Computer-aided multi-objective optimization in small molecule discovery

Jenna C. Fromer, Connor W. Coley

TL;DR

This paper reviews computer-aided multi-objective optimization (MMO) techniques for small molecule discovery. It highlights Pareto optimization as a more robust approach than scalarization for exploring trade-offs between competing molecular properties, and discusses various generative and library-based methods for implementing MMO.

What Problem Does This Paper Address?

Discovering new small molecules, whether for drugs, materials, or other applications, is often a complex multi-objective optimization problem. This means that researchers need to find molecules that simultaneously satisfy multiple, often conflicting, properties. For example, a new drug needs to be potent, safe, bioavailable, and synthesizable. Traditional approaches often simplify this by "scalarizing" the objectives – combining them into a single score using weighting factors. However, this approach requires prior knowledge of the relative importance of each property and obscures the inherent trade-offs. This paper addresses the need for methods that can effectively explore these trade-offs and identify a diverse set of optimal solutions without making strong assumptions about property importance.

Methods

The paper categorizes MMO methods into two main approaches:

  • Library-based: This involves searching through a pre-defined library of molecules. The authors highlight Bayesian Optimization (BO) as a key framework here. BO iteratively builds a surrogate model of the objective functions and uses an acquisition function to decide which molecules to evaluate next. For MMO, this involves adapting BO to handle multiple objectives, often using techniques like non-dominated sorting or hypervolume improvement.
  • De novo design: This approach aims to generate entirely new molecules. The paper discusses several generative model categories:
  • Iterative Distribution Learning: Generative models are iteratively retrained on promising subsets of generated molecules to shift their property distributions towards the Pareto front.
  • Genetic Algorithms (GAs): Molecules are evolved using operators like mutation and crossover, with selection criteria guiding the population towards optimal solutions. Non-dominated sorting is a key adaptation for MMO.
  • Reinforcement Learning (RL): Agents learn policies to generate molecules by maximizing a reward function. For MMO, the reward function can be designed to encourage exploration of the Pareto front.
  • Conditional Generation: Generators produce molecules with specific, user-defined property profiles, rather than directly optimizing them. While not strictly iterative optimization, it's discussed due to its prevalence and potential extension to MMO.

The paper also emphasizes the importance of defining the "chemical space" (library-based vs. generative) and how to mathematically represent objectives (continuous, thresholded, or Boolean).

Datasets

Not explicitly discussed in the paper. The authors refer to existing datasets and computational or experimental "oracles" (functions that evaluate molecular properties) used in various case studies.

Key Insights and Ideas

  • Pareto Optimization is Key: The paper strongly advocates for Pareto optimization over scalarization for MMO. Pareto optimization reveals the full spectrum of trade-offs between objectives, allowing for more informed decision-making without requiring a priori weighting.
  • Bayesian Optimization as a Foundation: For library-based approaches, BO provides a robust framework for efficiently exploring large chemical spaces by intelligently selecting candidates for evaluation.
  • Generative Models Offer Flexibility: De novo design methods, leveraging techniques like GANs, VAEs, and RL, can explore vast, potentially infinite, chemical spaces, overcoming the limitations of pre-defined libraries.
  • Diversity is Crucial: Both Pareto diversity (spread of solutions on the Pareto front) and molecular diversity (structural variety of generated molecules) are important for robust exploration and hedging bets against experimental validation failures.
  • Challenges in High Dimensions: As the number of objectives increases, standard Pareto optimization techniques like non-dominated sorting and hypervolume estimation become computationally challenging.

Key Results

The paper is a review and does not present new experimental results. It synthesizes findings from numerous cited studies, demonstrating the effectiveness of various MMO approaches across different applications, including:

  • Discovering transition metal complexes for redox flow batteries with maximized solubility and redox potential.
  • Designing selective inhibitors for biological targets.
  • Optimizing materials for gas separation with improved permeability and selectivity.
  • Generating drug-like molecules with desired property profiles.

The review highlights that MMO approaches generally outperform random search and scalarization methods in terms of efficiency and the quality of discovered solutions.

Limitations and Future Work

  • Scalability to Many Objectives: The paper acknowledges that current Pareto optimization methods struggle with a large number of objectives due to computational complexity. Future work may involve scalarizing certain objectives or converting them to constraints.
  • Incorporating Uncertainty in Generative Models: While BO explicitly uses uncertainty for exploration, many generative model optimization workflows bypass this by using ground-truth objectives for selection. The authors suggest an opportunity to integrate Bayesian principles and surrogate models into de novo design to better guide selection.
  • Benchmarking and Validation: The need for realistic oracle functions and challenging benchmarks that mimic practical applications is emphasized. Experimental validation remains crucial to confirm computational predictions.
  • Diversity-Aware Batching: Best practices for diversity-aware batching in MMO are not yet well-established, presenting an area for future research.
  • Defining "Good" MMO: Evaluating MMO performance solely on individual property values or constraints can be misleading. Metrics that capture the combination of properties and the overall quality of the Pareto front (e.g., density, hypervolume) are more informative.

Summary updated

Abstract

Molecular discovery is a multi-objective optimization problem that requires identifying a molecule or set of molecules that balance multiple, often competing, properties. Multi-objective molecular design is commonly addressed by combining properties of interest into a single objective function using scalarization, which imposes assumptions about relative importance and uncovers little about the trade-offs between objectives. In contrast to scalarization, Pareto optimization does not require knowledge of relative importance and reveals the trade-offs between objectives. However, it introduces additional considerations in algorithm design. In this review, we describe pool-based and de novo generative approaches to multi-objective molecular discovery with a focus on Pareto optimization algorithms. We show how pool-based molecular discovery is a relatively direct extension of multi-objective Bayesian optimization and how the plethora of different generative models extend from single-objective to multi-objective optimization in similar ways using non-dominated sorting in the reward function (reinforcement learning) or to select molecules for retraining (distribution learning) or propagation (genetic algorithms). Finally, we discuss some remaining challenges and opportunities in the field, emphasizing the opportunity to adopt Bayesian optimization techniques into multi-objective de novo design.
BibTeX
@inproceedings{fromer2023computer,
  author = {Jenna C. Fromer and Connor W. Coley},
  title = {{Computer-aided multi-objective optimization in small molecule discovery}},
  booktitle = {Patterns},
  year = {2023},
  url = {https://openreview.net/forum?id=rz2clo0sTV},
}
Figure