Research map / Probabilistic and scientific modeling / Molecules and proteins
CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CombiMOTS, a Pareto Monte Carlo Tree Search over a synthesizable Enamine REAL fragment/building-block space that generates dual-target inhibitors using vectorized docking and activity objectives.
Problem
Dual-target molecule generation needs to balance target affinity, drug-likeness and synthesizability, but existing methods collapse multiple objectives into a scalarized reward and often ignore synthetic routes.
Approach
Extends MCTS with Pareto UCB selection over vectorized objectives (predicted dual-activity and dual-docking scores), and reduces the search space to FGIB-extracted target-aware fragments mapped to Enamine REAL building blocks and reaction templates.
Key results
- On GSK3beta-JNK3, CombiMOTS reaches 88.67% diversity versus 64.52% (REINVENT) and 70.92% (RationaleRL), with 100% uniqueness.
- CombiMOTS yields 23.74% dual-active rate with FGIB fragments versus 14.45% (naive BRICS) and 10.90% (MiCaM) in the fragmentation ablation.
- Average normalized R2-distance to utopia on EGFR-MET is 0.8419 for CombiMOTS versus 0.8518 (REINVENT) and 0.8845 (MARS), with the largest Pareto set (268).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Pareto-based CombiMOTS rather than the strongest baseline | dual-target molecule generation · GSK3β-JNK3 | diversity (%): 70.92±0.01 -> 88.67±0.52 | Helped |
| use Pareto-based CombiMOTS rather than the strongest baseline | dual-target molecule generation · EGFR-MET | diversity (%): 91.89±0.10 -> 90.75±0.40 | Hurt |
| use Pareto-based CombiMOTS rather than the strongest baseline | dual-target molecule generation · PIK3CA-mTOR | diversity (%): 73.82±2.12 -> 90.88±0.36 | Helped |
| replace FGIB with BRICS fragments (ablation) | CombiMOTS · GSK3β-JNK3 | dual-active rate (%): 23.74% -> 14.45% | Hurt |
| replace FGIB with MiCaM fragments (ablation) | CombiMOTS · GSK3β-JNK3 | dual-active rate (%): 23.74% -> 10.90% | Hurt |
| lower the Tanimoto threshold from 0.4 to 0.3 (ablation) | CombiMOTS · GSK3β-JNK3 | dual-active rate (%): 23.74% -> 20.90% | Hurt |
| raise the Tanimoto threshold from 0.4 to 0.5 (ablation) | CombiMOTS · GSK3β-JNK3 | dual-active rate (%): 23.74% -> 11.86% | Hurt |
| raise the Tanimoto threshold from 0.4 to 0.6 (ablation) | CombiMOTS · GSK3β-JNK3 | diversity (%): 88.67±0.52 -> 78.73±1.11 | Hurt |
| replace Pareto optimization with equal-weight scalarization (ablation) | CombiMOTS · GSK3β-JNK3 | dual-active molecules found: 3,662 -> 1,488 | Hurt |
| replace Pareto optimization with equal-weight scalarization (ablation) | CombiMOTS · GSK3β-JNK3 | QED | Mixed |
| add QED and SA as two objectives (ablation) | CombiMOTS · GSK3β-JNK3 | dual-active molecules found: 8,606 -> 5,748 | Hurt |
| add QED and SA as two objectives (ablation) | CombiMOTS · GSK3β-JNK3 | docking score | Hurt |
Surprising findings
- A scalarized equal-weight version of CombiMOTS found 1,488 dual actives versus 3,662 with Pareto selection under the same budget.
- Adding QED and SA as objectives reduced dual-active count (5,748 vs 8,606) and slowed convergence.
Limitations
- Generated compounds are restricted to Enamine building blocks and reaction templates, limiting the reachable chemical space.
- Activity and docking are in silico predictions; some generated molecules are predicted toxic and no experimental validation is reported.
- Runtime grows with the number of objectives, and the search space size strongly affects performance via the similarity threshold.
Experimental setup
Datasets: GSK3beta-JNK3 benchmark ExCAPE-DB PubChem Enamine REAL Space ClinTox
Benchmarks: GSK3beta-JNK3 EGFR-MET PIK3CA-mTOR DHODH-RORgammat
Metrics: validity uniqueness novelty diversity activity success rate docking score QED SA score normalized R2-distance to utopia point Pareto optimal set size
Baselines: RationaleRL REINVENT MARS MolSearch Mothra AIxFuse
Builds on: Monte Carlo Tree Search Pareto MCTS Fragment-wise Graph Information Bottleneck BRICS REINVENT
Closest papers by meaning
- Drug Discovery with Dynamic Goal-aware Fragments (ICML 2024 · similarity 0.85)
- PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion (ICML 2025 · similarity 0.85)
- Procedural Synthesis of Synthesizable Molecules (ICLR 2025 · similarity 0.85)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.84)
- Double-Ended Synthesis Planning with Goal-Constrained Bidirectional Search (NeurIPS 2024 · similarity 0.84)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.