Research map / Probabilistic and scientific modeling / Molecules and proteins

On the Design Space of Discrete Diffusion Online Adaptation for Molecular Optimization

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based molecule design with diffusion · 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.

Systematic study of composable online-adaptation components (Thompson acquisition, CVaR and density-entropy reward shaping, replay, invalid-SMILES penalty) for discrete-diffusion molecular and protein optimization under oracle and GPU-hour budgets.

Problem

Online test-time adaptation of pretrained discrete diffusion models for molecular and protein optimization with a limited oracle budget; it is unclear how acquisition, reward shaping, debiasing, replay and validity control interact inside one loop.

Approach

Treats the online loop as a harness with swappable knobs and tests their complementarity via leave-one-out ablations; proposes a cheap single-sample estimator of the log-likelihood (masked-token log-softmax sum) as a debiasing reward term and combines the knobs into a recipe that beats offline fine-tuning and inference-time search.

Key results

• Online adaptation with the full loop reaches the best inference-time search baseline's final score in about 58x fewer oracle calls on small-molecule tasks (about 16x on protein tasks).
• Leave-one-out ablations show Density Entropy Regularization producing the largest drop on small molecules at later budgets; CVaR and debiasing effects are reported as close to additive on max reward.
• Without the invalid-SMILES penalty, validity collapses within the first hour of online training on small-molecule runs.

Surprising findings

• Adding inference-time search (best-of-N, beam, MCTS, DFKC, SMC) to the online loop did not improve over plain online fine-tuning at fixed wall budget because search overhead consumes the budget.
• Protein-fitness ablation effects are weak, and search baselines converge to the full loop by 5k calls there.

Limitations

• Model-debiasing uses a cheap single-mask log-softmax estimator rather than the unbiased ELBO, so it is only an approximation.
• Main results rely on FlashAffinity, a learned oracle; Boltz-2 is used only for selected comparisons.
• Hyperparameters were not tuned per method; ablations use one set of defaults with 5 seeds.

Experimental setup

Datasets: SAFE dataset V2 (GenMol pretraining) 6 PDB binding-pocket targets (2VT4, 5SDV, 6CM4, 7BKC, 7C7M, 7YLL) GB1, CreiLOV, TrpB fitness data

Benchmarks: six small-molecule binding-affinity tasks GB1 CreiLOV TrpB

Metrics: best reward so far (top-1) top-k mean top-10% mean normalized reward SMILES validity pairwise Tanimoto diversity QED SA

Baselines: offline DDPP-LB best-of-N beam search MCTS discrete Feynman-Kac correction SMC unconditional sampling

Builds on: DDPP VIDD CVaR reward shaping GenMol MDLM

Closest papers by meaning

• Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.88)
• Reward-Guided Iterative Refinement in Diffusion Models at Test-Time with Applications to Protein and DNA Design (ICML 2025 · similarity 0.88)
• Flow Density Control: Generative Optimization Beyond Entropy-Regularized Fine-Tuning (NeurIPS 2025 · similarity 0.87)
• Bridging Model-Based Optimization and Generative Modeling via Conservative Fine-Tuning of Diffusion Models (NeurIPS 2024 · similarity 0.87)
• Uncertainty-Aware Multi-Objective Reinforcement Learning-Guided Diffusion Models for 3D De Novo Molecular Design (NeurIPS 2025 · similarity 0.87)
• Fine-Tuning Discrete Diffusion Models via Reward Optimization with Applications to DNA and Protein Design (ICLR 2025 · similarity 0.87)