Boltz v0.3.2: Memory improvements

Bug Fixes

• Fixed max MSA length, which was causing some OOM for some users
• Changed default chunking threshold

Boltz v0.3.1: Few bug fixes

Bug Fixes

• Fixed an issue with automated MSA's when running boltz on multiple inputs
• Improved handling for smiles that fail conformer generation

Boltz v0.3.0: Confidence model, low memory mode & custom MSA pairing

New features

• We've added our confidence model, which can be used to rank multiple samples (as controlled by the --diffusion_samples flag). It dumps a json file with aggregate metrics as well as full PAE / PDE matrices on demand using --write_full_pae and --write_full_pde.
• We've added a chunking feature to lower memory requirements, it kicks in automatically after 512 tokens for little to no slowdown
• We now have a custom MSA pairing format using a CSV with sequence and key columns
• MSA server parameters are now surfaced to the user (thanks @YoelShoshan)

Bug fixes

• Fixed an issue with shallow MSA's on multimers when using the msa server
• Improved error messaging in a few places

Boltz v0.2.1: MSA generation now opt-in

Bolts v0.2.1 Release

Changes

We've made MSA generation an opt-in feature that is set using the --use_msa_server flag. This is important for privacy reasons.

Bug fixed

Fixed an issue with MSA's when running batched inputs.

Boltz v0.2.0: Automatic MSA generation & various improvements

New Features

• The boltz predict command will now automatically generate an MSA using the mmseqs2 server! Shoutout to the colabfold team for providing the server to the community. This works for both single and multi-chain complexes. This happens by default when no msa is passed in the input yaml / fasta.
• We have added an explicit single sequence mode that can be activated by passing in the special keyword empty. We generally advise against doing so as it will hurt performance.
• We've improved error messaging across the board to improve user experience

Bug fixes

• Fixed a failure on modified residues
• Fixed an indexing error for covalent bonds
• Fixed a gif writer bug when multiple unique smiles were provided in the input
• Fixed a versioning issue with RDkit for some python versions

Miscellaneous

• We've added many new example configs
• The model weights are not stored on huggingface hub. Thank you @Vaibhavs10

Boltz-1: First release

v0.1.0

v0.1.0