Research map / Probabilistic and scientific modeling / Molecules and proteins

Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design · 1 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.

Causal activation-patching, steering, and probing study of ESMFold, OpenFold, and Boltz-1 folding trunks finds a shared two-stage mechanism (early sequence-to-pair charge write-in, late pairwise distance geometry) and linearly alignable pairwise representations across models.

Problem

How folding models (ESMFold, OpenFold, Boltz-1) internally decide whether a region folds as a beta hairpin or alpha helix, and whether different architectures share the same internal mechanism.

Approach

Uses block-wise activation patching of sequence and pairwise representations between donor and target proteins to localize structural commitment, then links early-block charge information transfer (seq2pair) and late-block distance encoding (pairwise) to hairpin formation via linear steering, and shows cross-model Procrustes alignment enables substitution of pairwise states.

Key results

• Sequence patching at block 0 induces donor motif in ~40-45% of cases, while pairwise patching is effective in late blocks (peaking near block 35 in ESMFold and OpenFold).
• Linear distance probes on pairwise representations reach R^2 ~0.9 in late blocks across all three models.
• Procrustes-aligned OpenFold pairwise representations reach R^2 ~0.9 into ESMFold space, and Boltz-1 reaches ~0.75; cross-model patches reproduce the late-block window (ESMFold->ESMFold ~22%, OpenFold->ESMFold ~15%, Boltz-1->ESMFold ~10%).

Surprising findings

• Boltz-1 shows a broad pairwise patching window across nearly the entire trunk, attributed to the absence of seq2pair write-back in its Pairformer.
• Input intervention (literal sequence substitution) induces hairpins at a lower rate than trunk patching, indicating trunk representations carry information beyond raw sequence.

Limitations

• Limited to short local motifs (hairpins and helices in 12-25 residue windows); larger motifs untested.
• Counterfactual structures show low pLDDT and inverse folding fails to recover matching sequences; some interventions produce steric clashes.
• Does not apply to pairwise-representation-free architectures such as SimpleFold.

Experimental setup

Datasets: CATH Class 1 alpha-helical domains PDB PISCES culled chains (cullpdb_pc25.0) DSSP-annotated motifs

Metrics: DSSP-based hairpin/helix induction rate patching success rate ROC-AUC R^2 of linear distance probes CKA Procrustes R^2 cross-strand C-alpha distance

Baselines: Input intervention Encoder patching Structure module patching Shuffled-correspondence control

Builds on: activation patching difference-of-means steering vectors Centered Kernel Alignment whitened Procrustes alignment DSSP

Closest papers by meaning

• ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search (NeurIPS 2025 · similarity 0.86)
• Boltzmann-Aligned Inverse Folding Model as a Predictor of Mutational Effects on Protein-Protein Interactions (ICLR 2025 · similarity 0.86)
• Learning conformational ensembles of proteins based on backbone geometry (NeurIPS 2025 · similarity 0.86)
• Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.86)
• A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning (NeurIPS 2026 · similarity 0.86)
• FoldToken: Learning Protein Language via Vector Quantization and Beyond (AAAI 2025 · similarity 0.86)