google-deepmind/alphafold v2.3.0

AlphaFold v2.3.0

on GitHub

Version v2.3.0 updates the AlphaFold-Multimer model parameters. These new models are expected to be more accurate on large protein complexes but use the same model architecture and training methodology as our previously released AlphaFold-Multimer paper. See the v2.3.0 release note for more details.

Thanks to various memory optimisations, AlphaFold-Multimer now uses less GPU memory and it can therefore handle longer proteins.

A number of other bug fixes and small improvements have been made.

Change log

• Added new AlphaFold-Multimer models with better accuracy on large protein complexes.
• Added early stopping to recycling.
• Added filtering for non-protein sequences in the pdb_seqres download script to prevent template search errors.
• Fixed a bug where histidine residues had sometimes swapped atom coordinates after relaxation (thanks @avwillems).
• Updated MGnify to 2022_05, UniRef90 to 2022_01, UniClust30 to 2021_03, UniProt in Colab notebook to 2021_04.
• Used bf16 in multimer inference – reduces GPU memory usage.
• Upcast to fp32 when using bf16 in LayerNorm and replace hk.LayerNorm with common_modules.LayerNorm.
• Updated Jax to 0.3.25 and Haiku to 0.0.9 for consistency with the AlphaFold Colab notebook.
• Changed TriangleMultiplication to use fused projections and various other memory optimisations.
• Upgraded Python version in the AlphaFold Colab notebook to 3.8.
• AlphaFold Colab notebook usability improvements – multimers with up to 20 chains are now supported, higher sequence length limits, number of recycling iterations can now be controlled, and added an option to run single chains on the multimer model.
• Relaxation metrics are now saved in relax_metrics.json.
• Some Jax deprecation errors were addressed (thanks @jinmingyi1998).
• Various documentation and code improvements (thanks @mathe42).