MEMES: Machine learning framework for Enhanced MolEcular Screening

Abstract

In drug discovery applications, high throughput virtual screening exercises are routinely performed to determine an initial set of candidate molecules referred to as "hits". In such an experiment, each molecule...

MoleGuLAR: Molecule Generation Using Reinforcement Learning with Alternating Rewards

“…We further tested various machine-learning models for this regression task. Jaeger et al proposed the Mol2vec model for learning vector representations of SMILES strings that can then be used as input for further downstream tasks like binding affinity prediction as done by Mehta et al along with predicting other properties. Using these embeddings, a random forest model with 250 decision trees was trained for predicting the BA.…”

“…The identified hits are then optimized to get higher binding affinity, reduce toxicity, and improve oral bioavailibity. , The time and expense involved in this process give rise to alternate in silico approaches like virtual screening wherein small molecules from existing drug libraries are computationally evaluated by generating protein ligand complexes using docking calculations and ranking them using a scoring function. , However, these also come with the caveat that finding the most stable conformation of the complex is a nonconvex optimization problem, and it can take a very large amount of time (≈10 min per molecule) to find the most optimal conformation. These can be made faster using machine-learning-based approaches like the works of Aggarwal et al for detecting the ligand-binding site, Chelur and Priyakumar for binding residue detection, and Mehta et al for enhanced molecular sampling. − However, even the most exhaustive studies have been able to find binding affinities of ≈10 8 molecules on a single target which is minuscule in comparison to the vast magnitude of the chemical space with about 10 60 synthesizable molecules . This posits the argument for the de novo generation of molecules with high binding affinities to the required target instead of searching in existing libraries.…”

“…12,13 However, these also come with the caveat that finding the most stable conformation of the complex is a nonconvex optimization problem, and it can take a very large amount of time (≈10 min per molecule) to find the most optimal conformation. These can be made faster using machine-learning-based approaches like the works of Aggarwal et al 14 for detecting the ligand-binding site, Chelur and Priyakumar 15 for binding residue detection, and Mehta et al 16 for enhanced molecular sampling. 14−16 However, even the most exhaustive studies 17 have been able to find binding affinities of ≈10 8 molecules on a single target which is minuscule in comparison to the vast magnitude of the chemical space with about 10 60 synthesizable molecules.…”

Modern machine learning for tackling inverse problems in chemistry: molecular design to realization

“…This has led to the development of high throughput virtual screening (VS) methods that make it possible for us to narrow down possible candidates from a large library of molecules at a much faster rate than that possible with traditional methods. 57,58 In spite of these advancements, a significant computational effort is required to screen these huge libraries of molecules which may reach sizes beyond billions in number. 3,4 This calls for methods that generate molecules in a more targeted way and explore the chemical space more efficiently.…”

A transfer learning approach for reaction discovery in small data situations using generative model

“…A large body of work on the applications of deep generative models for molecular design has been focused on drug discovery ( Elton et al., 2019 ; Mendez-Lucio et al., 2020 ; Lim et al., 2020 ; Mehta et al., 2021 ). We wondered if the capabilities of these methods could suitably be tailored to fit the reaction discovery workflow.…”