Combining Cloud-Based Free-Energy Calculations, Synthetically Aware Enumerations, and Goal-Directed Generative Machine Learning for Rapid Large-Scale Chemical Exploration and Optimization

Abstract

The hit identification process usually involves the profiling of millions to more recently billions of compounds either via traditional experimental high-throughput screens (HTS) or computational virtual high-throughput screens (vHTS). We have previously demonstrated that, by coupling reaction-based enumeration, active learning, and free energy calculations, a similarly largescale exploration of chemical space can be extended to the hit-tolead process. In this work, we augment that approach by coupling large scale enumeration and cloud-based free energy perturbation (FEP) profiling with goal-directed generative machine learning, which results in a higher enrichment of potent ideas compared to large scale enumeration alone, while simultaneously staying within the bounds of predefined drug-like property space. We can achieve this by building the molecular distribution for generative machine learning from the PathFinder rules-based enumeration and optimizing for a weighted sum QSAR-based multiparameter optimization function. We examine the utility of this combined approach by designing potent inhibitors of cyclin-dependent kinase 2 (CDK2) and demonstrate a coupled workflow that can (1) provide a 6.4-fold enrichment improvement in identifying <10 nM compounds over random selection and a 1.5-fold enrichment in identifying <10 nM compounds over our previous method, (2) rapidly explore relevant chemical space outside the bounds of commercial reagents, (3) use generative ML approaches to "learn" the SAR from large scale in silico enumerations and generate novel idea molecules for a flexible receptor site that are both potent and within relevant physicochemical space, and (4) produce over 3 000 000 idea molecules and run 1935 FEP simulations, identifying 69 ideas with a predicted IC 50 < 10 nM and 358 ideas with a predicted IC 50 < 100 nM. The reported data suggest combining both reaction-based and generative machine learning for ideation results in a higher enrichment of potent compounds over previously described approaches and has the potential to rapidly accelerate the discovery of novel chemical matter within a predefined potency and property space.

Design of Organic Electronic Materials With a Goal-Directed Generative Model Powered by Deep Neural Networks and High-Throughput Molecular Simulations

“…We built a goal-directed generative model using the REINVENT ( Olivecrona et al, 2017 ) protocol, which has shown success in drug discovery applications ( Ghanakota et al, 2020 ) by generating tens of thousands of unique structures with targeted properties while only requiring a few hours of computing time. The deep reinforcement learning methodology applied in the REINVENT protocol is a robust solution for chemical enumeration while not consuming a prohibitive amount of computation cost ( Olivecrona et al, 2017 ; Ghanakota et al, 2020 ).…”

“…The second stage shifts the distribution based on a utility function, encoding the desired property ranges. In previous studies, we trained the REINVENT algorithm with a group of structures generated by the PathFinder algorithm ( Ghanakota et al, 2020 ). This work aims to use a design space for REINVENT to cover organic electronics, represented by the popular structural motifs shared among successful hole transport materials.…”

Comparison of structure- and ligand-based scoring functions for deep generative models: a GPCR case study

“…To our knowledge, few previous studies exist which have incorporated structural data into deep generative model scoring functions, compared to the ligand-based counterpart. Firstly, Ghanakota et al [ 48 ] combined high throughput free energy perturbation (FEP) with REINVENT to identify potential CDK2 inhibitors. To achieve this, they trained an AutoQSAR model [ 49 ] on a subset of 1,000 enumerated analogues of a potent inhibitor with the corresponding FEP predictions, which was subsequently used as the REINVENT scoring function.…”

“…(3) We directly use a physics-based scoring function (i.e. molecular docking) to obtain scores during the generative model training process, as opposed to predicting the outcome of said function via machine learning as in [ 48 , 52 ]. (4) In our approach, the model actively learns the conditional probability distribution of SMILES symbols that are associated with better docking scores and as such variable size distributions can be sampled (up to billions [ 57 ]) of molecules, as opposed to sampling a finite latent space as in [ 51 ].…”

Actively Searching: Inverse Design of Novel Molecules with Simultaneously Optimized Properties

“…31 This approach was extended in their recent follow-up to train a goal-directed generative model to generate promising ligands for further screening. 32 We propose formulating the entire goal of multi-target chemical optimization as an active learning problem. Rather than attempting to determine optimal regions of chemical space to sample or train on, we query to model to obtain its suggestions for compounds with the desired properties.…”