henhaobio
Sign in

bioRxivBioengineeringPreprint

Ambient Proteins: Training Diffusion Models on Low Quality Structures

Giannis Daras, Jeffrey Ouyang-ZhangiD, Krithika RavishankariD, William Daspit, Costis Daskalakis, Qiang LiuiD, Adam R. KlivansiD, Daniel J. DiaziD

PDFDOICodeData

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (5 Jul 2025) · CC BY 4.0

Treating low-confidence AlphaFold structures as noisy data at pLDDT-dependent diffusion times lets a 16.7M-parameter model generate long proteins with higher diversity, designability and novelty than Proteína.

AlphaFold-derived training sets are usually filtered by average pLDDT, which biases them toward small, simple folds and discards domain-level information in longer proteins. This work keeps low-confidence structures by assigning each one a noise-dependent range of usable diffusion times instead of discarding it.

Key results

  1. ShownOn long unconditional backbone generation, the model beats the 200M-parameter Proteína in both designability and diversity at lengths 700 and 800, with scores staying high across 300 to 800 residues.

    At 700 residues: diversity 45% to 86%, designability 68% to 86% vs Proteína. At 800 residues both metrics are 68%, with gains of 24% (designability) and 45% (diversity). Evaluation used 100 backbones per length (300 to 800). · Abstract; 4.1 Comparisons on unconditional generation of…

  2. ShownAgainst an Improved Genie2 with the same architecture, size and sampler but no ambient framework, the ambient training algorithm prevents the decline in designability and diversity seen on long proteins.

    At 800 residues, the number of designable clusters drops from 68% (Ambient) to 25% for the Improved Genie2 trained without the ambient framework. The two models perform similarly at 300 residues. · 4.2 Ablating the significance of Ambient…

  3. ShownOn short proteins (up to 256 residues), the model sets a new designability-diversity Pareto frontier over prior models including Proteína, with a much smaller network.

    5 structures per length from 50 to 256 residues (1,035 total), with the noise scale γ swept to trace the trade-off. Figure legend reports a 12.88× smaller model than Proteína. · 4.3 Comparisons on unconditional generation of…; Figure legends; Tables

  4. ShownThe model produces the least training-set-like designable backbones (lower TM-novelty is better) against both PDB and AFDB, using corrected FoldSeek v10 scoring that takes the maximum alnTM-score per query.

    Long (300 to 800 AAs, vs Proteína): 0.682 PDB and 0.740 AFDB, improvements of 18.4% and 16.2%. Short (≤256 AAs): beats Genie2 by 2.0% PDB and 1.6% AFDB, and Proteína by 7.0% and 4.8%. · 4.4 Novelty scores; Bug in novelty computation

  5. ShownWithout motif-specific optimization, the model matches the much larger, scaffolding-tuned Proteína on single-motif scaffolding and beats Genie2 on both single- and multi-motif tasks.

    Single-motif unique successful scaffolds at γ = 0.45: 1,923 vs Genie2 1,445 and Proteína 2,094 (1,000 samples per task, 24 tasks). Multi-motif: 89 unique across 5 of 6 problems vs Genie2 40 solving 4. · 4.5 Motif Scaffolding; A.1 Motif Scaffolding Results

Methods, data and reproducibility

AlphaFold structures are binned by average pLDDT (>90, 80 to 90, 70 to 80). Lower-quality ones are pre-noised once and used only at diffusion times where their distribution merges with clean data, with a rescaled loss and time-first sampling. Built on Genie2 with 8 triangle layers, trained on a geometrically reclustered AFDB set.

  • AFDB cluster representatives with average pLDDT > 70, reclustered with FoldSeek (TM-Align, TM threshold 0.5, coverage 0.75) into structurally diverse clusters · n = ~1.29M structures in; ~292K clusters out
  • Training set for the long-protein model (cluster representatives, three stages up to 712 residues; trained on proteins up to 768) · n = ~196,000, ~269,000, ~291,000 proteins across stages
  • Training set for the short-protein model (TM-Align threshold 0.4) · n = ~90K cluster representatives
  • Long-protein evaluation: sampled backbones at lengths 300, 400, 500, 600, 700 and 800 · n = 100 backbones per length
  • Short-protein evaluation: sampled backbones at lengths 50 to 256 · n = 1,035 structures (5 per length)
  • Novelty reference databases: PDB and the AFDB (588K) benchmark · n = AFDB (588K)
  • Motif scaffolding benchmark following Genie2 · n = 24 single-motif and 6 multi-motif tasks; 1,000 samples per task
Methods & tools
Genie2 (architecture and codebase), AlphaFold2 / AlphaFold Database, FoldSeek v10 (10-941cd33) and v9 (9-427df8a), MMseqs2, ProteinMPNN, ESMFold, Proteína (baseline), RFDiffusion (baseline), AdamW optimizer
Software & environment
Novelty evaluation uses patched FoldSeek v10 (release 10-941cd33); legacy comparisons use unpatched FoldSeek v9 (release 9-427df8a). Sampling follows Genie2 (1000 steps, γ = 0.6).
Assay conditions & compute
Training on 48 GH200 GPUs, taking 18, 48 and 48 hours for the three stages; Proteína is cited as 14 days on 128 A100 GPUs.
Reproducibility gaps
Code, models and datasets are promised at the listed repository as a future commitment, so availability could not be confirmed from the text. No separate download links or accession identifiers are given.
Limitations (6)
  • Quality is judged by the residue-averaged pLDDT, a coarse score that can itself be noisy or misleading; structures with pLDDT below 70 were discarded, so local corruption is not exploited. · stated by the authors
  • Only one representative per AFDB cluster is used, and the existing AFDB clustering is the starting point rather than reclustering the full 214M-structure dataset. · stated by the authors
  • The mapping from pLDDT to merging time was set by empirical tuning, without a systematic ablation or principled selection criterion. · stated by the authors
  • The generated structures have not been confirmed experimentally; evaluation is in silico (designability via ProteinMPNN and ESMFold). · stated by the authors
  • Short and long protein results come from two separately trained models with different training sets (TM threshold 0.4 vs 0.5), so a single model's generality across regimes is not shown. · noted in this summary
  • Gains over Proteína are quoted as percentages without stating whether they are relative or absolute, so the 91% and 26% figures at length 700 are hard to relate to the raw scores. · noted in this summary

Read this if you train generative protein structure models and want to use low-confidence AlphaFold structures, or need the corrected FoldSeek TM-novelty protocol.

Citations
4

Abstract

We present Ambient Protein Diffusion, a framework for training protein diffusion models that generates structures with unprecedented diversity and quality. State-of-the-art generative models are trained on computationally derived structures from AlphaFold2 (AF), as experimentally determined structures are relatively scarce. The resulting models are therefore limited by the quality of synthetic datasets. Since the accuracy of AF predictions degrades with increasing protein length and complexity, de novo generation of long, complex proteins remains challenging. Ambient Protein Diffusion overcomes this problem by treating low-confidence AF structures as corrupted data. Rather than simply filtering out low-quality AF structures, our method adjusts the diffusion objective for each structure based on its corruption level, allowing the model to learn from both high and low quality structures. Empirically, Ambient Protein Diffusion yields major improvements: on proteins with 700 residues, diversity increases from 45% to 86% from the previous state-of-the-art, and designability improves from 68% to 86%. We will make all of our code, models and datasets available under the following repository: https://github.com/jozhang97/ambient-proteins.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMahbub '25Lee '23Lisanza '23Wang '24Alamdari '23Villegas-Mo '25
76% semantic matchBioinformatics
Sazan Mahbub, Christoph Feinauer, Caleb N. Ellington+2 more·bioRxiv·Jun 30, 2025