DrugSynthMC: An Atom-Based Generation of Drug-like Molecules with Monte Carlo Search

Milo Roucairol, Alexios Georgiou, Tristan Cazenave, Filippo Prischi, Olivier E. Pardo

TL;DR

This paper introduces DrugSynthMC, a novel algorithm that uses Monte Carlo Search (MCS) combined with statistical priors to generate thousands of drug-like molecules per second. Unlike many deep learning methods, DrugSynthMC doesn't require training data and can generate interpretable molecules that are predicted to be synthesizable, significantly expanding the chemical space available for drug discovery.

What Problem Does This Paper Address?

The drug discovery process heavily relies on identifying novel molecules with desired properties. While in silico methods have become indispensable, a major bottleneck is the limited chemical diversity of existing compound libraries. Many current generative models, often based on deep learning, focus on designing molecules to bind specific targets. However, a significant challenge remains in generating de novo virtual libraries that are not only diverse and drug-like but also synthesizable. Existing approaches often struggle with synthesizability, leading to reduced chemical diversity or requiring complex, integrated synthesis planning tools that can inherit their own limitations. This paper aims to address this gap by developing a method that efficiently generates large, diverse, drug-like, and synthesizable molecule libraries without relying on training data or explicit synthesizability metrics during generation.

Methods

The core of DrugSynthMC is a Monte Carlo Search (MCS) algorithm. The search model operates by building molecules character by character, atom by atom, using a set of rules that ensure structural validity and respect chemical valency. This process is guided by statistical priors derived from existing drug databases (specifically, the FDA subset of the ZINC20 database) to inform the probability of using certain atoms and bond types, effectively mimicking common chemical patterns.

Key methodological aspects include:

  • Atom-Based Search Model: Molecules are constructed incrementally, with rules governing atom valency, subtree nesting, and cycle lengths (favoring 5-7 membered rings).
  • Statistical Priors: Frequency matrices derived from drug databases guide the selection of atoms and bonds, prioritizing common elements (C, O, N, F, S, Cl) and bond types. This helps ensure generated molecules resemble known drug-like structures.
  • Monte Carlo Search (MCS): The algorithm explores the vast chemical space by making sequential decisions (adding characters/atoms). Different "playout" strategies (how the search is explored) are investigated, including random, enforced, n-gram-based, and neural network-based.
  • Nested Monte Carlo Search (NMCS): This specific MCS variant is found to be particularly effective, outperforming other MCS algorithms like UCT (Upper Confidence Bounds applied to Trees) in certain scenarios, especially when more complex scoring functions are involved. NMCS is noted for its ability to explore the search tree more precisely and avoid local optima.
  • No Training Data Requirement: A significant departure from many deep learning methods, DrugSynthMC does not require a training dataset. Its "learning" comes from statistical priors derived from existing chemical databases.

Datasets

  • ZINC20 Database (FDA Subset): This dataset was used to derive statistical priors for atom and bond frequencies, as well as information on ring structures. This subset is representative of known drug-like molecules.
  • No new benchmark datasets were introduced.

Key Insights and Ideas

  • Synthesizability without Explicit Metrics: DrugSynthMC demonstrates that it's possible to generate molecules with a high predicted synthesizability without explicitly enforcing or rewarding it during the generation process. This is achieved through the use of statistical priors that reflect common chemical patterns found in synthesizable drugs.
  • N-gram Priors Outperform Neural Networks: Surprisingly, the simpler n-gram-based playout strategy, which relies on short character sequences, yielded better results in terms of synthesizability and drug-likeness compared to more complex neural network-based playouts. This suggests that for this task, mimicking established chemical patterns is more effective than learning complex, potentially overfitted, relationships.
  • NMCS Superiority: The Nested Monte Carlo Search (NMCS) algorithm consistently outperformed UCT, especially when combined with more sophisticated scoring functions, highlighting its robustness in exploring complex chemical spaces.
  • Explainability and Versatility: The algorithm's reliance on explicit rules and statistical priors makes it more explainable than many black-box deep learning models. Its modular design also allows for easy integration of different scoring functions for various drug design goals.
  • Efficient Expansion of Chemical Space: DrugSynthMC can generate thousands of molecules per second, enabling rapid exploration of a vast and novel chemical space that is currently underexplored.

Key Results

  • High Validity: Generated SMILES consistently achieved 100% validity when parsed by RDKit, indicating no syntax errors. This is superior to many other generative models.
  • High Novelty: Generated molecules showed a high degree of novelty, with over 99% not being present in the ZINC250 KS database.
  • Good Drug-Likeness: Molecules generated adhered to Lipinski's "Rule of 5" and exhibited favorable physicochemical properties (e.g., molecular weight, hydrogen bond donors/acceptors, clogP) comparable to FDA-approved drugs.
  • Promising Synthesizability: Using AiZynthFinder, a significant proportion of generated molecules were predicted to be synthesizable, with the n-gram playout yielding the highest percentage (32.2%).
  • Speed: The algorithm can generate thousands of molecules per second, making it highly efficient for virtual library generation. For instance, 1000 molecules were generated in under a second using the n-gram playout.

Limitations and Future Work

  • Synthesizability Underestimation: The authors note that AiZynthFinder might underestimate the synthesizability of molecules, suggesting that the actual synthesizability of DrugSynthMC-generated compounds could be even higher.
  • Larger Molecules and Synthesizability: Generating larger molecules (e.g., >30 characters in SMILES) made it more challenging for AiZynthFinder to predict synthesizability.
  • Neural Playout Speed: While promising, neural network playouts were significantly slower than n-gram playouts.
  • Future Work: The authors envision integrating more complex, target-specific scoring functions to tailor generated libraries for particular binding pockets or drug discovery objectives. They also suggest using pseudofree energy of binding as a reward in a reinforcement learning framework within DrugSynthMC to converge on compounds with desired pharmacophores and binding affinities.

Summary updated

Abstract

A growing number of Deep Learning (DL) methodologies have recently been developed to design novel compounds and expand the chemical space within virtual libraries. Most of these Neural Network approaches design molecules to specifically bind a target, based on its structural information and/or knowledge of previously identified binders. Fewer attempts have been made to develop approaches for de novo design of virtual libraries, as synthesizability of generated molecules remains a challenge. In this work, we developed a new Monte Carlo Search (MCS) algorithm, DrugSynthMC (Drug Synthetise using Monte Carlo), in conjunction with DL and statistical-based priors to generate thousands of interpretable chemical structures and novel drug-like molecules per second. DrugSynthMC produces drug-like compounds using an atom-based search model that builds molecules as SMILES, character by character. Designed molecules follow Lipinski’s “rule of 5”, show a high proportion of predicted-to-be synthesisable compounds and efficiently expand the chemical space within the libraries, without reliance on training datasets, synthesizability metrics or enforcing during SMILES generation. Our approach can function with or without an underlying Neural Network and is thus easily explainable and versatile. This ease in drug-like molecule generation allows for future integration of score functions aimed at different target- or job -oriented goals. Thus, DrugSynthMC is expected to enable the functional assessment of large compound libraries covering an extensive novel chemical space, overcoming the limitations of existing drug collections. The software is available at https://github.com/RoucairolMilo/DrugSynthMC
BibTeX
@article{roucairol2024drugsynthmc,
  author = {Milo Roucairol and Alexios Georgiou and Tristan Cazenave and Filippo Prischi and Olivier E. Pardo},
  title = {{DrugSynthMC: An Atom-Based Generation of Drug-like Molecules with Monte Carlo Search}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2024},
  doi = {10.1021/acs.jcim.4c01451},
}
Figure