AlphaFold v3.0.2

Version 3.0.2 provides new features, improves performance, improves documentation, and fixes bugs. AlphaFold v3.0.2 supports NVIDIA Blackwell GPUs out of the box. We thank everyone who reported issues, proposed new features, and submitted PRs!

Note that the AlphaFold 3 parameter files are compatible with any 3.0.x version. We bump the the major/minor version numbers only when a new model is released, otherwise only the patch version number is increased.

Change log

Below is a selection of new features, performance improvements and bug fixes. Full change log: v3.0.1...v3.0.2.

New features

• Better bond validation: residue upper bound, and validity of atom names within a residue are now checked.
• Better error messages if paths supplied instead of data in templates, MSA, or CCD.
• Allow supplying user-provided CCD using an external file.
• Added flag --force_output_dir to force writing into an output directory even if it exists.
• Added chirality checking utility.
• All output filenames are now unique, so flattening a directory of outputs is no longer an issue.
• Support for optionally saving the distogram using the --save_distogram flag. Thanks @PasqM.
• Added a flag to compress large output files (mmCIF and confidence JSON) using zstandard. Thanks @ntnn19.

Performance improvements

• Jackhmmer requires significantly less RAM when it finds excessive number of hits for a particular sequence.
• Support for searching against sharded databases in Jackhmmer and Nhmmer, enabling 10-30x MSA search when used. See https://github.com/google-deepmind/alphafold3/blob/main/docs/performance.md#sharded-genetic-databases for more details.
• Use Tokamax for the implementation of flash attention and SwiGLU. This also adds support for newer GPUs (e.g. NVIDIA Blackwell) and provides an option for further inference performance optimisations in the future. Big thanks to @rpachauri, @chr1sj0nes, @sbodenstein and other people working on Tokamax!
• Made saved embeddings 2x smaller by using float16 instead of float32.
• Significantly reduced the size of the Docker and Singularity containers by disabling pip cache.
• The Docker container build is faster thanks to using uv.
• Docker build should be slightly faster on machines with > 8 cores.

Bug fixes

• The output written in stdout is now cleaner and more consistent.
• Fixed token chain IDs in summary metrics to be the same as the input chain IDs.
• Fixed user CCD validation to check all provided entries.
• Fixed a crash in cases where Hmmsearch found no template hits (hapenning especially with short peptides).
• Added support for useStructureTemplate and maxTemplateDate in the AlphaFold Server JSON format.
• Embeddings saved as part of the output are now correctly cropped instead of including padding up to the bucket size.
• User-provided unpaired MSA is no longer deduplicated and thus preserves the exact provided pairing. Thanks @MGordon09.
• Make pseudo_beta_fn work with both np and jnp. Thanks @popfido.
• All CCD entries should now parse without errors by correctly wiring through the max reference date. Thanks @jvermaas.
• Atom names are now validated in the bonds input list. Thanks @maxh190.
• Removed dependency on the chex library.
• Output directory names now respect the casing in the folding job name. Thanks @adavasam.
• Fixed handling of CIF tokens that contain quotes surrounded by spaces in CifDict. Thanks @FilipSchymik.
• Reduced RAM use when there were multiple inputs provided with custom user CCD. Thanks @MichaelaBrezinova, @genius-0963.
• The pickle loader is now safe against maliciously crafted CCD pickles. Thanks @sanowl.
• Upgrade GitHub Actions to Node 24 compatible versions. Thanks @salmanmkc.
• The location of libcifpp (where AlphaFold 3 looks for the CCD) can be specified via the LIBCIFPP_DATA_DIR environment variable. Thanks @cgross95, @Ashutosh0x.
• Made sharded Jackhmmer result merging more robust in case of e-value ties

Documentation

• Documented the names, shapes, and dtypes of all AlphaFold 3 model parameters.
• The discrepancy in genetic search in AlphaFold 3 and AlphaFold Server is now documented in known issues. Thanks @stianale.
• Updated Docker build instructions for AlmaLinux/Rocky/RHEL users. Thanks @ocstx.

We thank the following people for their pull requests and/or great suggestions: @armanvnv, @Ashutosh0x, @ccoulombe, @genius-0963, @ishanrajsingh, @lucajovine, @MichaelaBrezinova, @noghte@ntnn19, @ojasshelke, @owainkenwayucl, @PasqM, @pavithra-rvce, @popfido, @salmanmkc, @sanowl, @zohaib-7035.

AlphaFold v3.0.1

Version 3.0.1 provides new features, improves performance, improves documentation, and fixes bugs. We thank everyone who reported issues, proposed new features, and submitted PRs!

Change log

Below is a selection of new features, performance improvements and bug fixes. Full change log: v3.0.0...v3.0.1.

New features

• It is now possible to run template-free, but still search for MSA.
• It is now possible to run template search when MSA provided.
• It is now possible to specify MSA and templates as external files instead of inline in the input JSON.
• Added these new flags: --max_template_date, --diffusion_num_samples, --num_recycles, --num_seeds.
• Exposed --conformer_max_iterations as a flag to allow increasing number of RDKit iterations.
• Added an option to return model embeddings and write them to a file.
• Made it possible to set which GPU to use on a multi-GPU system.
• The per-residue pLDDT is now added in the output mmCIF.

Performance improvements

• Replaced the database download script with a significantly faster one and added GCP post-processing scripts.
• Improved the Stockholm converter to not read more sequences than necessary when converting to A3M.
• Improved the Stockholm to A3M conversion to not read the whole file into RAM.
• Sped up template search by not reading and parsing irrelevant mmCIFs. Very significant if PDB stored on a slow filesystem.
• Refactored input JSON parsing as iterator, so that inputs are not all loaded into memory at once and cause an OOM.
• Added sequence deduplication when outputting the JSON, making it smaller for homomers.

Bug fixes

• The data pipeline is now skipped if blank MSA + templates provided.
• Silenced irrelevant Numpy warnings.
• Improved Singularity installation instructions.
• Fixed Pallas GLU kernel on large inputs and updated documentation for compilation buckets.
• Fixed setup to work with Python 3.12.
• If the output directory already exists, AlphaFold won't overwrite files in it, instead it will create a new one.
• Convert DATIVE bond type to SINGLE to match bond types used during training.
• Convert atom element to uppercase in embedding name to match training.
• Fixed user and permissions when untar-ing the PDB mmCIF files.
• Explicitly use conformer_id when getting generated conformers.
• Fixed A3M -> Stockholm conversion when description contains tabs.
• Sort the input files returned by glob - otherwise in arbitrary order.
• Ensure user CCD contains all necessary fields and improve docs.
• Fixed incorrect handling of two-letter atoms in SMILES ligands.
• Use OpenEye canonical SMILES when available.
• Raise clear errors if cannot create molecule definitions.
• Add XLA flag workaround for CUDA Capability 7.x GPUs.
• Small fix for numerical precision when handling PIDs in Pallas kernel.
• Deal properly with non-standard residues in the single-letter sequence.
• Fixed RASA calculation to work with arbitrary chain IDs.
• Use true chain type for paired MSA creation.

We thank the following people for their pull requests and/or great suggestions:
@alchemistcai, @amorehead, @aozalevsky, @ccoulombe, @dailypartita, @DanFosing, @davidecarlson, @eltociear, @hegelab @jkosinski, @jpcartailler, @jpcartailler, @linuxfold, @lucajovine, @noghte, @orbsmiv, @Saharsha-N, @smg3d, @vatese, @YoshitakaMo.

AlphaFold v3.0.0

Initial release of AlphaFold 3.

PiperOrigin-RevId: 695257954