Optimal Compound Downselection To Promote Diversity and Parallel Chemistry

Jenna C. Fromer, Alexandra D. Volkova, Connor W. ColeyJournal of Chemical Information and Modeling 2025 · OpenReview

TL;DR

This paper introduces SPARROW, an enhanced computational framework for selecting promising drug candidates. It goes beyond traditional methods by simultaneously optimizing for compound utility, synthetic feasibility, molecular diversity, and amenability to parallel synthesis, making it a more holistic tool for accelerating drug discovery.

What Problem Does This Paper Address?

In early-stage drug discovery, researchers face a critical challenge: choosing which of the vast number of potential molecules to synthesize and test. This decision is complex, involving trade-offs between a molecule's potential therapeutic value (utility), how easy it is to make (synthetic feasibility and cost), how different it is from other candidates (diversity), and how efficiently it can be produced in batches (parallel chemistry). Existing computational tools often focus on only a subset of these factors, failing to capture the full complexity of a chemist's decision-making process. This can lead to suboptimal selections, potentially missing promising drug candidates or incurring unnecessary synthesis costs and delays.

Methods

The core of this work is the enhancement of the SPARROW (Synthesis Planning and Rewards-based Route Optimization Workflow) framework. The authors introduce several key improvements:

  • Expected Cumulative Reward: Instead of a simple weighted sum, SPARROW now optimizes for an "expected cumulative reward." This metric intelligently combines a compound's predicted utility with the probability of successfully synthesizing it, accounting for the risk of failure at each step. This is a more principled way to balance reward and risk.
  • Cluster-Based Diversity Formulation: To promote molecular diversity, SPARROW now uses a cluster-based approach. Candidates are grouped into clusters (initially based on structural similarity using Morgan fingerprints and the Butina clustering algorithm), and the optimization objective is modified to encourage the selection of compounds from a wider range of clusters. This allows for more flexible definitions of diversity beyond just structural similarity.
  • Parallel Chemistry Amenability: The framework is extended to favor reactions that are amenable to parallel synthesis. This is achieved by constraining the number of distinct reaction classes used in the synthesis routes, encouraging the selection of reactions that can be run simultaneously in batches.

These new formulations are integrated into SPARROW, which then uses optimization algorithms to select a batch of compounds and their corresponding synthetic routes.

Datasets

The primary dataset used for evaluation consists of 9,621 candidate antibiotics for Staphylococcus aureus, generated by the AutoMolDesigner tool. This dataset is publicly available. The authors also leverage reaction data from Reaxys and building block information from Enamine's catalog.

Key Insights and Ideas

  • Holistic Optimization is Key: The most significant insight is that optimizing for multiple, often competing, criteria (utility, risk, diversity, parallelizability) simultaneously leads to more effective compound selection than optimizing for them individually.
  • Risk-Aware Synthesis Planning: Incorporating the likelihood of synthetic success directly into the reward calculation is crucial. This moves beyond simply finding a route to finding a reliable route, which is critical for experimental success.
  • Diversity as a Strategic Tool: Promoting diversity isn't just about exploring chemical space; it's a strategy to mitigate the risk of an entire chemical series failing experimental validation or revealing similar liabilities.
  • Parallel Chemistry as a Bottleneck Solver: Enabling parallel synthesis is not just about speed; it's about efficiently exploring larger chemical spaces and accelerating the discovery pipeline. SPARROW's ability to incorporate this constraint directly addresses a practical bottleneck in chemical synthesis.

Key Results

  • The enhanced SPARROW framework successfully identified diverse batches of compounds that were also amenable to parallel synthesis.
  • By incorporating expected cumulative reward, the framework more accurately captures the risk associated with synthetic routes.
  • The cluster-based diversity formulation allowed for the selection of compounds spanning a greater number of distinct clusters, demonstrating improved diversity.
  • Constraining reactions to specific classes significantly improved amenability to parallel chemistry, although this sometimes came with a slight decrease in overall expected reward.
  • The authors demonstrated that it's possible to balance the trade-off between diversity and parallel chemistry, achieving both to a significant degree. For instance, in one example, combining diversity and parallel chemistry considerations led to selections spanning eight distinct clusters, all using the same reaction class.

Limitations and Future Work

  • Computational Cost: While improvements have been made, optimizing for these multiple criteria can still be computationally intensive, especially for very large retrosynthetic networks and candidate sets. The paper mentions memory issues encountered with certain optimization formulations.
  • Heuristic for Parallel Chemistry: The current approach to parallel chemistry amenability is heuristic, relying on classifying reactions into classes. Future work could involve more direct modeling of parallel synthesis capabilities.
  • Definition of Diversity: While the cluster-based approach offers flexibility, the definition of "diversity" can still be application-dependent. Future work could explore more tailored diversity metrics.
  • Generalizability: While demonstrated on a specific antibiotic discovery case study, further validation on diverse chemical spaces and drug discovery projects would be beneficial.

The authors suggest that SPARROW's flexibility in defining clusters and reaction classes allows it to be applied to a wide range of molecular design campaigns.

Summary updated

Abstract

Click to copy section linkSection link copied!
BibTeX
@article{fromer2025optimal,
  author = {Jenna C. Fromer and Alexandra D. Volkova and Connor W. Coley},
  title = {{Optimal Compound Downselection To Promote Diversity and Parallel Chemistry}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2025},
  url = {https://openreview.net/forum?id=AX0NXrkItY},
}
Figure