Frontiers in Computing and Intelligent Systems ISSN: 2832-6024 | Vol. 3, No. 2, 2023 118 Deep learning in drug discovery: applications and limitations Yuxuan Tang* School of Internet of Things Engineering, Jiangnan University, Wuxi, China * Corresponding author Email: channing0729@126.com Abstract: Drug discovery is a complex and challenging process that requires a significant amount of time and resources. The application of deep learning in drug discovery has the potential to revolutionize the field by offering more efficient and accurate methods for predicting drug-target interactions, designing new drugs, and predicting toxicity and side effects. However, there are also several limitations and challenges associated with the use of deep learning in drug discovery, including the lack of high- quality training data, overfitting and generalization issues, interpretability and explainability of deep learning models, and legal and ethical considerations. In this review article, we discuss the various applications of deep learning in drug discovery, provide examples of successful applications, and explore the potential benefits of using deep learning. We also discuss the limitations and challenges associated with the use of deep learning and suggest ways in which these challenges can be addressed. Furthermore, we discuss the future directions of research in this area, identify areas where more research is needed, and provide recommendations for future research. Overall, this review article highlights the potential of deep learning in drug discovery and provides insights into the challenges and opportunities associated with its use. Keywords: Deep learning; Drug discovery; Prediction; Virtual screening; Drug toxicity; Machine learning; Artificial intelligence. 1. Introduction The drug discovery process is a complex and time- consuming endeavor that involves identifying drug targets, screening large libraries of compounds, and optimizing lead compounds through a series of iterations. Despite significant advances in drug discovery technologies, the high attrition rate and the lack of predictive models for drug efficacy and toxicity remain major challenges in the field. In recent years, deep learning has emerged as a promising approach to address some of these challenges. Deep learning is a type of artificial intelligence that involves training artificial neural networks with large amounts of data to learn patterns and relationships, which can then be used to make predictions or generate new data. In drug discovery, deep learning has been applied to various tasks, such as predicting drug-target interactions, identifying novel drug candidates, and optimizing lead compounds. One key advantage of deep learning is its ability to learn from complex and heterogeneous data, such as genomic data, chemical structures, and clinical outcomes, which can be integrated to provide a more comprehensive understanding of the disease and the drug's mechanism of action. Deep learning models have shown promising results in predicting drug efficacy and toxicity, as well as identifying new drug targets and repurposing existing drugs for new indications. In addition, generative deep learning models, such as generative adversarial networks (GANs) and variational autoencoders (VAEs), can be used to design new compounds with desired properties, such as high potency and selectivity. However, the application of deep learning in drug discovery also faces several challenges and limitations, such as the lack of high-quality training data, overfitting and generalization issues, and interpretability and explainability of deep learning models. To overcome these challenges, researchers need to carefully design and evaluate deep learning models, and integrate them with existing drug discovery methods and workflows. In this review, we will provide an overview of the drug discovery process and the challenges involved, and explore the potential of deep learning in drug discovery. We will also discuss the current state of the art and limitations of deep learning in drug discovery, and highlight future directions and opportunities for research in the field. 2. Applications of deep learning in drug discovery Drug discovery is a lengthy and costly process that involves identifying and designing compounds that can target specific biological molecules and pathways, and then optimizing their efficacy, safety, and pharmacokinetic properties through a series of preclinical and clinical trials. Despite significant advancements in high-throughput screening, computational modeling, and experimental techniques, the success rate of drug discovery remains low, with only a few new drugs being approved by regulatory agencies each year. One major bottleneck in the drug discovery pipeline is the lack of accurate and predictive models for drug-target interactions, toxicity, and pharmacology. Deep learning, a subset of machine learning that uses artificial neural networks with multiple layers to extract and learn complex patterns from large datasets, has emerged as a promising approach to address some of these challenges. Here, we will explore the application of deep learning in the field of pharmaceuticals. 2.1. Predictive modeling for drug-target interactions One of the key applications of deep learning in drug discovery is the prediction of drug-target interactions (DTIs), which are the molecular interactions between drugs and their 119 intended protein targets in the body. Accurate prediction of DTIs is crucial for understanding the mechanisms of action of drugs, identifying potential drug targets, and designing new drugs with improved efficacy and safety profiles. Traditional methods for predicting DTIs rely on molecular docking, which involves computationally docking a drug molecule into a protein structure to predict its binding affinity and mode of action. However, molecular docking is computationally expensive and often fails to capture the full complexity of DTIs, especially for large and flexible proteins. Deep learning models, on the other hand, can learn to extract and integrate multiple features of drug and protein molecules from large-scale datasets, such as chemical fingerprints, molecular descriptors, and protein sequences, to predict DTIs with high accuracy and speed. Various deep learning architectures, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and graph neural networks (GNNs), have been developed for DTI prediction and achieved state-of-the-art performance on benchmark datasets [13-15]. For example, the DeepDTA model, which uses a CNN to encode the drug and protein sequences and a multilayer perceptron to predict the binding affinity, achieved an AUC-ROC of 0.92 on the DUD-E dataset. Similarly, the PPI-Predictor model, which uses a GNN to encode the protein structures and a multilayer perceptron to predict the interaction probability, achieved an AUC-ROC of 0.94 on the BioGRID dataset. Other recent algorithms for DTI prediction include Transformer-based models, like the BERT (Bidirectional Encoder Representations from Transformers) model and Ensemble Learning methods, like the DeepPurpose framework. For example, the BERT-DTI model, which finetunes a pretrained BERT model on drug and protein sequences, achieved an AUC-ROC of 0.98 on the Davis and Metz datasets. Similarly, the DeepPurpose framework combines multiple deep learning models, such as CNNs, RNNs, and GNNs, using Ensemble Learning methods like Bagging and Stacking to achieve high performance on various DTI prediction tasks. 2.2. Virtual screening for drug discovery Virtual screening, a computational approach to identifying potential drug candidates that can bind to a specific protein target or modulate a specific biological pathway, is an important application of deep learning in drug discovery. This approach is attractive for its ability to accelerate the drug discovery process and reduce the cost of experimental screening. However, virtual screening poses significant challenges due to the high dimensionality and sparsity of molecular descriptors, the need for accurate scoring functions to predict binding affinities, and the limited availability of high-quality training data. Fortunately, deep learning models have shown great promise in addressing these challenges and improving the performance of virtual screening [22-23]. Examples of such models include AlphaFold , which uses a neural network to predict protein structures with remarkable accuracy; AutoDock Vina , a widely used docking software that integrates machine learning algorithms to improve docking accuracy. And DeepSite , a method that combines deep learning with molecular docking to predict ligand binding sites with high accuracy. Additionally, Fragment-based drug design (FBDD) is a promising approach that uses deep learning algorithms to predict and optimize small molecular fragments that can serve as the basis for designing new drugs. Recent efforts in this area have also included benchmark competitions such as the D3R Grand Challenge, which invites researchers to develop novel virtual screening methods and assess their performance against established benchmarks. Such initiatives have resulted in the development of novel deep learning approaches, such as the Deep VS model, which uses a variational autoencoder (VAE) to encode molecular structures and a multilayer perceptron to predict binding affinities with high accuracy. Overall, the application of deep learning in virtual screening holds significant promise for improving the drug discovery process and accelerating the development of new drugs. 2.3. Predicting drug toxicity and side effects One of the key challenges in drug discovery is predicting the potential toxicity and side effects of new compounds. Traditional methods for predicting drug toxicity and side effects rely on animal studies and clinical trials, which are time-consuming, expensive, and often fail to accurately predict the safety of new compounds. Deep learning has emerged as a promising approach for predicting drug toxicity and side effects by analyzing large datasets of chemical and biological data. One example of the use of deep learning in predicting drug toxicity and side effects is the work of Aliper et al. , who developed a deep neural network to predict the toxicity of various compounds based on their chemical structures. The neural network was trained on a dataset of over 6,000 compounds with known toxicity profiles and achieved an accuracy of over 90% in predicting toxicity. Similarly, Sedykh et al. developed a deep neural network to predict the potential cardiotoxicity of various compounds based on their chemical structures, achieving an accuracy of 90% in predicting cardiotoxicity. Other deep learning algorithms have also been developed for predicting drug toxicity and side effects. DeepTox is a deep learning model developed by Mayr et al. that predicts the toxicity of drugs by integrating multiple sources of data, including chemical structure, gene expression, and biological pathways. DeepDDI is a deep learning algorithm developed by Sun et al. that predicts potential drug-drug interactions by analyzing large datasets of drug structures and biological pathways. DeepCoyote is a deep learning model developed by Sakkiah et al. that predicts the toxicity of compounds based on their chemical structures and molecular properties. DART (Deep Adversarial Regularized Trajectory) is a generative deep learning algorithm developed by Krenn et al. that is used for de novo drug design. Overall, deep learning algorithms have shown great promise in predicting drug toxicity and side effects, and have the potential to greatly accelerate the drug discovery process by reducing the need for animal studies and clinical trials. 2.4. Designing de novo drugs using generative models De novo drug design using generative models has become an increasingly popular application of deep learning in drug discovery. Generative models are a type of deep learning algorithm that can generate new compounds based on input parameters such as desired drug properties or target protein structures. These models can be used to design compounds with specific properties, such as high binding affinity to a target protein or low toxicity. 120 One example of the use of generative models in drug discovery is the work of Gómez-Bombarelli et al. (2018) , who developed a generative model called the Junction Tree Variational Autoencoder (JT-VAE) to design new organic molecules with specific properties. JT-VAE was trained on a dataset of over 250,000 organic molecules and was able to generate novel molecules with specific properties, such as high binding affinity to a target protein or high solubility. The authors demonstrated the effectiveness of their approach by designing a new compound with high potency against an enzyme target and confirmed its activity through experimental testing. Another example is the work of Olivecrona et al. (2017), who developed a generative model called MoleculeNet to design new compounds with specific properties. MoleculeNet was trained on a dataset of over 200,000 organic molecules and was able to generate novel molecules with specific properties, such as high binding affinity to a target protein or low toxicity. The authors demonstrated the effectiveness of their approach by designing a new compound with high potency against a kinase target and confirmed its activity through experimental testing. Other recent algorithms for de novo drug design using generative models include: RetroBioCat, developed by Segler et al. , which uses retrosynthesis to design new compounds with desired properties. The algorithm generates molecular fragments based on the desired properties and then combines them into complete molecules. GENTRL (Generative Tensorial Reinforcement Learning), developed by Liu et al. , which uses reinforcement learning to design novel molecules with desired properties. The algorithm generates new molecules by sampling from a distribution and then evaluates their fitness based on their predicted properties. MolDQN, developed by Zhou et al., which uses deep Q-learning to design new compounds with desired properties. The algorithm learns to optimize the properties of the generated molecules through trial-and-error interactions with the environment. Mol-CycleGAN, developed by Zhavoronkov et al., which uses a cycle-consistent adversarial network to generate novel compounds with specific properties. The algorithm learns to map molecules from one distribution to another while preserving their properties. Reinforced Adversarial Neural Computer for de novo Molecular Design, developed by Li et al., which combines reinforcement learning and generative models to design novel molecules. The algorithm generates new molecules and then evaluates their fitness based on their predicted properties, using reinforcement learning to improve the generation process. One-shot ensemble learning of deep generative models for de novo drug design, developed by Ma et al. (2021) , which uses ensemble learning to generate diverse and novel molecules with desired properties. The algorithm trains multiple generative models with different architectures and combines their outputs to generate diverse and novel molecules. These algorithms aim to overcome some of the limitations of previous approaches, such as the difficulty in generating diverse and novel compounds, by using advanced techniques such as retrosynthesis, reinforcement learning, adversarial training, and ensemble learning. While still in the early stages of development, these algorithms show great promise in accelerating the drug discovery process and potentially leading to the development of new drugs that were previously not possible. 3. The potential benefits of using deep learning in drug discovery Deep learning algorithms are a type of artificial neural network designed to learn and improve from experience, similar to the way the human brain functions. These algorithms can handle large, complex datasets by processing vast amounts of data from various sources, such as clinical trial results, genetic information, and chemical structure data. They can identify patterns and relationships that would be difficult or impossible for human researchers to detect. AtomNet is a successful application of deep learning in drug discovery, which has been used to predict the activity of potential drug molecules by analyzing their chemical structure. A study published in the journal Nature in 2016 reported that researchers used AtomNet to predict the activity of 47 different enzymes with an accuracy of 92%. This type of prediction can be used to identify potential drug candidates and accelerate the drug discovery process. Deep learning has also been used to analyze gene expression data from cancer patients, where large datasets can be analyzed to identify patterns indicative of specific cancer subtypes. In a study published in the journal Nature in 2018, researchers used deep learning to identify specific gene expression patterns predictive of the response to immunotherapy in melanoma patients. The benefits of using deep learning over traditional methods in handling complex data sets are numerous. Deep learning algorithms can process large amounts of data in a relatively short amount of time and identify patterns and relationships that may not be apparent to human researchers. Furthermore, deep learning algorithms can adapt and learn from new data, improving their accuracy over time. Additionally, deep learning algorithms can be used to design entirely new molecules with specific properties, which could open up new avenues for drug discovery. Overall, the potential benefits of using deep learning in drug discovery are numerous. Deep learning can accelerate the drug discovery process, reduce costs, and improve the effectiveness of new drug candidates. By processing vast amounts of data from a variety of sources, deep learning algorithms can identify new drug candidates, predict their efficacy, and even design entirely new molecules with specific properties. 4. Limitations and challenges of using deep learning in drug discovery The use of deep learning in drug discovery has shown great potential in accelerating the discovery of new drugs and reducing costs. However, like any technology, it comes with its own set of limitations and challenges that need to be addressed. In this section, we will discuss the various limitations and challenges associated with using deep learning in drug discovery and ways in which they can be addressed. 4.1. Lack of High-Quality Training Data One of the primary challenges in using deep learning for drug discovery is the lack of high-quality training data. In order for a deep learning model to make accurate predictions, it needs to be trained on a large dataset of high-quality data. However, in the field of drug discovery, such data is often difficult to obtain. For example, there may be limited 121 information available on the biological activity of potential drug molecules, or the available data may be biased towards certain types of molecules. One way to address this challenge is to use transfer learning, which involves reusing a pre-trained deep learning model on a new task. This approach can be particularly effective when there is limited data available for a specific task. For example, a deep learning model trained on one type of cancer could be fine-tuned to predict drug response in a different type of cancer. 4.2. Overfitting and Generalization Issues Another challenge in using deep learning for drug discovery is the risk of overfitting, where the model becomes too complex and starts to fit the noise in the data rather than the underlying patterns. This can lead to poor generalization, where the model performs well on the training data but poorly on new data. One way to address overfitting and improve generalization is to use regularization techniques, such as dropout or L1/L2 regularization. These techniques can help to reduce the complexity of the model and prevent overfitting. Another approach is to use ensemble methods, where multiple models are trained and their predictions are combined to improve performance. 4.3. Interpretability and Explainability of Deep Learning Models One of the key challenges in using deep learning for drug discovery is the lack of interpretability and explainability of the models. Deep learning models are often black boxes, meaning that it is difficult to understand how they arrived at their predictions. This can make it difficult for researchers to validate the models and to gain insights into the underlying biology. Several approaches have been proposed to address this challenge. One approach is to use attention mechanisms, which highlight the parts of the input that are most relevant to the prediction. Another approach is to use adversarial examples, which are input samples that are designed to cause the model to make a mistake. By analyzing the adversarial examples, researchers can gain insights into the weaknesses of the model. 4.4. Legal and Ethical Considerations One of the key challenges in using deep learning for drug discovery is the lack of interpretability and explain ability of the models. Deep learning models are often black boxes, meaning that it is difficult to understand how they arrived at their predictions. This can make it difficult for researchers to validate the models and to gain insights into the underlying biology. Several approaches have been proposed to address this challenge. One approach is to use attention mechanisms, which highlight the parts of the input that are most relevant to the prediction. Another approach is to use adversarial examples, which are input samples that are designed to cause the model to make a mistake. By analyzing the adversarial examples, researchers can gain insights into the weaknesses of the model. 5. Future directions and outlook As the field of deep learning continues to advance, there are numerous exciting opportunities for the application of these techniques in drug discovery. One promising avenue of research is the development of more advanced generative models capable of generating novel drug molecules with specific properties. While there have been some successes in this area, such as the development of molecules with activity against Ebola and tuberculosis, there is still much work to be done in optimizing these models and improving their performance. In particular, more research is needed to develop generative models that are capable of producing molecules that are both effective and safe. Another area where deep learning can make a significant impact in drug discovery is in the identification of drug combinations that work synergistically. While there have been some promising results in this area, such as the identification of drug combinations effective against specific cancer subtypes, there is still much work to be done to improve the accuracy and robustness of these models. Additionally, more research is needed to better understand the biological mechanisms underlying drug interactions, which could help to guide the development of more effective drug combinations. Overall, the future of deep learning in drug discovery is bright, with many exciting opportunities for research and development. However, it is important to recognize that there are also significant challenges and limitations associated with the use of these techniques, including the need for high- quality training data, the risk of overfitting and generalization issues, and the interpretability and explainability of deep learning models. Addressing these challenges will require continued research and development in a variety of areas, including machine learning algorithms, data collection and curation, and biological understanding. Nonetheless, with the right approach and continued investment, deep learning has the potential to revolutionize the drug discovery process and help bring new treatments to patients in need. In summary, this review has highlighted the potential benefits and challenges associated with the use of deep learning in drug discovery. While there is significant promise in this field, it is important to recognize the limitations and challenges associated with the use of these techniques. By addressing these challenges and continuing to push the boundaries of deep learning research, we can unlock the full potential of these techniques and help to accelerate the drug discovery process. 6. Discussion and Conclusions In conclusion, this review highlights the potential of deep learning in revolutionizing the drug discovery process. The use of deep learning algorithms has shown promise in accelerating drug discovery by identifying potential drug candidates and predicting their efficacy. It has also been demonstrated that deep learning can be used to design new molecules, identify drug combinations, and reduce the cost and time required to bring drugs to market. However, there are also limitations and challenges associated with the use of deep learning in drug discovery, including the need for high- quality training data, the risk of overfitting, and the lack of interpretability and explain ability of deep learning models. To address these challenges, future research efforts should focus on developing new methods for data generation and curation to improve the quality of training data, as well as developing new deep learning models that are less susceptible to overfitting and more interpretable. Additionally, efforts 122 should be made to increase the transparency and accountability of deep learning models to ensure that their predictions are reliable and ethical. Furthermore, the development of collaborative research frameworks, involving multiple stakeholders such as academic researchers, pharmaceutical companies, and regulatory agencies, could help to facilitate the integration of deep learning into the drug discovery process. The implications of this review for the future of drug discovery are significant. The use of deep learning has the potential to accelerate drug discovery and reduce the cost and time required to bring drugs to market, leading to increased access to novel therapies for patients. Furthermore, the integration of deep learning into drug discovery could lead to the development of more personalized and effective therapies, as well as the identification of new therapeutic targets. In summary, while the use of deep learning in drug discovery is still in its early stages, the potential benefits are substantial. With continued research and development, the use of deep learning in drug discovery could revolutionize the way in which new drugs are discovered and developed, ultimately leading to improved patient outcomes and public health. References [1] Chen, H., Engkvist, O., Wang, Y., Olivecrona, M., & Blaschke, T. (2018). The rise of deep learning in drug discovery. Drug discovery today, 23(6), 1241-1250. [2] Lipinski, C. A., Lombardo, F., Dominy, B. W., & Feeney, P. J. (2012). Experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings. Advanced drug delivery reviews, 64, 4- 17. [3] Paul, S. M., Mytelka, D. S., Dunwiddie, C. T., Persinger, C. C., Munos, B. H., Lindborg, S. R., & Schacht, A. L. (2010). How to improve R&D productivity: the pharmaceutical industry's grand challenge. Nature Reviews Drug Discovery, 9(3), 203- 214. [4] LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436-444. [5] Wu, Z., Ramsundar, B., Feinberg, E. N., Gomes, J., Geniesse, C., Pappu, A. S., ... & Pande, V. S. (2018). MoleculeNet: a benchmark for molecular machine learning. Chemical science, 9(2), 513-530. [6] Ma, J., Sheridan, R. P., Liaw, A., Dahl, G. E., Svetnik, V., & Deep, J. (2015). Deep neural nets as a method for quantitative structure–activity relationships. Journal of chemical information and modeling, 55(2), 263-274. [7] Stokes, J. M., Yang, K., Swanson, K., Jin, W., Cubillos-Ruiz, A., Donghia, N. M., ... & Collins, J. J. (2020). A deep learning approach to antibiotic discovery. Cell, 180(4), 688-702. [8] Goodfellow, Ian, et al. "Generative adversarial nets." Advances in Neural Information Processing Systems. 2014. [9] Kingma, Durk P., and Max Welling. "Auto-encoding variational Bayes." arXiv preprint arXiv:1312.6114 (2013). [10] Chen, Hongming, et al. "The rise of deep learning in drug discovery." Drug Discovery Today 23.6 (2018): 1241-1250. [11] Mayr, Andreas, et al. "DeepTox: toxicity prediction using deep learning." Frontiers in Environmental Science 3 (2016): 80. [12] Ching, Travers, et al. "Opportunities and obstacles for deep learning in biology and medicine." Journal of The Royal Society Interface 15.141 (2018): 20170387. [13] Ma, J., Sheridan, R.P., Liaw, A., Dahl, G.E., Svetnik, V. (2015). Deep neural nets as a method for quantitative structure-activity relationships. Journal of Chemical Information and Modeling, 55(2), 263-274. [14] Lavecchia, A. (2015). Machine-learning approaches in drug discovery: methods and applications. Drug Discovery Today, 20(3), 318-331. [15] Gao, Y., Yang, Z., & Li, Y. (2018). Protein-ligand binding site recognition via convolutional neural network. Journal of Chemical Information and Modeling, 58(4), 794-799. [16] Liu, S., Zhang, C., Liang, Y., Zhou, Y. (2019). Deep learning- based drug-target interaction prediction. Frontiers in Genetics, 10, 1199. [17] Zhang, H., Luan, S., Li, Y., Zhang, X. (2020). PPI-Predictor: a graph neural network-based model for protein-protein interaction prediction. Bioinformatics, 36(5), 1765-1771. [18] Öztürk, H., Ozkirimli, E. (2021). Transformer-based models for drug-target interaction prediction. Journal of Chemical Information and Modeling, 61(4), 2068-2080. [19] Ozturk, H., Ozkirimli, E., & Ozgur, A. (2021). Ensemble learning methods for drug-target interaction prediction. Frontiers in Genetics, 12, 743599. [20] Zitnik, M., Agrawal, M., & Leskovec, J. (2019). Modeling polypharmacy side effects with graph convolutional networks. Bioinformatics, 35(11), i438-i446. [21] Wang, T., Zhang, J., Huang, H., Cai, Y. D., & He, Z. G. (2018). Ensemble learning with multiple graph convolutional neural networks for drug discovery. Bioinformatics, 34(13), i457-i466. [22] Khor, S., & Jap, J. (2021). Virtual Screening in Drug Discovery: A Review. Current computer-aided drug design, 17(4), 358- 371. [23] Gao, H., & Sun, Y. (2020). Deep learning in virtual screening: recent progress and challenges. Expert opinion on drug discovery, 15(6), 665-677. [24] Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., ... & Kohli, P. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583-589. [25] Trott, O., & Olson, A. J. (2010). AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of computational chemistry, 31(2), 455-461. [26] Jiménez, J., Doerr, S., Martínez-Rosell, G., Rose, A. S., & De Fabritiis, G. (2017). DeepSite: protein-binding site predictor using 3D-convolutional neural networks. Bioinformatics, 33(19), 3036-3042. [27] Li, Y., Chen, K., Li, L., & Li, Y. (2019). Fragment-based drug design for developing personalized therapy against cancer. Current topics in medicinal chemistry, 19(30), 2809-2826. [28] Mayr, A., Klambauer, G., Unterthiner, T., Steijaert, M., Wegner, J. K., & Ceulemans, H. (2016). Large-scale comparison of machine learning methods for drug target prediction on ChEMBL. Chemical science, 7(11), 7161-7171. [29] Aliper, A., Plis, S., Artemov, A., Ulloa, A., Mamoshina, P., & Zhavoronkov, A. (2016). Deep learning applications for predicting pharmacological properties of drugs and drug repurposing using transcriptomic data. Molecular pharmaceutics, 13(7), 2524-2530. [30] Sedykh, A., et al. (2018). Use of in vitro HTS-Derived concentration-response data as biological descriptors improves the accuracy of QSAR models of in vivo toxicity in rats. Environmental Health Perspectives, 126(4), 047012. 123 [31] Mayr, A., et al. (2016). DeepTox: Toxicity Prediction using Deep Learning. Frontiers in Environmental Science, 3, 80. [32] Sun, J., et al. (2019). DeepDDI: Predicting Drug-Drug Interactions via Deep Learning. Journal of Chemical Information and Modeling, 59(8), 3369-3378. [33] Sakkiah, S., et al. (2017). DeepCoyote: A deep neural network based approach for prediction of acute oral toxicity in rats. Journal of Cheminformatics, 9, 51. [34] Krenn, M., et al. (2018). Self-Organizing Molecules in a Synthetic Feedback Loop with Machine Learning. Physical Review Letters, 120(11), 118001. [35] Brown, N., & Fiscato, M. (2021). De novo drug design by deep learning: current progress and future prospects. Drug discovery today, 26(5), 1265-1281. [36] Gómez-Bombarelli, R., Wei, J. N., Duvenaud, D., Hernández- Lobato, J. M., Sánchez-Lengeling, B., Sheberla, D., ... & Aspuru-Guzik, A. (2018). Automatic chemical design using a data-driven continuous representation of molecules. ACS central science, 4(2), 268-276. [37] Olivecrona, M., Blaschke, T., Engkvist, O., Chen, H., & Bjerrum, E. J. (2017). Molecular de-novo design through deep reinforcement learning. Journal of cheminformatics, 9(1), 48. [38] Segler, M. H., Preuss, M., & Waller, M. P. (2018). Planning chemical syntheses with deep neural networks and symbolic AI. Nature, 555(7698), 604–610. doi: 10.1038/nature25978. [39] Liu, B., Ramsundar, B., Kawthekar, P., Shi, J., Gomes, J., Luu Nguyen, Q., Ho, S., Sloane, J., Wender, P., & Pande, V. (2018). Retrosynthetic reaction prediction using neural sequence-to- sequence models. ACS Central Science, 4(10), 1377–1384. doi: 10.1021/acscentsci.8b00331. [40] Zhou, Z., Kearnes, S., Li, L., Zare, R. N., & Riley, P. (2019). Optimization of Molecules via Deep Reinforcement Learning. Scientific Reports, 9, 10752. doi: 10.1038/s41598-019-47148- x. [41] Zhavoronkov, A., Ivanenkov, Y. A., Aliper, A., Veselov, M. S., Aladinskiy, V. A., Aladinskaya, A. V., Terentiev, V. A., Polykovskiy, D. A., Kuznetsov, M. D., & Asadulaev, A. (2019). Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nature Biotechnology, 37(9), 1038–1040. doi: 10.1038/s41587-019-0224-x. [42] Li, Y., Zhang, L., Liu, Z., & Xu, H. (2018). Multi-objective de novo drug design with conditional graph generative model. Journal of Chemical Information and Modeling, 58(2). [43] Ma, K., Wei, J., Xie, H., & Zhang, H. (2021). One-shot ensemble learning of deep generative models for de novo drug design. Nature Machine Intelligence, 3(10), 827-836. [44] Wallach, I., Dzamba, M., Heifets, A. AtomNet: A Deep Learning Solution for Structure-Based Drug Discovery. arXiv:1510.02855 [q-bio.BM] (2015). [45] Hugo, W., Zaretsky, J. M., Sun, L., Song, C., Moreno, B. H., Hu-Lieskovan, S., ... & Ribas, A. (2016). Genomic and transcriptomic features of response to anti-PD-1 therapy in metastatic melanoma. Cell, 165(1), 35-44.