Research map / Vision-language and multimodal / Human and embodied modeling

A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning

NeurIPS 2026 · Vision-language and multimodal › Human and embodied modeling · approach: Medical vision-language models · 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.

Probes and distills the atom-level ligand representations of the co-folding model Boltz2 into small-molecule tasks (ADMET prediction, generative modeling, structure-guided ligand optimization), finding it competitive with specialized molecular foundation models.

Problem

Small-molecule foundation models are pretrained on standalone molecules, whereas protein-ligand co-folding models expose atom-level ligand-protein interaction supervision; it is unclear whether co-folding representations transfer to standalone small-molecule tasks.

Approach

Repurpose Boltz2's ligand pair and single representations (from the Pairformer trunk) as frozen features for standalone molecular property prediction, and use representation-alignment distillation to accelerate diffusion-based generation and to provide dense supervision in online RL for ligand discovery.

Key results

• On the TDC ADMET benchmark, Boltz2 achieves the best average among non-ensemble methods on 9 of 22 tasks, and Boltz2 with MiniMol ensemble is best on 9 tasks.
• Representation alignment with Boltz2 improves GruM* generation: FCD drops from 1.49 to 1.31 and NSPDK from 0.0007 to 0.0005, with reported 2x faster training.
• Boltz2 (0.30 Caco2 MAE) outperforms UniMol2, a 1B Pairformer pretrained on standalone 3D molecules, on most ADMET tasks despite less data (750K complexes vs 884M molecules).

Surprising findings

• Boltz2 underperforms on metabolism substrate prediction (e.g., CYP2D6 substrate 0.52 vs 0.74 for best baseline), and UniMol2 shows the same weakness, suggesting a limitation of 3D structure-based learning.
• Combining Boltz2 with the less aligned MolE yields larger gains on some benchmarks than combining with the more aligned and stronger-performing MiniMol.

Limitations

• Substrate prediction tasks remain weak, indicating limits of structure-based representations for local atomistic mechanisms.
• Only a frozen backbone is used via probing and distillation; backbone fine-tuning is left for future work.
• Some baseline gains may depend on bioassay label leakage in pretraining datasets (MiniMol), which complicates comparisons.

Experimental setup

Datasets: TDC ADMET benchmark ZINC250k

Benchmarks: TDC ADMET benchmark ZINC250k unconditional generation SynFlowNet-Boltz structure-guided ligand discovery

Metrics: AUROC AUPRC MAE Spearman Validity FCD NSPDK Novelty Uniqueness Scaffold similarity

Baselines: MolE KPGT MiniMol QIP UniMol2 MolGPS UniQSAR GruM GBD DeFoG TopBF Marginal SID

Builds on: Boltz representation alignment GruM SynFlowNet-Boltz CKNNA

Closest papers by meaning

• Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.88)
• Self-supervised Pocket Pretraining via Protein Fragment-Surroundings Alignment (ICLR 2024 · similarity 0.87)
• Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.86)
• Few-Step Cofolding with All-Atom Flow Maps (NeurIPS 2026 · similarity 0.86)
• Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.86)
• BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.86)