[Paper Review] ChemBO: Bayesian Optimization of Small Organic Molecules with Synthesizable Recommendations

Ksenia Korovina, Sailun Xu+5 more|arXiv (Cornell University)|Aug 5, 2019

Computational Drug Discovery MethodsComputer Science46 references39 citations

TL;DR

ChemBO introduces a Gaussian process-based Bayesian optimization framework that searches synthesizable small organic molecules via an optimal-transport inspired kernel and a synthesis-graph guided acquisition, achieving competitive QED and Pen-logP scores with fewer evaluations.

ABSTRACT

In applications such as molecule design or drug discovery, it is desirable to have an algorithm which recommends new candidate molecules based on the results of past tests. These molecules first need to be synthesized and then tested for objective properties. We describe ChemBO, a Bayesian optimization framework for generating and optimizing organic molecules for desired molecular properties. While most existing data-driven methods for this problem do not account for sample efficiency or fail to enforce realistic constraints on synthesizability, our approach explores the synthesis graph in a sample-efficient way and produces synthesizable candidates. We implement ChemBO as a Gaussian process model and explore existing molecular kernels for it. Moreover, we propose a novel optimal-transport based distance and kernel that accounts for graphical information explicitly. In our experiments, we demonstrate the efficacy of the proposed approach on several molecular optimization problems.

Motivation & Objective

• Motivation: to design molecules with desirable properties while ensuring synthesizability.
• Goal: to maximize objective properties (QED, Pen-logP) using sample-efficient BO that respects synthesis constraints.
• Contribution: develop a GP model with novel kernels and a synthesis-guided acquisition process.
• Goal alignment: provide actionable synthesis recipes along with candidate molecules to chemists.

Proposed method

• Model: Gaussian process for structure-property relations in molecules.
• Kernel: fingerprint-based kernel plus a novel optimal-transport based dissimilarity between graphs, combined as an ensemble kernel.
• Acquisition: acquisition optimization performed on a synthesis graph to ensure synthesizability and provide synthesis recipes.
• Synthesis graph: nodes are molecules; edges/parents defined by reagents and reaction conditions; random-walk based exploration to maximize acquisition.
• Optimization setup: compare fingerprint kernel, ot-dist kernel, and their sum (sum-kernel) against a random-walk baseline.
• Evaluation: two objectives (QED and Pen-logP) with 80 BO iterations (100 total evaluations) and initial 20 molecules from ChEMBL.

Experimental results

Research questions

• RQ1 Can GP-based Bayesian optimization with synthesis-aware constraints efficiently optimize molecular properties?
• RQ2 Does incorporating an optimal-transport based molecular dissimilarity improve performance over standard fingerprint kernels in ChemBO?
• RQ3 How does a synthesis-graph guided acquisition compare to naive or random synthesis exploration in identifying high-quality molecules?
• RQ4 What are the final achievable QED and Pen-logP scores under synthesis constraints with limited evaluations?

Key findings

• ChemBO variants outperformed the random-walk baseline on both QED and Pen-logP benchmarks.
• ot-dist yields higher QED; fingerprint yields higher Pen-logP; sum-kernel provides a robust trade-off with lower variance.
• Final results (80 evaluations): QED: rand 0.90±0.01, fingerprint 0.91±0.01, ot-dist 0.93±0.01, sum-kernel 0.94±0.01; Pen-logP: rand 6.81±0.34, fingerprint 9.79±2.26, ot-dist 8.10±1.01, sum-kernel 8.65±0.43.
• ChemBO discovers novel molecules (high novelty) with 78–96% novelty depending on objective and kernel.
• Synthesis-path sanity checks show end-molecule SA scores within reasonable range compared to ChEMBL/ZINC datasets.
• ChemBO achieves competitive or better scores than several state-of-the-art methods with significantly fewer evaluations (100 BO iterations, 5 trials).