Trends in Deep Learning for Property-driven Drug Design

Abstract

: It is more pressing than ever to reduce the time and costs for developing lead compounds in the pharmaceutical industry. The co-occurrence of advances in high-throughput screening and the rise of deep learning (DL) have enabled the development of large-scale multimodal predictive models for virtual drug screening. Recently, deep generative models have emerged as a powerful tool for exploring the chemical space and raising hopes to expedite the drug discovery process. Following this progress in chemocentric approaches for generative chemistry, the next challenge is to build multimodal conditional generative models that leverage disparate knowledge sources when biochemical mapping properties to target structures. Here, we call the community to bridge drug discovery more closely with systems biology when designing deep generative models. Complementing the plethora of reviews on the role of DL in chemoinformatics, we herein specifically focus on the interface of predictive and generative modeling for drug discovery. Through a systematic publication keyword search on PubMed and a selection of preprint servers (arXiv, biorXiv, chemRxiv, and medRxiv), we quantify trends in the field and find that molecular graphs and VAEs have become the most widely adopted molecular representations and architectures in generative models, respectively. We discuss progress on DL for toxicity, drug-target affinity, and drug sensitivity prediction and specifically focus on conditional molecular generative models that encompass multimodal prediction models. Moreover, we outline prospects in the field and identify challenges such as the integration of deep learning systems into experimental workflows in a closed-loop manner or the adoption of federated machine learning techniques to overcome data sharing barriers. Other challenges include, but are not limited to interpretability in generative models, more sophisticated metrics for the evaluation of molecular generative models, and, following up on that, community-accepted benchmarks for both multimodal drug property prediction and property-driven molecular design.

Unsupervised Learning in Drug Design from Self-Organization to Deep Chemistry

“…The potential profits of DL in rational drug discovery were recently reviewed in the references [ 40 , 41 , 42 , 43 ]. Multilayer structures enable the extraction of cascade features that work with nonlinear functions [ 2 ].…”

“…The DL lexicon uses the term generative models (generative chemistry) for unsupervised algorithms to stress the difference between the classical design based on the local domain molecular exploration vs. the DL systematic continuous screening. Born and Manica [ 42 ] predicted that multimodal deep learning chemistry using disparate sources to generate molecules would be the next challenge in DL in the near future. The Variational Autoencoder (VAE) method [ 44 ] was developed as an algorithm to learn continuous molecular representations.…”

A Systematic Review of Deep Learning Methodologies Used in the Drug Discovery Process with Emphasis on In Vivo Validation

“…Deep learning can be employed for the prediction of drug–target interactions (DTIs), de novo molecular design, synthesis prediction, etc. [ 14 , 15 , 16 ]. While several studies have reviewed the contribution of deep learning in drug development, in vivo evaluation of the published algorithms is limited.…”

Active Site Sequence Representations of Human Kinases Outperform Full Sequence Representations for Affinity Prediction and Inhibitor Generation: 3D Effects in a 1D Model

“…54 In that work, deep learning methods discovered potential DDR1 inhibitors of which two were found active in vitro and one even in vivo. 54 The most common type of deep generative model for molecular design is variational autoencoders (VAEs), 55 which can be seen as a global search method in the chemical space. VAEs have been coupled with Bayesian optimization (BO) principles by means of Gaussian processes (GPs) to optimize chemocentric properties such as drug-likeness.…”

“…In that work, deep learning methods discovered potential DDR1 inhibitors of which two were found active in vitro and one even in vivo . The most common type of deep generative model for molecular design is variational autoencoders (VAEs), which can be seen as a global search method in the chemical space. VAEs have been coupled with Bayesian optimization (BO) principles by means of Gaussian processes (GPs) to optimize chemocentric properties such as drug-likeness. , While GP optimization allows us to maximize computationally costly functions efficiently, no previous work has, according to the best of our knowledge, incorporated an additional modality (like proteins) into the evaluation function and thus optimized a more complex, biochemical property.…”