Research map / Probabilistic and scientific modeling / Molecules and proteins
Sidechain conditioning and modeling for full-atom protein sequence design with FAMPNN
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.
Introduces FAMPNN, a full-atom MPNN-GVP sequence design model that jointly generates sequence identity (cross-entropy) and sidechain coordinates (per-token diffusion) via iterative masked unmasking.
Problem
Fixed-backbone protein sequence design models ignore sidechain conformation during sequence generation, so they must infer sidechain interactions only from backbone geometry and sequence labels.
Approach
Model per-token discrete amino acid identity and continuous sidechain coordinates jointly with a combined categorical cross-entropy and per-token Euclidean diffusion loss in a masked language modeling framework, sampling sequence and sidechains iteratively (MaskGIT-style).
Key results
- FAMPNN achieves 49.66% single-step sequence recovery on CATH 4.2, above ProteinMPNN (45.96%).
- On CASP15 sidechain packing, FAMPNN (0.0 Å) reaches 0.690 Å all-atom RMSD, better than FlowPacker (0.765 Å).
- Zero-shot SKEMPIv2 binding prediction is significantly better than unsupervised baselines; stability and antibody-antigen gains are modest but consistent.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the sidechain packing objective (ablation) | MPNN-GVP · CATH 4.2 test set | sequence accuracy | Helped |
| add full-atom conditioning (ablation) | MPNN-GVP · CATH 4.2 test set | sequence accuracy | Helped |
| add full-atom conditioning and sidechain packing (ablation) | FAMPNN · SKEMPIv2 | Spearman correlation | Helped |
| add full-atom conditioning and sidechain packing (ablation) | FAMPNN · antibody-antigen binding affinity datasets | Spearman correlation | No clear effect |
| provide partial sidechain context along with sequence context instead of partial sequence context alone (ablation) | FAMPNN · protein-protein interfaces | sequence accuracy | Helped |
| increase available partial sequence or sidechain context (ablation) | FAMPNN · CASP targets | sidechain packing RMSD | Helped |
| use 10 iterative unmasking steps instead of 1 (ablation) | FAMPNN · RFdiffusion-generated de novo backbones | self-consistency | Helped |
Surprising findings
- Full-atom conditioning does not improve antibody-antigen binding prediction, which the authors attribute to inaccurate input structures.
- Single-step sampling is insufficient for self-consistency; about 10 iterative steps are needed.
Limitations
- Requires a starting backbone; cannot design backbone jointly.
- Does not yet condition on ligands or model ligand-sidechain interactions.
- Full-atom context did not help antibody-antigen binding prediction, likely due to inaccurate input structures.
Experimental setup
Datasets: CATH 4.2 CASP13-15 (excluded homologues) RFdiffusion de novo backbones SKEMPIv2 S669 Megascale FireProtDB Bushuiev et al. SKEMPIv2 split
Benchmarks: CATH 4.2 CASP15 RFdiffusion de novo backbones (ProteinBench) SKEMPIv2 S669 Megascale FireProtDB
Metrics: sequence recovery perplexity scTM scRMSD atom RMSD Spearman correlation
Baselines: ProteinMPNN GVP PiFold ESM-IF1 Frame2Seq AlphaDesign 3DCNN ChromaDesign LigandMPNN FlowPacker AttnPacker
Builds on: ProteinMPNN GVP MaskGIT EDM diffusion loss for autoregressive models
Closest papers by meaning
- Full-Atom Peptide Design based on Multi-modal Flow Matching (ICML 2024 · similarity 0.87)
- Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings (NeurIPS 2025 · similarity 0.87)
- Pre-Training Protein Bi-level Representation Through Span Mask Strategy On 3D Protein Chains (ICML 2024 · similarity 0.87)
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.87)
- P(all-atom) Is Unlocking New Path For Protein Design (ICML 2025 · similarity 0.86)
- CarbonNovo: Joint Design of Protein Structure and Sequence Using a Unified Energy-based Model (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.