relatively more rigid linkers compared to those with poorer efficiency.\u003c/p\u003e \u003cp\u003eThese statistical results provide insight into the current trends in PROTAC linker design. Over half of the reported PROTAC molecules have linkers with 5 to 9 heavy atoms and 3 to 7 rotatable bonds, regardless of whether the molecules are classified as \"active\" or \"inactive.\"\u003c/p\u003e \u003cp\u003e \u003cb\u003ePrediction Efficiency of DeepPROTACs 2.0 Compared to Multiple Methods.\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe initial dataset used in the previous DeepPROTACs study comprised 2,832 labeled PROTACs from PROTAC-DB. These were categorized based on their \u003cem\u003eDC\u003c/em\u003e₅₀ and \u003cem\u003eD\u003c/em\u003e\u003csub\u003e\u003cem\u003emax\u003c/em\u003e\u003c/sub\u003e values, with PROTACs having \u003cem\u003eDC\u003c/em\u003e₅₀ below 100 nM and \u003cem\u003eD\u003c/em\u003e\u003csub\u003emax\u003c/sub\u003e above 80% labeled as \"good,\" while all others were labeled as \"bad.\" This resulted in 988 \"good\" and 1,844 \"bad\" data points. Upon expanding the dataset, the total number of PROTACs increased to 4,140, with 1,474 \"good\" and 2,666 \"bad\" data points.\u003c/p\u003e \u003cp\u003eIn previous work using DeepPROTACs, the model achieved an accuracy of 77.15% and an AUROC of 0.8246 on the 2,832 data points with optimal parameter settings. After expanding the dataset, the same model yielded an improved accuracy of 81.82% and an AUROC of 0.8734 (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, with the ROC curve presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA). This significant enhancement in accuracy following the database expansion demonstrates the effectiveness of our data collection and processing.\u003c/p\u003e \u003cp\u003eIn our current work, we further optimized the original model by incorporating inter-atomic distances as a feature. Experimental results showed that the new model achieved an accuracy of 79.68% and an AUROC of 0.8493 on the original 2,832 data points. On the expanded dataset of 4,140 data points, the model achieved an accuracy of 83.45% and an AUROC of 0.9001. In both cases, the new model outperformed the previous version. Overall, our findings indicate that, compared to the original DeepPROTACs, our model's accuracy improved by over 8% and the AUROC by over 9%.\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\u003ePerformance comparison of DeepPROTACs models on different datasets. The table displays average accuracy and AUROC scores for DeepPROTACs 1.0 and DeepPROTACs 2.0 using the PROTAC-DB and PROTAC-Databank datasets.\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=\"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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAverage Accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAUROC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDeepPROTACs 1.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePROTAC-DB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e77.15%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8246\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDeepPROTACs 1.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePROTAC-Databank\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e81.82%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8734\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDeepPROTACs 2.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePROTAC-DB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e79.68%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8493\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDeepPROTACs 2.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePROTAC-Databank\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e83.45%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9001\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\u003eWhen compared to traditional machine learning models such as Support Vector Machines (SVM) \u003csup\u003e\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u003c/sup\u003e and Random Forests (RF) \u003csup\u003e\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e\u003c/sup\u003e, which utilized MACCS and Morgan molecular fingerprints for PROTAC ligands and ligases\u003csup\u003e\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e, \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e, our DeepPROTACs 2.0 approach achieved superior accuracy and AUROC (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, with the ROC curve presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eB).\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\u003eThe average accuracy and AUROC scores for DeepPROTACs 2.0, SVM with MACCS and Morgan fingerprints, and RF with MACCS and Morgan fingerprints.\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=\"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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFingerprints\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAverage Accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAUROC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDeepPROTACs 2.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e83.45%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.9001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMACCS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e72.54%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7333\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMorgan\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e77.78%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7989\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMACCS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e78.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8550\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMorgan\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e78.66%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8746\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\u003eWe also assessed the generalization ability of our model. We removed data for specific ligands from the training set and used the remaining data to train the model. After training, the removed data served as a test set to evaluate model predictions. The results showed that, after optimizing the dataset and model, the prediction accuracy for specific target proteins exceeded 90% (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). This surpasses the performance reported in the original DeepPROTACs study, indicating that our improved model has enhanced generalization capabilities.\u003c/p\u003e \u003cp\u003eTo further validate our model's predictive capacity, we used an experimental database from DeepPROTACs, which includes data on 16 PROTACs. While DeepPROTACs accurately predicted 11 out of 16 PROTACs (68.75% accuracy), our current model successfully predicted 12 PROTACs, achieving a 75% accuracy rate.\u003c/p\u003e \u003cp\u003eIn the expanded dataset of 4,140 PROTACs, there were 1,474 \"good\" data points and 2,666 \"bad\" data points. In addition to the original data distribution, we employed both over-sampling and under-sampling techniques, adjusting the data ratios to 1,474:1,474 and 2,666:2,666, respectively. For each sampling method, we independently trained our model four times and conducted predictions, calculating the average accuracy and AUROC for each method. The results, as shown in the figure, indicate that models trained on oversampled data exhibit better performance in both accuracy and AUROC (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). This aligns with conclusions drawn from previous DeepPROTACs work.\u003c/p\u003e \u003cp\u003eWe have also launched a model prediction service, allowing users to input data using the same PROTAC processing methods described in this article to obtain degradation rate predictions. The dataset used for training the model is accessible on our website (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://bailab.siais.shanghaitech.edu.cn/services/deepprotac-db\u003c/span\u003e\u003cspan address=\"https://bailab.siais.shanghaitech.edu.cn/services/deepprotac-db\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e ), where users can search by PROTAC ID, Target, and E3 ligases.\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\u003eAverage accuracy and AUROC of DeepPROTACs 2.0 on four different targets(EZH2, eIF4E, STAT, FLT)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTarget\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAverage Accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAUROC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEZH2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e90.32%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.96131\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eeIF4E\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e92.70%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.95635\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSTAT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e92.70%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.94048\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFLT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e91.07%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.92653\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 \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\u003eAverage accuracy and AUROC of different sample methods\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSampling Method\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAverage Accuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAUROC\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnder-sampling\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e78.15%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8524\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNormal-sampling\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e83.45%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOver-sampling\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e84.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.9109\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 \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we developed PROTAC-bank, a comprehensive database of predicted ligand-protein complex structures for existing PROTAC molecules. The database aggregates and processes data from PROTAC-DB and relevant literature, constructing and refining both \"warhead-target protein\" and \"E3 ligase-E3 ligand\" structures for each PROTAC. These structures are labeled according to their corresponding PROTAC degradation levels. The statistical analysis of PROTAC-Bank, as presented above, highlights the distribution of target proteins and E3 ligases, and underscores the distinctions between PROTACs and traditional small molecule drugs. Moreover, we utilized PROTAC-Bank to enhance the prediction accuracy of DeepPROTACs by expanding the dataset. Additionally, we optimized the AI model underlying DeepPROTACs, leading to a significant improvement in accuracy from 77.15\u0026ndash;83.45%.\u003c/p\u003e \u003cp\u003ePROTAC-Bank is a unique resource that focuses on protein-ligand structures, providing datasets for target-warhead and E3 ligase-E3 ligand interactions. By linking PROTAC degradation efficiency with these ligand-protein structures, PROTAC-bank offers a significant advantage: the ability to incorporate structural information and ligand-protein interactions into existing training sets as new features. This facilitates the future development and refinement of AI models for designing and evaluating PROTAC molecules. Furthermore, the database serves as a valuable tool for optimizing and updating existing models and can contribute to theoretical research on PROTAC degradation mechanisms and the design of novel PROTACs.\u003c/p\u003e \u003cp\u003eThe database also reveals the chemical property differences between PROTACs and traditional small molecule drugs. The statistical analysis shows that most PROTACs have a molecular weight exceeding 500 Da, often failing to comply with Lipinski's rule of five. Moreover, when considering molecular properties such as logP, H-bond acceptor and donor counts, and the number of rotatable bonds, the majority of PROTAC molecules do not meet the criteria set by Lipinski\u0026rsquo;s rules, indicating their limited druggable potential. Addressing the administration of PROTACs is thus a crucial challenge that needs to be addressed. Future research directions should focus on refining PROTACs to optimize their ADMET properties and developing effective strategies for their delivery.\u003c/p\u003e \u003cp\u003eThe updated DeepPROTACs model has achieved a higher accuracy in predicting PROTAC degradation efficiency, making it a valuable tool for screening and evaluating designed PROTACs before molecule synthesis and biological experimentation. This can significantly reduce both the cost and time associated with PROTAC discovery.\u003c/p\u003e \u003cp\u003eDespite the significant contributions of PROTAC-Bank to PROTAC research and AI model enhancement, there is still room for improvement. During data collection, some target proteins lacked experimental or predicted structures, and others were missing pocket site information, leading to their exclusion from PROTAC-Bank. Additionally, with the rapid increase in PROTAC-related literature over the past five years, maintaining an up-to-date PROTAC-bank requires continuous effort. Currently, data collection is done manually, but more efficient methods for updating PROTAC data are needed. A long-term goal is to create a platform that allows researchers to contribute their discovered PROTACs to the database.\u003c/p\u003e \u003cp\u003eIn conclusion, we have established the PROTAC ligand-protein structure database, PROTAC-Bank, and used it to update and retrain DeepPROTACs, achieving a higher accuracy rate. These advancements will greatly facilitate the design of novel PROTACs and support theoretical research in the field.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eConflicts of interest/Competing interests\u003c/h2\u003e \u003cp\u003eThe authors have no relevant financial or non-financial interests to disclose.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e \u003cp\u003eThis work was supported by National Key R\u0026amp;D Program of China (Grant IDs: 2022YFC3400501 and 2022YFC3400500), Shanghai Science and Technology Development Fund (22ZR1441400), start-up package from ShanghaiTech University, and Shanghai Frontiers Science Center for Biomacromolecules and Precision Medicine at ShanghaiTech University. We are grateful for the support from HPC Platform of ShanghaiTech University.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eLu J\u003cem\u003e, et al.\u003c/em\u003e Hijacking the E3 Ubiquitin Ligase Cereblon to Efficiently Target BRD4. \u003cem\u003eChemistry \u0026amp; biology\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 755-763 (2015).\u003c/li\u003e\n\u003cli\u003eLyu X\u003cem\u003e, et al.\u003c/em\u003e The global landscape of approved antibody therapies. \u003cem\u003eAntibody therapeutics\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 233-257 (2022).\u003c/li\u003e\n\u003cli\u003eAlshaer W\u003cem\u003e, et al.\u003c/em\u003e siRNA: Mechanism of action, challenges, and therapeutic approaches. \u003cem\u003eEuropean journal of pharmacology\u003c/em\u003e \u003cstrong\u003e905\u003c/strong\u003e, 174178 (2021).\u003c/li\u003e\n\u003cli\u003eSakamoto KM, Kim KB, Kumagai A, Mercurio F, Crews CM, Deshaies RJ. Protacs: chimeric molecules that target proteins to the Skp1-Cullin-F box complex for ubiquitination and degradation. \u003cem\u003eProceedings of the National Academy of Sciences of the United States of America\u003c/em\u003e \u003cstrong\u003e98\u003c/strong\u003e, 8554-8559 (2001).\u003c/li\u003e\n\u003cli\u003eZou Y, Ma D, Wang Y. The PROTAC technology in drug development. \u003cem\u003eCell biochemistry and function\u003c/em\u003e \u003cstrong\u003e37\u003c/strong\u003e, 21-30 (2019).\u003c/li\u003e\n\u003cli\u003eHe M\u003cem\u003e, et al.\u003c/em\u003e PROTACs: great opportunities for academia and industry (an update from 2020 to 2021). \u003cem\u003eSignal transduction and targeted therapy\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 181 (2022).\u003c/li\u003e\n\u003cli\u003eNeklesa TK, Winkler JD, Crews CM. Targeted protein degradation by PROTACs. \u003cem\u003ePharmacology \u0026amp; therapeutics\u003c/em\u003e \u003cstrong\u003e174\u003c/strong\u003e, 138-144 (2017).\u003c/li\u003e\n\u003cli\u003ePettersson M, Crews CM. PROteolysis TArgeting Chimeras (PROTACs) \u0026mdash; Past, present and future. \u003cem\u003eDrug Discovery Today: Technologies\u003c/em\u003e \u003cstrong\u003e31\u003c/strong\u003e, 15-27 (2019).\u003c/li\u003e\n\u003cli\u003eToure M, Crews CM. Small-Molecule PROTACS: New Approaches to Protein Degradation. \u003cem\u003eAngewandte Chemie (International ed in English)\u003c/em\u003e \u003cstrong\u003e55\u003c/strong\u003e, 1966-1973 (2016).\u003c/li\u003e\n\u003cli\u003eSchneekloth AR, Pucheault M, Tae HS, Crews CM. Targeted intracellular protein degradation induced by a small molecule: En route to chemical proteomics. \u003cem\u003eBioorganic \u0026amp; Medicinal Chemistry Letters\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e, 5904-5908 (2008).\u003c/li\u003e\n\u003cli\u003eGalla MS, Sharma N, Mishra P, Shankaraiah N. Recent insights of PROTAC developments in inflammation-mediated and autoimmune targets: a critical review. \u003cem\u003eRSC medicinal chemistry\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 2585-2600 (2024).\u003c/li\u003e\n\u003cli\u003eAlugubelli YR\u003cem\u003e, et al.\u003c/em\u003e Discovery of First-in-Class PROTAC Degraders of SARS-CoV-2 Main Protease. \u003cem\u003ebioRxiv\u003c/em\u003e, (2023).\u003c/li\u003e\n\u003cli\u003eLi D, Yu D, Li Y, Yang R. A bibliometric analysis of PROTAC from 2001 to 2021. \u003cem\u003eEuropean journal of medicinal chemistry\u003c/em\u003e \u003cstrong\u003e244\u003c/strong\u003e, 114838 (2022).\u003c/li\u003e\n\u003cli\u003eNeklesa T\u003cem\u003e, et al.\u003c/em\u003e ARV-110: An oral androgen receptor PROTAC degrader for prostate cancer. \u003cem\u003eJournal of Clinical Oncology\u003c/em\u003e \u003cstrong\u003e37\u003c/strong\u003e, 259-259 (2019).\u003c/li\u003e\n\u003cli\u003eGao X\u003cem\u003e, et al.\u003c/em\u003e Phase 1/2 study of ARV-110, an androgen receptor (AR) PROTAC degrader, in metastatic castration-resistant prostate cancer (mCRPC). \u003cem\u003eJournal of Clinical Oncology\u003c/em\u003e \u003cstrong\u003e40\u003c/strong\u003e, 17-17 (2022).\u003c/li\u003e\n\u003cli\u003eB\u0026eacute;k\u0026eacute;s M, Langley DR, Crews CM. PROTAC targeted protein degraders: the past is prologue. \u003cem\u003eNature Reviews Drug Discovery\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 181-200 (2022).\u003c/li\u003e\n\u003cli\u003eChen W, Liu X, Zhang S, Chen S. Artificial intelligence for drug discovery: Resources, methods, and applications. \u003cem\u003eMolecular therapy Nucleic acids\u003c/em\u003e \u003cstrong\u003e31\u003c/strong\u003e, 691-702 (2023).\u003c/li\u003e\n\u003cli\u003eFleming N. How artificial intelligence is changing drug discovery. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e557\u003c/strong\u003e, S55-s57 (2018).\u003c/li\u003e\n\u003cli\u003eFrye L, Bhat S, Akinsanya K, Abel R. From computer-aided drug discovery to computer-driven drug discovery. \u003cem\u003eDrug discovery today Technologies\u003c/em\u003e \u003cstrong\u003e39\u003c/strong\u003e, 111-117 (2021).\u003c/li\u003e\n\u003cli\u003eMendez D\u003cem\u003e, et al.\u003c/em\u003e ChEMBL: towards direct deposition of bioassay data. \u003cem\u003eNucleic Acids Research\u003c/em\u003e \u003cstrong\u003e47\u003c/strong\u003e, D930-D940 (2019).\u003c/li\u003e\n\u003cli\u003eKnox C\u003cem\u003e, et al.\u003c/em\u003e DrugBank 6.0: the DrugBank Knowledgebase for 2024. \u003cem\u003eNUCLEIC ACIDS RESEARCH\u003c/em\u003e, (2023).\u003c/li\u003e\n\u003cli\u003eBerman HM\u003cem\u003e, et al.\u003c/em\u003e The Protein Data Bank. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 235-242 (2000).\u003c/li\u003e\n\u003cli\u003eKim S\u003cem\u003e, et al.\u003c/em\u003e PubChem in 2021: new data content and improved web interfaces. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e49\u003c/strong\u003e, D1388-d1395 (2021).\u003c/li\u003e\n\u003cli\u003eWei L\u003cem\u003e, et al.\u003c/em\u003e Hit Identification Driven by Combining Artificial Intelligence and Computational Chemistry Methods: A PI5P4K-\u0026beta; Case Study. \u003cem\u003eJ Chem Inf Model\u003c/em\u003e \u003cstrong\u003e63\u003c/strong\u003e, 5341-5355 (2023).\u003c/li\u003e\n\u003cli\u003eYou Y\u003cem\u003e, et al.\u003c/em\u003e Artificial intelligence in cancer target identification and drug discovery. \u003cem\u003eSignal transduction and targeted therapy\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 156 (2022).\u003c/li\u003e\n\u003cli\u003eWei M\u003cem\u003e, et al.\u003c/em\u003e HobPre: accurate prediction of human oral bioavailability for small molecules. \u003cem\u003eJournal of Cheminformatics\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 1 (2022).\u003c/li\u003e\n\u003cli\u003eBlaschke T\u003cem\u003e, et al.\u003c/em\u003e REINVENT 2.0: An AI Tool for De Novo Drug Design. \u003cem\u003eJ Chem Inf Model\u003c/em\u003e \u003cstrong\u003e60\u003c/strong\u003e, 5918-5922 (2020).\u003c/li\u003e\n\u003cli\u003eImrie F, Bradley AR, van der Schaar M, Deane CM. Deep Generative Models for 3D Linker Design. \u003cem\u003eJournal of Chemical Information and Modeling\u003c/em\u003e \u003cstrong\u003e60\u003c/strong\u003e, 1983-1995 (2020).\u003c/li\u003e\n\u003cli\u003eZheng S\u003cem\u003e, et al.\u003c/em\u003e Accelerated rational PROTAC design via deep learning and molecular simulations. \u003cem\u003eNature Machine Intelligence\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, 739-748 (2022).\u003c/li\u003e\n\u003cli\u003eWeng G\u003cem\u003e, et al.\u003c/em\u003e PROTAC-DB: an online database of PROTACs. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e49\u003c/strong\u003e, D1381-d1387 (2021).\u003c/li\u003e\n\u003cli\u003eWeng G\u003cem\u003e, et al.\u003c/em\u003e PROTAC-DB 2.0: an updated database of PROTACs. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, D1367-d1372 (2023).\u003c/li\u003e\n\u003cli\u003eLi F\u003cem\u003e, et al.\u003c/em\u003e DeepPROTACs is a deep learning-based targeted degradation predictor for PROTACs. \u003cem\u003eNature Communications\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 7133 (2022).\u003c/li\u003e\n\u003cli\u003eWeininger D. SMILES, a chemical language and information system. 1. introduction to methodology and encoding rules. \u003cstrong\u003e28\u003c/strong\u003e, 31\u0026ndash;36 (1988).\u003c/li\u003e\n\u003cli\u003eVaradi M\u003cem\u003e, et al.\u003c/em\u003e AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e50\u003c/strong\u003e, D439-d444 (2022).\u003c/li\u003e\n\u003cli\u003eJumper J\u003cem\u003e, et al.\u003c/em\u003e Highly accurate protein structure prediction with AlphaFold. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e596\u003c/strong\u003e, 583-589 (2021).\u003c/li\u003e\n\u003cli\u003eG L. RDKit: Open-source cheminformatics., (2022).\u003c/li\u003e\n\u003cli\u003eSastry GM, Adzhigirey M, Day T, Annabhimoju R, Sherman W. Protein and ligand preparation: parameters, protocols, and influence on virtual screening enrichments. \u003cem\u003eJournal of computer-aided molecular design\u003c/em\u003e \u003cstrong\u003e27\u003c/strong\u003e, 221-234 (2013).\u003c/li\u003e\n\u003cli\u003eLu C\u003cem\u003e, et al.\u003c/em\u003e OPLS4: Improving Force Field Accuracy on Challenging Regimes of Chemical Space. \u003cem\u003eJournal of Chemical Theory and Computation\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 4291-4300 (2021).\u003c/li\u003e\n\u003cli\u003eSchr\u0026ouml;dinger Release 2023-1: LigPrep, Schr\u0026ouml;dinger, LLC, New York, NY, 2023.\u003c/li\u003e\n\u003cli\u003eJacobson MP\u003cem\u003e, et al.\u003c/em\u003e A hierarchical approach to all-atom protein loop prediction. \u003cem\u003eProteins\u003c/em\u003e \u003cstrong\u003e55\u003c/strong\u003e, 351-367 (2004).\u003c/li\u003e\n\u003cli\u003eSchr\u0026ouml;dinger Release 2023-1: Prime, Schr\u0026ouml;dinger, LLC, New York, NY, 2023.\u003c/li\u003e\n\u003cli\u003eJacobson MP, Friesner RA, Xiang Z, Honig B. On the Role of the Crystal Environment in Determining Protein Side-chain Conformations. \u003cem\u003eJ Mol Biol\u003c/em\u003e \u003cstrong\u003e320\u003c/strong\u003e, 597-608 (2002).\u003c/li\u003e\n\u003cli\u003eWang C, Zhang Y, Deng J, Liang B, Xing D. Developments of PROTACs technology in immune-related diseases. \u003cem\u003eEuropean journal of medicinal chemistry\u003c/em\u003e \u003cstrong\u003e249\u003c/strong\u003e, 115127 (2023).\u003c/li\u003e\n\u003cli\u003eOrigin, Version 2020. OriginLab Corporation, Northampton, MA, USA.\u003c/li\u003e\n\u003cli\u003eAsh S, Cline MA, Homer RW, Hurst T, Smith GBJJoci, sciences c. SYBYL line notation (SLN): A versatile language for chemical structure representation. \u003cstrong\u003e37\u003c/strong\u003e, 71-79 (1997).\u003c/li\u003e\n\u003cli\u003eHomer RW, Swanson J, Jilek RJ, Hurst T, Clark RDJJoci, modeling. SYBYL line notation (SLN): a single notation to represent chemical structures, queries, reactions, and virtual libraries. \u003cstrong\u003e48\u003c/strong\u003e, 2294-2307 (2008).\u003c/li\u003e\n\u003cli\u003eO\u0026apos;Boyle NM, Banck M, James CA, Morley C, Vandermeersch T, Hutchison GR. Open Babel: An open chemical toolbox. \u003cem\u003eJournal of Cheminformatics\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 33 (2011).\u003c/li\u003e\n\u003cli\u003eSterling T, Irwin JJ. ZINC 15 \u0026ndash; Ligand Discovery for Everyone. \u003cem\u003eJournal of chemical information and modeling\u003c/em\u003e \u003cstrong\u003e55\u003c/strong\u003e, 2324-2337 (2015).\u003c/li\u003e\n\u003cli\u003eKipf TN, Welling MJapa. Semi-supervised classification with graph convolutional networks. (2016).\u003c/li\u003e\n\u003cli\u003eVelickovic P, Cucurull G, Casanova A, Romero A, Lio P, Bengio YJs. Graph attention networks. \u003cstrong\u003e1050\u003c/strong\u003e, 10-48550 (2017).\u003c/li\u003e\n\u003cli\u003ePaszke A\u003cem\u003e, et al.\u003c/em\u003e Pytorch: An imperative style, high-performance deep learning library. \u003cstrong\u003e32\u003c/strong\u003e, (2019).\u003c/li\u003e\n\u003cli\u003eFey M, Lenssen JEJapa. Fast graph representation learning with PyTorch Geometric. (2019).\u003c/li\u003e\n\u003cli\u003eSanchez-Lengeling B, Reif E, Pearce A, Wiltschko ABJD. A gentle introduction to graph neural networks. \u003cstrong\u003e6\u003c/strong\u003e, e33 (2021).\u003c/li\u003e\n\u003cli\u003eAbbasi K, Razzaghi P, Poso A, Amanlou M, Ghasemi JB, Masoudi-Nejad AJB. DeepCDA: deep cross-domain compound\u0026ndash;protein affinity prediction through LSTM and convolutional neural networks. \u003cstrong\u003e36\u003c/strong\u003e, 4633-4642 (2020).\u003c/li\u003e\n\u003cli\u003eKruse R, Mostaghim S, Borgelt C, Braune C, Steinbrecher M. Multi-layer Perceptrons. In: \u003cem\u003eComputational Intelligence: A Methodological Introduction\u003c/em\u003e (eds Kruse R, Mostaghim S, Borgelt C, Braune C, Steinbrecher M). Springer International Publishing (2022).\u003c/li\u003e\n\u003cli\u003eNair V, Hinton GE. Rectified linear units improve restricted boltzmann machines. In: \u003cem\u003eProceedings of the 27th international conference on machine learning (ICML-10)\u003c/em\u003e) (2010).\u003c/li\u003e\n\u003cli\u003eLi F, Hu Q, Zhou Y, Yang H, Bai F. DiffPROTACs is a deep learning-based generator for proteolysis targeting chimeras. \u003cem\u003eBriefings in bioinformatics\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, (2024).\u003c/li\u003e\n\u003cli\u003eZhao C, Zhang H, Zhang X, Liu M, Hu Z, Fan BJT. Application of support vector machine (SVM) for prediction toxic activity of different data sets. \u003cstrong\u003e217\u003c/strong\u003e, 105-119 (2006).\u003c/li\u003e\n\u003cli\u003eCano G\u003cem\u003e, et al.\u003c/em\u003e Automatic selection of molecular descriptors using random forest: Application to drug discovery. \u003cstrong\u003e72\u003c/strong\u003e, 151-159 (2017).\u003c/li\u003e\n\u003cli\u003ePattanaik L, Coley CWJC. Molecular representation: going long on fingerprints. \u003cstrong\u003e6\u003c/strong\u003e, 1204-1207 (2020).\u003c/li\u003e\n\u003cli\u003eCereto-Massagu\u0026eacute; A, Ojeda MJ, Valls C, Mulero M, Garcia-Vallv\u0026eacute; S, Pujadas GJM. Molecular fingerprint similarity search in virtual screening. \u003cstrong\u003e71\u003c/strong\u003e, 58-63 (2015).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-5021266/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5021266/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eProteolysis-targeting chimera(PROTAC), is an emerging therapeutic modality with the potential to treat disease-driven proteins \u0026nbsp;that emerged in the new century . Since the first PROTAC drug, ARV-110, entered Phase II clinical trials in 2019, research in this field has surged. The growing body of data has created opportunities for developing artificial intelligence (AI) methods in PROTAC design. In this study, we present the largest PROTAC database to date, named as PROTAC-Databank. Each molecule in the database is labeled with manually reviewed and uniformly formatted degradation efficiency data and protein-ligand complex structure information, providing a valuable resource for AI-driven PROTAC modeling or design. Additionally, we have enhanced the prediction accuracy of the DeepPROTACs, a widely used tool for predicting PROTAC degradation efficiency developed by our group two years ago. The update version, \u0026nbsp;DeepPROTACs 2.0, which is developed based on a complete new deep learning framework. Compared with the previous version, DeepPROTACs 2.0 shows a significant improvement in prediction accuracy, increasing from 77.15% to 83.45%. The introduction of PROTAC-Databank and DeepPROTACs 2.0 offers optimized tools for PROTAC design, streamlining the screening process and reducing both time and costs.\u003c/p\u003e","manuscriptTitle":"PROTAC-Databank: The Present Largest Integrated Resource of PROTACs, Enabling the Enhanced DeepPROTACs 2.0 for Degradation Prediction","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-09-16 06:51:30","doi":"10.21203/rs.3.rs-5021266/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":"edf32685-0ad8-453a-93a2-77d08c2e5f36","owner":[],"postedDate":"September 16th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":37578455,"name":"Biological sciences/Computational biology and bioinformatics/Databases"},{"id":37578456,"name":"Biological sciences/Computational biology and bioinformatics/Computational models"}],"tags":[],"updatedAt":"2024-11-14T13:40:47+00:00","versionOfRecord":[],"versionCreatedAt":"2024-09-16 06:51:30","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5021266","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5021266","identity":"rs-5021266","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}