Research map / Probabilistic and scientific modeling / Molecules and proteins
P(all-atom) Is Unlocking New Path For Protein Design
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.
Pallatom generates all-atom protein structures and sequences jointly by diffusing an atom14 point cloud with a dual-track residue/atom decoder, trained on PDB and AFDB.
Problem
Protein generation usually models structure and sequence in separate steps (backbone then sequence, or hallucination plus sequence redesign), and existing all-atom generators such as Protpardelle yield sequences that fail to encode the intended fold, so the joint distribution P(structure, seq) is no
Approach
Learn P(all-atom) directly with an EDM-style diffusion over a fixed-size atom14 point cloud (virtual atoms pad every residue to 14 atoms), using a dual-track residue/atom AtomDecoder with traversing atom embeddings and differentiable recycling, and predict sequence from aggregated atom embeddings with a linear head.
Key results
- Pallatom reaches 85.03% DES-aa in CO-DESIGN 1 mode, versus 62.74% for Multiflow and 43.14% for ProteinGenerator, with 291 structural clusters out of 1466 designable sequences.
- In PMPNN 1 mode, Pallatom achieves 89.89% DES-bb (w) and 318 DIV-str clusters, the highest diversity among compared methods.
- Ablation: replacing atom14 with hybrid14 drops DES-aa from 87% to 5%; removing differentiable recycling drops DES-aa to 21%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use atom14 instead of hybrid14 representation (ablation) | Pallatom · protein generation | DES-aa: 5% -> 87% | Helped |
| use atom14 instead of hybrid14 representation (ablation) | Pallatom · protein generation | DES-bb(w): 68% -> 95% | Helped |
| use atom14 instead of hybrid14 representation (ablation) | Pallatom · protein generation | DES-bb(wo): 82% -> 95% | Helped |
| add differentiable recycling (ablation) | Pallatom · protein generation | DES-aa: 21% -> 87% | Helped |
| add differentiable recycling (ablation) | Pallatom · protein generation | DES-bb(w): 93% -> 95% | Helped |
| add differentiable recycling (ablation) | Pallatom · protein generation | DES-bb(wo): 95% -> 95% | No clear effect |
| increase sampling step scale η from 1.75 to 2.75 | Pallatom · L=100 sampling | DES-aa: 57% -> 94% | Helped |
| increase sampling step scale η from 1.75 to 2.75 | Pallatom · L=100 sampling | DIV-str: 56 -> 55 | Mixed |
| reduce additional noise level γ0 from 0.2 to 0.1 | Pallatom · L=100 sampling | DES-aa: 87% -> 89% | Helped |
| reduce additional noise level γ0 from 0.2 to 0.1 | Pallatom · L=100 sampling | DIV-str: 64 -> 52 | Hurt |
| use Pallatom rather than Multiflow | all-atom protein generation · CO-DESIGN 1 | DES-aa: 62.74% -> 85.03% | Helped |
| use Pallatom rather than Multiflow | protein generation · PMPNN 1 | DIV-str: 184 -> 318 | Helped |
Surprising findings
- Without any fixed-backbone design training, the linear sequence head reaches sequence quality comparable to ProteinMPNN, according to the authors.
- The hybrid14 sequence-guided alternative to atom14 sharply reduced designability (DES-aa 5%), contrary to the expectation that sequence guidance would help.
Limitations
- Evaluated only on small monomer proteins of lengths 60-120 in the main results; generality to complex systems is untested.
- Sequence is decoded from the atom point cloud with a low-temperature softmax, and sequence quality is compared against ProteinMPNN only on designability proxies.
- Designability relies on ESMFold/pLDDT and aaRMSD proxies rather than experimental validation.
Experimental setup
Datasets: PDB AlphaFold Database
Benchmarks: CO-DESIGN 1 mode (DES-aa, DIV-str/seq, NOV-str) PMPNN 1 mode (DES-bb, DIV-str, NOV-str)
Metrics: DES-aa DES-bb DIV-str DIV-seq NOV-str aaRMSD pLDDT
Baselines: Protpardelle ProteinGenerator Multiflow RFdiffusion ProteinMPNN
Builds on: EDM AlphaFold3 Protpardelle RFdiffusion
Closest papers by meaning
- An All-Atom Generative Model for Designing Protein Complexes (ICML 2025 · similarity 0.87)
- Generative Modeling of Full-Atom Protein Conformations using Latent Diffusion on Graph Embeddings (NeurIPS 2025 · similarity 0.87)
- ESM All-Atom: Multi-Scale Protein Language Model for Unified Molecular Modeling (ICML 2024 · similarity 0.86)
- Sidechain conditioning and modeling for full-atom protein sequence design with FAMPNN (ICML 2025 · similarity 0.86)
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.86)
- Elucidating the Design Space of Multimodal Protein Language Models (ICML 2025 · 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.