[3dem] CryFold: automated protein model building

[A. Amunts]

Dear colleagues,
Forwarding this message from Jianyi Yang, the developer of CryFold.
Alexey

— 

We are pleased to share that the full version of CryFold, the automated protein model building tool is available.
The open source code and model weights can be accessed at: https://urldefense.com/v3/__https://github.com/SBQ-1999/CryFold__;!!Mih3wA!ENiW_V8YqUPCC10WeZUq0JRUMG1ZlLLRaLOm2coLVkgPH3gMcWy7rma4FGKfk_kj87LN0XlVWG0Vu1NHJ4PcPRVuGQ$ 

Key Features:

- CryFold builds on the state-of-the-art method ModelAngelo by integrating some aspects of AF2 network.

- Innovations include the implementation of a local attention mechanism and 3D rotary position embedding (3D-RoPE).

- Benchmark tests demonstrate more complete protein models, reduced resolution requirements, and faster modeling processes.

- CryFold can generate models using only density map information (no sequence), inferring the most probable sequences. With an optional database, CryFold identifies homologous sequences for further refinement. The iterative process is particularly useful for unknown sequence discovery, as outlined in the preprint: https://urldefense.com/v3/__https://www.biorxiv.org/content/10.1101/2024.11.13.623164v3__;!!Mih3wA!ENiW_V8YqUPCC10WeZUq0JRUMG1ZlLLRaLOm2coLVkgPH3gMcWy7rma4FGKfk_kj87LN0XlVWG0Vu1NHJ4MPN4268g$ 

- Modelling typically takes minutes to hours on a single A800 GPU, depending on the map's size.

We hope CryFold will be a useful addition to your toolkit.

Baoquan Su, Kun Huang, Zhenling Peng, Alexey Amunts, Jianyi Yang.