Research map / Probabilistic and scientific modeling / Molecules and proteins

Sample Efficient Generative Molecular Optimization with Joint Self-Improvement

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation · 0 citations (OpenAlex)

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 Joint Self-Improvement, which trains a joint generative-predictive model (Hyformer) and samples molecules with a self-improving tilted-sampling scheme that uses the predictor to bias the generator, improving sample efficiency for offline and online molecular optimization under small oracle budgets.

Problem

Generative molecular optimization needs many expensive oracle evaluations; in the offline setting, decoupled surrogates suffer distribution shift and objective hacking as the generator drifts, and RL fine-tuning is high-variance and sample-inefficient.

Approach

Align generator and surrogate through a joint model p(x,y)=p(y|x)p(x) with shared parameters trained by likelihood loss, and bias generation at inference time with the predictive part via iterative global-advantage logit tilting over stochastic beam search samples, instead of RL policy-gradient updates.

Key results

• Offline setting: Joint Self-Improvement reaches 76.94 hit ratio on 5HT1B vs 56.91 for GeneticGFN and 62.93 for RaM under a 3,000-evaluation budget.
• Online setting: achieves 69.40 hit ratio on PARP1 vs 57.98 for SATURN and 84.91 on 5HT1B vs 68.19 for SATURN.
• Ablation: removing self-improving sampling drops PARP1 hit ratio from 50.31 to 2.90, while removing joint modeling lowers it to 46.11.

Surprising findings

• Replacing the self-improving sampler with Best-of-N rejection sampling yields near-zero hit ratios across targets in the ablation, despite the same predictor.

Limitations

• Surrogate inaccuracy may limit performance on challenging tasks.
• No uncertainty estimates incorporated into the sampling scheme.
• Hit ratio used instead of multi-objective metrics such as hypervolume.

Experimental setup

Datasets: ZINC250K

Benchmarks: docking score optimization on PARP1, FA7, 5HT1B, BRAF, JAK2

Metrics: Hit Ratio IntDiv1 Time per molecule

Baselines: REINVENT SATURN GeneticGFN RaM JT-VAE FREED-QS MOOD Augmented Memory GEAM

Builds on: Hyformer REINVENT RL-as-inference Gumbel-Top-k trick Best-of-N sampling

Closest papers by meaning

• Offline Model-based Optimization for Real-World Molecular Discovery (ICML 2025 · similarity 0.87)
• Empower Structure-Based Molecule Optimization with Gradient Guided Bayesian Flow Networks (ICML 2025 · similarity 0.86)
• On the Design Space of Discrete Diffusion Online Adaptation for Molecular Optimization (NeurIPS 2026 · similarity 0.86)
• Domain-Agnostic Molecular Generation with Chemical Feedback (ICLR 2024 · similarity 0.84)
• DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · similarity 0.84)
• FlashMol: High-Quality Molecule Generation in as Few as Four Steps (NeurIPS 2026 · similarity 0.84)