Research map / Probabilistic and scientific modeling / Molecules and proteins

Few-Step Cofolding with All-Atom Flow Maps

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design · 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.

Distills the all-atom cofolding diffusion models Boltz-1 and Pearl into few-step sigma-space flow maps (DeCAF) with an SE(3)-aligned denoiser parametrization, and pairs them with a flow-map-lookahead reward-guided search (DeCAF-SEARCH).

Problem

All-atom cofolding diffusion models need many (~200) network evaluations per sample, making sampling and reward-guided inference-time search expensive.

Approach

Define a flow map in EDM sigma-space with a denoiser parametrization that predicts the endpoint and applies Kabsch SE(3) alignment before the loss, enabling stable distillation of EDM cofolding teachers; the flow map's clean-space lookahead then provides reward estimates for search.

Key results

• DeCAF-Pearl matches its Pearl-2026.1.dev teacher on best@5 success rate (76.0 vs 77.0) using 40 NFEs instead of >=200, and outperforms it on PB-valid at best@5 (99.5 vs 98.5).
• On PoseBusters, DeCAF-SEARCH at 30 NFEs exceeds Boltz-1x at 600 NFEs on success rate, a claimed 20x inference cost reduction.
• On Runs N' Poses, DeCAF-SEARCH (MC-GRAD) at 150 NFEs matches Boltz-1x at 800 NFEs on per-target mean RMSD, at 5.3x less compute.

Surprising findings

• Velocity parametrization and consistency distillation fail to sample valid poses (PB-valid 9.2% and 0.0%) while the x0-aligned denoiser parametrization reaches 90.4%.
• Boltz-1x with default settings produces 0% on every metric at 2 steps, requiring step-scale tuning.

Limitations

• Reward signal inherited from Boltz-1x, so failure modes such as non-planar sp2 bonds persist.
• Non-monotone relation between NFE budget and per-target success suggests adaptive search is needed.
• Pearl-2026.1.dev is a development checkpoint, and the Pearl distillation is evaluated on a post-2023 subset of 196 structures.

Experimental setup

Datasets: RCSB PDB structures (pre-2021-09-30) Runs N' Poses PoseBusters

Benchmarks: Runs N' Poses PoseBusters

Metrics: RMSD<2A success rate PB-Valid lDDT-PLI Success Rate best@1 best@5

Baselines: Boltz-1x AlphaFold3 Chai-1 Boltz-2 Pearl-2026.1.dev

Builds on: flow maps EDM MeanFlow consistency trajectory models Boltz-1 Feynman-Kac steering

Closest papers by meaning

• AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.89)
• SE(3)-Stochastic Flow Matching for Protein Backbone Generation (ICLR 2024 · similarity 0.89)
• Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings (NeurIPS 2025 · similarity 0.89)
• Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional (ICML 2025 · similarity 0.89)
• Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.89)
• Energy-Based Flow Matching for Generating 3D Molecular Structure (ICML 2025 · similarity 0.89)