CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation

Technical Summary: CombiMOTS (Combinatorial Multi-Objective Tree Search)

CombiMOTS is a novel framework designed for dual-target molecule generation, a critical task in drug discovery aimed at finding single agents capable of interacting with two distinct protein targets. This approach is intended to improve therapeutic efficacy, mitigate drug resistance, and reduce side effects compared to traditional single-target drugs or combination therapies.

1. The Problem: Challenges in Dual-Target Design

Existing generative models for dual-target drugs face two primary bottlenecks:

• Scalarization Loss: Most current methods simplify multi-objective optimization (e.g., balancing target affinity, drug-likeness, and synthetic accessibility) into a single scalar reward. This "averaging" fails to capture the complex trade-offs between competing objectives, often resulting in molecules that excel in one area but are biologically or physically impractical.
• Synthesizability Gap: Many deep generative models propose "ideal" molecules that are chemically impossible or extremely difficult to synthesize in a laboratory setting.

2. Methodology: The CombiMOTS Framework

To address these issues, the authors propose a framework that combines Fragment-Based Drug Discovery (FBDD) with Pareto Monte Carlo Tree Search (PMCTS).

A. Synthesizable Search Space Reduction
Instead of searching the near-infinite chemical space, CombiMOTS narrows the search to a "synthesizable space":

• Target-Aware Fragment Extraction: Using Fragment-wise Graph Information Bottleneck (FGIB), the model extracts molecular substructures from known active inhibitors that are most informative regarding target properties.
• Mapping to Building Blocks: These fragments are mapped to industry-available building blocks from the Enamine REAL Space using Tanimoto similarity. This ensures that the "actions" taken during the search correspond to real-world chemical reactions.

B. Pareto Monte Carlo Tree Search (PMCTS)
The core engine is an extension of MCTS that utilizes vectorized optimization rather than scalar rewards:

• Pareto Dominance: Instead of a single score, the model maintains a vector of objectives (e.g., Docking Score A, Docking Score B, QED, SA). A solution is only considered "better" if it dominates another across the Pareto front (i.e., it is no worse in any objective and better in at least one).
• Pareto Upper Confidence Bound (PUCB): The selection phase of the tree search is modified to use a vectorized UCB formula. This allows the algorithm to explore the "Pareto front"—the set of optimal trade-off solutions—rather than converging on a single, potentially imbalanced point.
• Reaction-Based Expansion: The tree grows by applying valid chemical reaction templates to the building blocks, ensuring every node in the tree represents a chemically plausible intermediate or product.

3. Key Contributions

• Novel Algorithm: The first application of Pareto MCTS to the specific problem of dual-target molecule generation.
• Synthesizability-by-Design: By grounding the search in the Enamine REAL Space and reaction templates, the model generates molecules with high practical utility.
• Multi-Objective Robustness: The framework avoids the pitfalls of scalarization, allowing researchers to choose from a diverse set of optimal candidates depending on whether they prioritize potency, safety, or ease of synthesis.

4. Experimental Results

The authors evaluated CombiMOTS on three protein target pairs: GSK3β-JNK3 (Alzheimer’s), EGFR-MET (Lung Cancer), and PIK3CA-mTOR (Cancer).

• Superior Trade-offs: CombiMOTS consistently outperformed baselines (like REINVENT, MARS, and RationaleRL) in generating molecules that balanced high docking scores with high drug-likeness (QED) and high synthetic accessibility (SA).
• Diversity and Novelty: The model produced a wider and more diverse range of unique, novel molecules compared to existing methods, which tended to cluster around specific structural motifs.
• Pareto Optimality: Quantitative analysis showed that CombiMOTS found a larger number of solutions on the true Pareto front, closer to the "utopia point" (the theoretical ideal where all objectives are maximized).
• Case Studies: In silico docking confirmed that generated compounds (e.g., Compound 1 and 2 for GSK3β-JNK3) successfully engaged key binding pockets and hinge regions, mimicking known inhibitors.

5. Significance

CombiMOTS represents a significant step toward rational, automated drug design. By integrating the mathematical rigor of Pareto optimization with the practical constraints of synthetic chemistry, it provides a powerful tool for medicinal chemists to discover complex, multi-functional therapeutic agents that are actually capable of being manufactured and tested in clinical settings.