ColabFold 1.6.1

• Fix compatibility with numpy 2.4 by pulling in fix from AF2 upstream

v1.6.0

ColabFold 1.6.0

New features and enhancements

• AlphaFold3 output integration New --af3-json flag in colabfold_search and colabfold_batch to generate AlphaFold3-compatible input JSON files from MSA search results (edc9758, 554c268, bc64530, c417d83, 3456682)
• New MSA pairing pipeline Improved complex mode with --pair-mode option (unpaired, paired, unpaired_paired) and faster pairing search (e8d94b2, 07644a8, 26c0d46, 18081af, a77f28d, cffc8d0)
• Skip output feature New --skip-output flag allows selective skipping of MSA, plots, and PAE JSON output to reduce disk usage and runtime (5d9710b)
• Lighter colabfold_search colabfold_search now works with few heavy dependencies installed and has reduced startup overhead (dfbe59d, 62b5286, 77a8204, f0bf465)
• Expanded CLI parameters Expose --max-template-hits and --max-template-date to control template control cut-offs (ea3292b)
• Debug logging mode New --debug-logging flag for verbose output (e462021)
• Initial structure guess New --initial-guess parameter for warm-starting predictions from known structures (1403b76, 27b6e3e, 404772e)
• Environmental pairing Experimental environmental sequence pairing feature for improved MSA generation (07644a8, 57b220e). Currently not supported on the public MSA server
• Per-entry custom templates in CSV Support for templatepath column in CSV input for sequence-specific templates (972f6a4)
• MMseqs2-GPU support GPU-accelerated local MSA search via --gpu 1 in setup_databases.sh and colabfold_search. GPU databases are much smaller since they don't require the k-mer prefilter datastructures. Requires MMseqs2 release 16 or later (88e5efa, 64c8fd6, 379d101, 0cf2621, 2bec5c7)
• Ungapped prefilter mode New --prefilter-mode 1 for CPU-based exhaustive ungapped alignment prefilter, providing higher sensitivity at the cost of runtime (be06b20)
• Docker improvements New and improved Docker container (8bce4c4)
• Faster MSA read-in Faster unserialize_msa (093bb55)
• AWS S3 database download Databases are now downloaded from AWS S3 by default instead of FTP (4ed8bc8)
• PDB100 database Switched from PDB70 to PDB100 for template search (7dd7631)
• Added optional orjson support for faster JSON serialization
• actifpTM Added support for actifpTM score (bc30247) Thanks @gezmi
• aarch64 support ColabFold is now pip-installable on ARM64/aarch64 Linux (1ccca5a)
• OpenMM and pdbfixer dependencies installable directly via pip (905fdfd, ec82aa5)
• MMSEQS_IGNORE_INDEX environment variable to force colabfold_search to skip existing prebuilt index (09c251b)

Bug fixes

• Fix pandas dtype guessing breaking on invalid job names in CSV input (fdf3b23) Thanks @staszekdh
• Fix endless retry loops in database download (a134f6a) Thanks @heya5
• Fix unpacking of query sequences (ecc7d10, 747aa90) Thanks @k-ujihara
• Fix template database for multimer predictions (55ac663)
• Fix colabfold_search now always treating input as protein sequences (c5b9621)

v1.5.2

v1.5.2

• bugfix - same random seed was used between recycle, resulting in identical dropouts (if --use-dropouts was enabled).
• various modifications to reduce GPU RAM used and minimize memory leaks between recycles/models/inputs.

See change log

v1.5.1

v1.5.0

• See change log

v1.5.1

• bugfix --save-recycles/--save-all option was broken

v1.4.0

pin tensorflow version

v1.3.0

Update README.md

v1.2.0

v1.2.0

ColabFold

Making Protein folding accessible to all via Google Colab!