MMseqs2 Release 18 at a glance: new Forward–Backward aligner, re-enabled substitution matrix parameter estimation, faster ARM64 alignments, improved GPU support.

Breaking changes

• CMake ≥ 3.15 and <4 is required to compile MMseqs2 (da0b2c3).
• gpuserver no longer accepts the --gpu parameter (3b5d13e).
• databases generated by the databases module are now GPU compatible. This might slightly alter search/clustering results since the sequence order was changed.

New features and enhancements

• New Forward‑Backward (FWBW) aligner fwbw (003fabc, fb687b7) by @Gyuuul2 @elpis51613 @lasseReifenrath
• Custom substitution matrices are supported again through a new lambda calculator (5ebd6e9, efad625, b76ebc4) by @edawson.
• Proximity‑aware pairing: pairaln can now match sequences that are physically close in accession space (--pairing-filter 1 and --pairing-prox-dist e019185, c9107ba, 835acb9, 1970db6, f6e9636, 60a894b).
• createdb and databases accept --gpu parameter to directly produce GPU databases (0578939, 90ee542). Sequence databases generated by the databases module use the flag default and are GPU compatible.
• Speedup aarch64 SIMD alignment with new/improved simd_any, simd_eq_all, simd_hmax* instructions (103fe79). Thanks @nskyav
• aarch64 GPU binaries mmseqs-linux-gpu-arm64.tar.gz are now built with Clang 20 and wide Neon registers for additional speedup (62cf4d0, 9564601, 1668032). Thanks to @nskyav. mmseqs-linux-arm64.tar.gzis still the compiled with the older slower configuration
• createdmptaxonomy allows converting taxonomy databases back to .dmp files (bc0f9cb).
• taxonomyreport can now emit one database per query with --report-mode 3 (8284a8b, 033d5f5).
• Reduce thread start overhead in expandaln, pairaln, subtractdbs and unpackdb (18d8ddc).

Bug fixes

• Improve error handling for createindex with only ungapped prefilter (9668e96, 829003a).
• Fix precomputed index being slightly too large (b98f207).
• expandaln now skips entries lacking alignments (9c13275) and finds correct representatives (#691, 8783404).
• clusterupdate preserves members correctly (#961, e7f5852, 296d912).
• MPI nucleotide clustering was crashing (defe1af)
• SAM start coordinate was wrong sometimes (c13eef0)
• Systems with old mawk could result in corrupted databases due to large‑int printing (eaecacf)
• Wrong createdb mode message (#955, 48143e7)
• Fix createdb could crash if FD 0 was closed (99a025e). Thanks @jnooree

Developer notes

• GPU binaries are now compiled with prebuilt GPU containers available at https://github.com/steineggerlab/build-containers (and other CI improvements; 12a0d9a, 887b441, 4289ae3, 0eb08bf).
• GPU regression tests added to CI (bebc956, 4289ae3).