Model Book/OmegaFold

OmegaFold

B

HeliXon

Structure Predictionnon-commercial
Updated 2 months ago
NextIn take

OmegaFold's single-sequence approach is elegant and fast, but in practice, the accuracy gap versus MSA-based methods means you'd only reach for it when you literally can't build an alignment.

Specifications

ArchitectureSingle-sequence structure prediction — no MSA required
Parameters~100M
Training DataPDB structures, UniRef sequences
LicenseNon-commercial academic license. Commercial use requires HeliXon agreement.
Hardware1x GPU with 16GB+ VRAM
Inference CostSelf-hosted — fast inference, low compute cost
API AvailableNo
Weights AvailableYes

Benchmark Performance

BenchmarkScore
CAMEO0.72 lDDT (avg)

When to Use This

  • +Orphan proteins with no detectable homologs
  • +Quick structural screening where MSA construction is too slow

When NOT to Use This

  • Any target where MSA construction is feasible — Boltz-1 or AF3 will be more accurate
  • Complex prediction — single-sequence models don't handle multimers

Production Readiness

research

Known Users

  • Academic labs working on novel proteins

Grade Rationale

B

A useful single-sequence predictor that trades accuracy for speed and simplicity. Solid B — it fills a niche for orphan proteins without homologs, but MSA-based methods are strictly better when alignments are available.

Sources

Update History

2026-01-15Initial entry