Research map / Probabilistic and scientific modeling / Molecules and proteins
UniIF: Unified Molecule Inverse Folding
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 UniIF, a unified block-graph GNN with frame-based block representations and virtual blocks that performs inverse folding for proteins, RNA, and materials.
Problem
Molecule inverse folding has been studied separately for macromolecules (proteins, RNA) and small molecules/materials, using different unit representations and featurizers, leading to redundant efforts; quadratic cost of full attention also limits scaling to large systems.
Approach
A frame-based block graph that represents amino acids, nucleotides, and atoms as fixed-size blocks with local frames, combined with a geometric block attention GNN that adds directed virtual inter-atom geometric interactions, gated edge attention, and global virtual blocks for long-range dependencies.
Key results
- On CATH4.3 full-length protein design, UniIF reaches 0.53 recovery versus 0.51 for PiFold, and 0.66 on NovelPro versus 0.57 for PiFold.
- On RNA design, UniIF (drop 0.05) achieves 48.94% overall recovery versus 41.53% for RDesign.
- On CHILI-3K material design, UniIF reaches 75.3% recovery versus 63.2% for EdgeCNN, the best baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UniIF instead of KWDesign | protein inverse folding · NovelPro | recovery: 0.64 -> 0.66 | Helped |
| use UniIF instead of KWDesign | protein inverse folding · CATH4.3 Full | recovery: 0.60 -> 0.53 | Hurt |
| use UniIF instead of KWDesign | protein inverse folding · CASP | recovery: 0.56 -> 0.51 | Hurt |
| use UniIF instead of RDesign, with tertiary structure input only | RNA inverse folding · RNA design All | recovery: 41.53 ± 0.38 -> 48.94 ± 0.37 | Helped |
| use UniIF instead of EdgeCNN | material inverse folding · CHILI-3K | recovery: 63.2±0.9 -> 75.3±1.2 | Helped |
| replace geometric interaction with VFN's operation (ablation) | UniIF · CATH4.3 Full | recovery: 0.53 -> 0.52 | Hurt |
| add geometric dot product features (ablation) | UniIF · CATH4.3 Full | recovery: 0.52 -> 0.53 | Helped |
| use gated edge attention instead of PiGNN attention (ablation) | UniIF · CATH4.3 Full | recovery: 0.52 -> 0.53 | Helped |
| add global virtual frames (ablation) | UniIF · CATH4.3 Full | recovery: 0.52 -> 0.53 | Helped |
| use dropout rate 0.05 instead of 0.0 (ablation) | UniIF · RNA design All | recovery: 44.29 ± 0.29 -> 48.94 ± 0.37 | Helped |
| use dropout rate 0.05 instead of 0.1 (ablation) | UniIF · RNA design All | recovery: 46.00 ± 0.38 -> 48.94 ± 0.37 | Helped |
| use dropout rate 0.05 instead of 0.2 (ablation) | UniIF · RNA design All | recovery: 47.19 ± 0.45 -> 48.94 ± 0.37 | Helped |
Surprising findings
- Padding proteins to maximum length for transformer-style implementation greatly increased GPU and runtime cost, so the authors use GNN without padding.
Limitations
- RNA design uses only tertiary structure input, not secondary structure, which is harder than baselines' setting.
- Overfitting tendency: the model fits training data better than PiFold, requiring dropout regularization.
- Material design evaluated only on the CHILI-3K mono-metal oxide dataset.
Experimental setup
Datasets: CATH4.3 CASP15 NovelPro RDesign RNA dataset CHILI-3K
Benchmarks: CATH4.3 CASP15 NovelPro RDesign RNA benchmark CHILI-3K
Metrics: recovery rate (%)
Baselines: StructGNN GraphTrans GCA GVP AlphaDesign ProteinMPNN PiFold LMDesign KWDesign VFN RDesign SeqRNN
Builds on: PiFold AlphaFold2 GVP PiGNN
Closest papers by meaning
- All-atom inverse protein folding through discrete flow matching (ICML 2025 · similarity 0.87)
- AlphaFold Database Debiasing for Robust Inverse Folding (NeurIPS 2025 · similarity 0.85)
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.85)
- UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery (ICLR 2025 · similarity 0.85)
- Bridge-IF: Learning Inverse Protein Folding with Markov Bridges (NeurIPS 2024 · similarity 0.85)
- Two Stages of Folding: Convergent Mechanisms in AI Protein Folding Trunks (NeurIPS 2026 · similarity 0.85)
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.