[Paper Review] Generative Active Learning for the Search of Small-molecule Protein Binders

Maksym Korablyov, Chenghao Liu+32 more|arXiv (Cornell University)|May 2, 2024

Monoclonal and Polyclonal Antibodies Research4 citations

TL;DR

This paper introduces LambdaZero, a generative active learning framework that leverages deep reinforcement learning to efficiently search the vast chemical space for novel, synthesizable small-molecule inhibitors. By combining a learned generative policy with a surrogate model for molecular docking, LambdaZero achieves an exponential speedup—identifying high-affinity sEH inhibitors equivalent to screening 100 billion molecules with only ~10,000 docking evaluations, and experimentally validating a lead compound (UM0152893) with sub-micromolar inhibition.

ABSTRACT

Despite substantial progress in machine learning for scientific discovery in recent years, truly de novo design of small molecules which exhibit a property of interest remains a significant challenge. We introduce LambdaZero, a generative active learning approach to search for synthesizable molecules. Powered by deep reinforcement learning, LambdaZero learns to search over the vast space of molecules to discover candidates with a desired property. We apply LambdaZero with molecular docking to design novel small molecules that inhibit the enzyme soluble Epoxide Hydrolase 2 (sEH), while enforcing constraints on synthesizability and drug-likeliness. LambdaZero provides an exponential speedup in terms of the number of calls to the expensive molecular docking oracle, and LambdaZero de novo designed molecules reach docking scores that would otherwise require the virtual screening of a hundred billion molecules. Importantly, LambdaZero discovers novel scaffolds of synthesizable, drug-like inhibitors for sEH. In in vitro experimental validation, a series of ligands from a generated quinazoline-based scaffold were synthesized, and the lead inhibitor N-(4,6-di(pyrrolidin-1-yl)quinazolin-2-yl)-N-methylbenzamide (UM0152893) displayed sub-micromolar enzyme inhibition of sEH.

Motivation & Objective

• Address the 'needle-in-a-haystack' problem in drug discovery, where the chemical space exceeds 10^60 molecules, making exhaustive screening infeasible.
• Overcome limitations of traditional virtual screening and active learning, which are constrained by biased chemical libraries and limited search space.
• Develop a de novo generative method that ensures synthesizability and drug-likeness while efficiently exploring combinatorially large molecular spaces.
• Achieve high sample efficiency in identifying potent small-molecule inhibitors by integrating reinforcement learning with surrogate modeling of expensive docking simulations.
• Discover novel molecular scaffolds for protein inhibition, specifically targeting soluble epoxide hydrolase 2 (sEH), with experimental validation of lead compounds.

Proposed method

• LambdaZero employs a deep reinforcement learning-based generative policy to explore the combinatorial space of small molecules, guided by a surrogate model of molecular docking scores.
• The surrogate model approximates the computationally expensive molecular docking oracle, enabling fast evaluation of candidate molecules during the search process.
• A synthesizability constraint is enforced through a differentiable molecular generation policy that ensures generated molecules are synthetically accessible.
• The method uses active learning: batches of generated molecules are evaluated via the docking oracle, and the policy is updated based on feedback to focus on high-scoring regions.
• The framework iteratively enriches a candidate library through multiple rounds of generation, evaluation, and policy refinement.
• The approach is applied to design inhibitors for soluble epoxide hydrolase 2 (sEH), with a focus on achieving high binding affinity and drug-like properties.

Experimental results

Research questions

• RQ1 Can a reinforcement learning-based generative policy efficiently explore the vast chemical space of small molecules to discover novel, high-affinity protein binders?
• RQ2 To what extent can a surrogate model of molecular docking reduce the number of expensive docking evaluations required to identify potent inhibitors?
• RQ3 Can the method generate molecules that are not only potent but also synthesizable and drug-like, avoiding overfitting to known chemical scaffolds?
• RQ4 How does the performance of this generative active learning approach compare to traditional virtual screening in terms of sample efficiency and hit rate?
• RQ5 Can de novo-designed molecules from novel scaffolds achieve experimental validation with sub-micromolar inhibition constants?

Key findings

• LambdaZero discovered a novel quinazoline-based scaffold for sEH inhibition, with the lead compound UM0152893 showing sub-micromolar enzyme inhibition (IC50 = 0.67 μM).
• The method achieved docking scores equivalent to screening 100 billion molecules using only ~10,000 docking evaluations, demonstrating an exponential speedup.
• The generative policy successfully explored novel chemical space, producing molecules with distinct scaffolds not present in existing libraries.
• In vitro validation confirmed the experimental feasibility and potency of the de novo-designed inhibitors, with multiple compounds showing IC50 values below 10 μM.
• The surrogate model enabled accurate and efficient guidance of the search, maintaining high sample efficiency while ensuring synthesizability and drug-likeness.
• The approach outperformed traditional virtual screening and active learning methods by combining generative search with iterative, oracle-guided refinement.