Research map / Probabilistic and scientific modeling / Molecules and proteins
Offline Model-based Optimization for Real-World Molecular Discovery
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 MolStitch, an offline multi-objective molecular optimization framework that stitches offline molecules with a learned StitchNet, ranks them with a pairwise rank-based proxy, and fine-tunes REINVENT via IPO-style preference optimization.
Problem
Molecular discovery usually relies on online oracle queries (wet-lab experiments) that are slow and costly; offline multi-objective molecular optimization must improve a generative model using only a fixed offline dataset without repeated oracle evaluation.
Approach
Stitches molecule pairs from the offline dataset into new molecules with a StitchNet (pretrained by imitating a rule-based crossover, then refined with proxy-derived chemical feedback), replaces score regression with a pairwise rank-based proxy, samples objective weights via Dirichlet priority sampling, and fine-tunes REINVENT with IPO-style preference optimization.
Key results
- On GSK3β+JNK3+QED+SA, MolStitch reaches HV 0.352±0.080 and R2 2.953, versus 0.281 HV for the best prior baseline Saturn (in Table 1, RaM 0.249).
- On docking score optimization, MolStitch achieves the highest HV on all five targets, e.g. 0.560 on parp1 versus 0.544 for BootGen.
- Under the semi-offline setting, MolStitch reaches HV 0.493 on the four-objective MPO task versus 0.382 for Saturn.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MolStitch rather than the strongest reported baseline | Offline molecular property optimization · GSK3β+JNK3 | HV: 0.540±0.113 -> 0.579±0.070 | Helped |
| use MolStitch rather than the strongest reported baseline | Offline molecular property optimization · GSK3β+JNK3+QED | HV: 0.309±0.087 -> 0.403±0.065 | Helped |
| use MolStitch rather than the strongest reported baseline | Offline molecular property optimization · GSK3β+JNK3+QED+SA | HV: 0.281±0.058 -> 0.352±0.080 | Helped |
| use MolStitch rather than the strongest reported baseline | Offline docking score optimization · parp1, jak2, braf, fa7, 5ht1b | HV: 0.544±0.032 -> 0.560±0.037; 0.498±0.030 -> 0.515±0.041; 0.528±0.034 -> 0.554±0.042; 0.445±0.053 -> 0.451±0.061; 0.545±0. | Helped |
| add StitchNet to the no-component baseline (ablation) | Offline MOMO (GSK3β+JNK3+QED+SA) · GSK3β+JNK3+QED+SA | HV: 0.171±0.026 -> 0.193±0.053 | Helped |
| add Priority Sampling to StitchNet without Rank-based Proxy (ablation) | Offline MOMO (GSK3β+JNK3+QED+SA) · GSK3β+JNK3+QED+SA | HV: 0.193±0.053 -> 0.220±0.054 | Helped |
| add Rank-based Proxy to the setup with StitchNet and Priority Sampling (ablation) | Offline MOMO (GSK3β+JNK3+QED+SA) · GSK3β+JNK3+QED+SA | HV: 0.220±0.054 -> 0.352±0.080 | Helped |
| add StitchNet instead of stochastic sampling for data augmentation (ablation) | Offline MOMO (GSK3β+JNK3+QED+SA) · GSK3β+JNK3+QED+SA | HV: 0.251±0.084 -> 0.352±0.080 | Helped |
| replace the rule-based crossover operator with StitchNet for data augmentation (ablation) | Offline MOMO (GSK3β+JNK3+QED+SA) · GSK3β+JNK3+QED+SA | HV: 0.302±0.072 -> 0.352±0.080 | Helped |
| use IPO instead of RLHF for preference optimization (ablation) | StitchNet-based offline MOMO · GSK3β+JNK3+QED+SA | HV: 0.232±0.071 -> 0.344±0.082 | Helped |
| use IPO instead of DPO for preference optimization (ablation) | StitchNet-based offline MOMO · GSK3β+JNK3+QED+SA | HV: 0.327±0.081 -> 0.344±0.082 | Helped |
| add a separate Rank-based Proxy to StitchNet with IPO (ablation) | StitchNet-based offline MOMO · GSK3β+JNK3+QED+SA | HV: 0.344±0.082 -> 0.352±0.080 | Helped |
Surprising findings
- Score-based regression proxies underperform a pairwise rank-based proxy; replacing RP with a score-based proxy lowers HV from 0.352 to 0.289 in the ablation.
Limitations
- Relies on the assumption that structurally similar stitched molecules have similar objective scores, used to approximate their labels.
- Evaluation is on in silico oracles and benchmarks rather than wet-lab validation.
Experimental setup
Datasets: ZINC MPO offline dataset (GSK3β, JNK3, QED, SA) docking score dataset (parp1, fa7, jak2, braf, 5ht1b)
Benchmarks: molecular property optimization (GSK3β, JNK3, QED, SA) docking score optimization (parp1, jak2, braf, fa7, 5ht1b)
Metrics: hypervolume (HV) R2 indicator average property score (top-10/top-100) scaffold and carbon-skeleton diversity
Baselines: REINVENT AugMem GraphGA Saturn GeneticGFN Grad COMs IOM RoMA Ensemble Proxy BIB BootGen
Builds on: REINVENT DPO IPO offline model-based optimization trajectory stitching
Closest papers by meaning
- Sample Efficient Generative Molecular Optimization with Joint Self-Improvement (NeurIPS 2026 · similarity 0.87)
- Domain-Agnostic Molecular Generation with Chemical Feedback (ICLR 2024 · similarity 0.87)
- On the Design Space of Discrete Diffusion Online Adaptation for Molecular Optimization (NeurIPS 2026 · similarity 0.86)
- Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · similarity 0.86)
- UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery (ICLR 2025 · similarity 0.85)
- Multi-Objective Molecular Design Through Learning Latent Pareto Set (AAAI 2025 · similarity 0.85)
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.