OpenFold
A trainable, Apache-2.0 PyTorch reimplementation of AlphaFold2 that matched the original and opened it up.
About
OpenFold is a faithful PyTorch reimplementation of DeepMind's AlphaFold2 from Mohammed AlQuraishi's lab at Columbia, developed with the OpenFold Consortium, a 40-plus-member organization now hosted by the Open Molecular Software Foundation. Unlike DeepMind's public inference-only code, OpenFold is trainable: it supports training with DeepSpeed and mixed precision, and the consortium retrained the model from scratch on its OpenProteinSet data.
The Nature Methods 2024 paper reports that the from-scratch model matched AlphaFold2's accuracy with improved speed and memory efficiency, and used the retraining experiments to study what the network actually learns. Code is Apache 2.0, and model parameters plus OpenProteinSet are on the AWS Registry of Open Data under CC BY 4.0. OpenFold2 is also served as an NVIDIA NIM. Full training remains enormously expensive, but inference and fine-tuning are practical for labs.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (Ahdritz et al., Nature Methods 2024): retrained-from-scratch model matched AlphaFold2 accuracy on CAMEO and CASP15 validation sets with improved speed and memory efficiency
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: Nature Methods 2024 paper (DOI 10.1038/s41592-024-02272-z), repo github.com/aqlaboratory/openfold, and the openfold.io consortium site confirm the trainable reimplementation, the from-scratch parity results, the Apache 2.0 license, and the open dataset. Strengths: permissive licensing, genuinely trainable, open training data, active consortium backing. Docked for: (1) full training is out of reach for most groups; only inference and fine-tuning are practical; (2) inference still needs MSA and template database setup; (3) AF3-class models have since moved the field forward. License: Apache 2.0 (code); CC BY 4.0 (data and parameters).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Trainable PyTorch AF2 reimplementation; from-scratch model matched AF2 accuracy (Nature Methods 2024).
- · Full training is out of reach for most groups; inference and fine-tuning are the practical uses.
- · AF3-class models have since moved the field forward.