Research map / Probabilistic and scientific modeling / Molecules and proteins

AlphaFold Database Debiasing for Robust Inverse Folding

NeurIPS 2025 · 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.

Proposes DeSAE, an SE(3) structure autoencoder pretrained to repair corrupted PDB backbones, and applies it to AFDB structures to produce debiased training data that improves inverse folding recovery on CATH and TS benchmarks.

Problem

AlphaFold DB structures differ systematically from experimental PDB structures in local geometry (narrower dihedral and bond-length distributions), and models for inverse folding trained on AFDB generalize worse to experimental proteins.

Approach

Train an SE(3) frame-based autoencoder to reconstruct native PDB backbones from locally corrupted inputs (one backbone atom replaced by the centroid of the other three), then pass AFDB structures through it to obtain debiased training structures for inverse folding.

Key results

• Debiased AFDB improves recovery over raw AFDB for all five inverse folding models on CATH 4.3, e.g., PiFold 17.74% to 35.38% and GVP 24.35% to 31.15%.
• Debiased AFDB-XL (DeSAE pretrained on full PDB) gives further modest gains, e.g., PiFold 35.76% and ProteinMPNN 35.55%, and lower perplexity than Debiased AFDB.
• DeSAE has 5.9M parameters and processes about 20k AFDB structures in 3 minutes on a single A100.

Measured effects reported in the paper

Surprising findings

• Models trained on raw AFDB perform best on an AFDB-version CATH4.2 test set, and Debiased AFDB trained models do not reach PDB-trained models on several architectures there (e.g., ProteinMPNN 45.34 vs 41.87 is better, but PiFold 46.35 vs 44.31 also better; AFDB-trained ProteinMPNN is worst at 30.48)
• Debiased AFDB-trained models still fall short of PDB-trained models on experimental CATH 4.3 (e.g., ProteinMPNN 35.22 vs 42.33 PDB).

Limitations

• Focuses only on backbone geometry, ignoring sidechain orientation and higher-order details.
• Debiased AFDB still underperforms PDB-trained inverse folding models on experimental benchmarks.
• Larger-scale DeSAE pretraining (XL) yields only marginal recovery gains, which the authors attribute to local-feature focus.

Experimental setup

Datasets: AFDB PDB CATH 4.2 CATH 4.3 TS50 TS500

Benchmarks: CATH 4.2 CATH 4.3 TS50 TS500

Metrics: sequence recovery rate perplexity RMSD

Baselines: StructGNN GraphTrans GVP ProteinMPNN PiFold

Builds on: AlphaFold2 Chroma PiFold ProteinMPNN

Closest papers by meaning

• Inverse problems with experiment-guided AlphaFold (ICML 2025 · similarity 0.88)
• Fast Uncovering of Protein Sequence Diversity from Structure (ICLR 2025 · similarity 0.87)
• Protein Inverse Folding From Structure Feedback (NeurIPS 2025 · similarity 0.86)
• Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.86)
• Few-Step Cofolding with All-Atom Flow Maps (NeurIPS 2026 · similarity 0.86)
• All-atom inverse protein folding through discrete flow matching (ICML 2025 · similarity 0.86)