Source-linked AI summary
ChemBO: Bayesian Optimization of Small Organic Molecules with Synthesizable Recommendations
Ksenia Korovina, Sailun Xu, Kirthevasan Kandasamy, Willie Neiswanger, Barnabas Poczos, Jeff Schneider, Eric P. Xing
TL;DR
Molecular design requires finding desirable compounds from an enormous chemical space using costly experiments while respecting synthesis constraints. ChemBO addresses this with Gaussian-process Bayesian optimization over a synthesis graph and graph-aware molecular kernels, achieving competitive benchmark performance and strong novelty with fewer evaluations. Its practical reliability remains bounded by imperfect reaction-outcome prediction and potentially inefficient synthesis paths.
Problem
Molecular optimization must identify desirable molecules from a vast chemical space using few costly evaluations while producing synthesizable recommendations.
Method
ChemBO uses Gaussian-process Bayesian optimization, synthesis-graph exploration, and an optimal-transport molecular kernel that explicitly incorporates graph structure.
Results
ChemBO methods outperform random walk on QED and penalized logP, while the optimal-transport and fingerprint variants trade off across tasks and their sum-kernel combination performs well on both benchmarks.
Takeaways & Limitations
ChemBO achieves competitive molecular-optimization results with substantially fewer evaluations and explores novel molecules despite synthesizability constraints.
Takeaways & Limitations
Because reaction predictors are imperfect, ChemBO can recommend unsynthesizable molecules or incorrect, inefficient synthesis recipes.
Abstract
from arXiv · showhide
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.
1 Introduction
ChemBO frames molecular design as sample-efficient Bayesian optimization while requiring recommendations to be chemically valid and synthesizable. It combines Gaussian-process modeling, molecular kernels, and synthesis-graph exploration, with experiments showing competitive benchmark performance using fewer evaluations.
- ChemBO treats molecule design as black-box optimization, seeking high-value molecules with as few costly laboratory evaluations as possible.
- ChemBO combines Bayesian optimization with synthesis-graph exploration so recommendations are chemically valid and accompanied by synthesis recipes.
- The method uses a Gaussian-process structure–property model with existing fingerprint kernels and a new optimal-transport similarity measure that treats molecules as graphs.
- ChemBO outperforms simpler nons probabilistic synthesizeable-optimization alternatives, while its QED and penalized partition-coefficient results are competitive with state-of-the-art methods using less data and fewer evaluations.
2 Related Work
Prior molecular-optimization methods use string or graph representations and often require many samples or rely on imperfect synthesizability proxies. ChemBO instead jointly explores synthesis paths and optimizes molecular properties with sample efficiency as a central goal.
- Optimization: Molecular optimization methods commonly use SMILES strings or molecular graphs, with graph-based approaches often relying on generative modeling or reinforcement learning.
- Optimization: Because laboratory evaluations can be expensive, conventional generative and reinforcement-learning approaches may require too many samples for stateless molecular optimization.
- Synthesizable recommendations: Synthesizability is often enforced through proxy scores or objective constraints, but such scores can produce unrealistic structures such as excessively large carbon rings.
- Joint optimization and synthesis: ChemBO jointly performs property optimization and synthesis-path exploration, distinguishing it from approaches that learn generative representations without designing for sample-efficient goal-directed optimization.
- Kernels on molecules: Molecular graph kernels can miss chemical structure or be computationally expensive, motivating ChemBO’s explicit graph-aware optimal-transport kernel alongside fingerprint kernels.
3 Method
ChemBO models molecular properties with Gaussian processes and optimizes acquisition values over a synthesis graph. Its kernel suite includes fingerprints and an optimal-transport graph distance, while reaction prediction enables synthesizable exploration but introduces reliability and path-efficiency limitations.
- ChemBO iteratively fits a posterior Gaussian process, selects the molecule with the highest acquisition value, evaluates it, and updates the dataset.
- 3.2 Kernel: ChemBO combines fingerprint kernels with an optimal-transport dissimilarity that matches weighted atoms while penalizing atom-type, bond-type, and unmatched structure differences.
- 3.2 Kernel: The optimal-transport distance can overstate dissimilarity for larger molecules, whereas its unnormalized form better captures size-related drug-property differences than the normalized counterpart.
- 3.2 Kernel: An exponential-sum kernel combines four weighting and normalization variants, with component weights learned as kernel hyperparameters.
- The framework optimizes acquisition by randomly walking through reagent and process-condition choices, using reaction outcomes to expand the pool of synthesizable candidates.
- 3.3 Exploring the Space of Synthesizable Molecules and Optimizing the Acquisition: ChemBO relies on imperfect reaction predictors, so it may recommend unsynthesizable molecules or inefficient synthesis recipes reached through circuitous random walks.
4 Experiments
ChemBO is evaluated on QED and Pen-logP optimization using multiple kernels, synthesis-constrained search, and comparisons with random exploration, prior methods, and virtual screening. The experiments show strong optimization performance, plausible synthesis paths, and broad exploration under limited evaluations.
- Experimental setup: The evaluation uses QED and Pen-logP, with 20 initial ChEMBL molecules followed by 80 optimization iterations and 100 total evaluations per run.QED ranges from 0 to 1, while Pen-logP combines logP, synthetic accessibility, and a ring penalty.
- Main Results: ChemBO's fingerprint, ot-dist, and sum-kernel variants all outperform random walk on QED and Pen-logP, with sum-kernel offering a strong trade-off and lower variance.ot-dist performs better on QED, fingerprint on Pen-logP, while the sum-kernel works well on both benchmarks.
- Reliability of synthesis paths: ChemBO produces synthesis paths whose endpoint synthetic-accessibility scores are within reasonable ranges of curated ChEMBL and ZINC averages, while path minima are substantially lower.Actual laboratory synthesis would be required for thorough validation of the proposed paths.
- Novel Molecules: Novel molecules comprise 95.64% of ot-dist QED outputs, 96.84% of fingerprint QED outputs, 78% of ot-dist Pen-logP outputs, and 87.67% of fingerprint Pen-logP outputs.These results indicate exploration beyond molecules appearing in the full ChEMBL dataset despite synthesizability constraints.
- Comparison with existing work: ChemBO reaches better or comparable QED and Pen-LogP values than prior methods using 1,000 total function evaluations, versus at least 5,000 or roughly 250,000 for several alternatives.The comparison uses best reported values across prior work, although those methods use different settings and do not impose the same synthesizability conditions.
- Virtual screening baseline: Virtual screening performs worse than ChemBO in the reported comparison even when allowed more samples and access to compounds outside the starting pool.This baseline can access a larger dataset search space by purchasing randomly sampled compounds in addition to the initial pool.
5 Conclusion
ChemBO targets automated molecular optimization by combining Bayesian recommendations with synthesizability and synthesis recipes. The authors report competitive or better benchmark scores with fewer objective evaluations, while identifying synthesis-predictor reliability and broader molecular scope as remaining challenges.
- ChemBO combines Bayesian optimization with synthesizable recommendations and synthesis recipes for iterative molecular design.
- ChemBO can outperform naive alternatives, while its optimal-transport kernel sometimes outperforms standard kernels and a combined kernel offers lower variance across benchmarks.
- On two benchmark objectives, ChemBO achieves competitive or better scores than existing work using significantly fewer objective evaluations.
- The framework remains constrained by current synthesis predictors, motivating improvements in predictor reliability and chemical-space exploration.
- Future directions include counting synthesis steps in the optimization budget and extending the framework to biologics and other molecular optimization problems.
A Some Additional Details on the Dissimilarity Measure
The molecular dissimilarity measure is reformulated as an optimal-transport program over atom-match weights and penalties. An augmented cost matrix and marginal vectors encode atom and bond mismatches alongside non-matching penalties.
- The matching matrix U assigns weights between atoms of two molecules, while molecule weights aggregate atom weights for the transport marginals.
- The dissimilarity computation is reformulated as an optimal-transport program equivalent to the original linear program.
- The augmented cost matrix combines atom-type and bond-type penalties in its main block with unit non-matching penalties in its final row and column.
- The transport variable U′ has the original atom-matching matrix as its first block and satisfies row and column marginal constraints.
- The formulation follows a formally established optimal-transport equivalence in related work.
A.2 T-SNE visualizations for the OT distance
The authors evaluate the proposed optimal-transport dissimilarity by embedding distances among 200 molecules in two dimensions and coloring points by molecular properties. The embeddings show relatively continuous property variation and clusters, with OT distance appearing slightly better than comparison distances.
- The experiment compares four base combinations of settings for the proposed OT distance.
- Distances among 200 randomly sampled molecules are embedded with two-dimensional t-SNE and colored by QED and synthetic accessibility scores.
- The induced embedding space shows relatively continuous dependencies between molecular structure and properties, including clusters with similar values.
- Compared with fingerprint-based and inverted-Tanimoto alternatives, OT distance appears slightly better at separating low and high property values.
A.3 Some Known Limitations
ChemBO’s graph-based dissimilarity cannot distinguish stereoisomers because molecules with identical connectivity receive zero distance. Combining the kernel with features encoding three-dimensional structure is proposed as a possible remedy.
- The graph-based dissimilarity assigns zero distance to stereoisomers with identical formulas and bonded atoms, despite their different three-dimensional structures and potentially different physical properties.The paper illustrates this with D-Glucose and L-Glucose, which differ in digestibility.
- The kernel could be combined with other three-dimensional features using a sum or product kernel to address stereoisomer indistinguishability.
B Some Implementation Details
ChemBO fits Gaussian-process hyperparameters by marginal-likelihood maximization and uses an ensemble of acquisition functions. Acquisition optimization repeatedly expands the searchable molecule pool across Bayesian-optimization iterations.
- Gaussian-process hyperparameters are fitted by maximizing the marginal likelihood, while acquisition uses an ensemble of EI, UCB, and TTEI.
- The explorer runs for 20 iterations per Bayesian-optimization iteration and adds newly found molecules to the initial pool for subsequent searches.This maintains a large pool throughout the optimization routine.
Experiments with low starting value
ChemBO is evaluated from low-quality starting pools to test whether optimization remains effective when initial molecules have limited objective values. It performs well and retains its earlier relative behavior against the baseline and between kernels.
- ChemBO performs well from randomly selected 20-molecule pools capped at QED 0.7 and penalized LogP 3, outperforming the baseline with the same pattern as before.The caps correspond approximately to the 60th percentiles in ChEMBL.
- The fingerprint kernel performs worse than the OT-distance kernel on QED but better on the penalized LogP task.
Synthesis Paths
The paper visualizes synthesis paths for optimized molecules and explains how starting reagents and reaction parents are represented. It also reports artifacts and caveats involving initial-pool composition, large rings, and the interpretation of supporting figures.
- Synthesis Paths: Synthesis-path visualizations show optimized molecules, with boxed molecules representing members of the initial pool and multiple parent arrows indicating reactions that combine those parents.
- Synthesis Paths: Some paths combine complex molecules into simpler products such as methane, an artifact attributed to the initial pool that may be reduced through careful pool selection.
- Synthesis Paths: Large rings appear in all shown paths and may be difficult to synthesize because they are not necessarily stable; the paper attributes this possibility to Rexgen.
- Synthesis Paths: Red boxes in molecular drawings arise from RDKit’s layout algorithm overlaying two atoms, which is likely for large molecules.
- Synthesis Paths: The ChEMBL histograms contextualize optimization by showing QED and penalized LogP distributions, with optimized values lying in the dataset’s highest percentiles.
- Synthesis Paths: Figure 5 compares three methods across iterations using highest-found QED and penalized LogP, averaging five runs with one-standard-error shading.
- Synthesis Paths: The displayed paths include examples targeting penalized LogP values of 11.988 and 8.306 and QED values of 0.92 and 0.93.
- Synthesis Paths: The t-SNE figures compare OT distance configurations colored by QED or SA score and contrast OT-distance embeddings with fingerprint-based embeddings.