Optional analyses¶
Frameshift screening, TMVec2, Boltz, and InterProScan are opt-in. Each feature needs a separate runtime or data set.
| Feature | Install or data requirement | Run option |
|---|---|---|
| Frameshift screening | Pinned BATH and Easel builds | --frameshift-screening |
| TMVec2 | Structural Pixi environment and the pinned Lobster-24M/BFVD resource set | --tmvec. Add --tmvec-gpu to require CUDA. |
| Boltz and Foldseek | Boltz runtime and a Foldseek database prefix | --boltz --no-skip-structural |
| InterProScan | An external InterProScan directory with interproscan.sh |
--interproscan |
Frameshift screening¶
Build the pinned BATH tools:
Test them with the shipped three-contig Trichomonas vaginalis G3 input:
The row must have status=success, predictions=5, and accepted=2. The
frameshift screening guide describes the rescue
files and validation steps.
TMVec2¶
Install the structural environment:
Install the pinned TMVec2 models and matching BFVD embeddings:
pixi run virosync orchestrate setup \
--config config/orchestration.yaml \
--no-write-config \
--tmvec
The default target is resources/virosync-optional/tmvec. Setup downloads the
pinned BFVD bundle and the pinned Lobster-24M and TMVec2 model files. It checks
each SHA-256 before it activates the target.
Use --tmvec-dir PATH to select another target. The command-line path has
priority over phase3.tmvec_database_dir in the config. A custom
--tmvec-url also needs --tmvec-resource-sha256. An explicit TMVec2 setup
request exits with status 1 if a download or validation step fails.
ViroSync uses Lobster-24M to create 408-column residue features. TMVec2 then creates one 512-column protein vector. The installed manifest binds those model revisions to the BFVD embeddings. BFVD supplies the structural evidence used by the confidence score.
The setup command downloads the Lobster-24M weights from
asalam91/lobster_24M and the
TMVec2 weights from
scikit-bio/TMVec-2. The
Lobster code uses the Apache-2.0
license. The TMVec2 model repository uses the BSD-3-Clause license. The
Lobster-24M model repository does not state a model-weight license.
Run the model-parity and BFVD query check:
The check selects the CPU or CUDA reference vector from compute.device. It
then runs the recorded sequence through the production BFVD search path and
requires the recorded target and score.
Run ViroSync with TMVec2:
pixi run -e structural virosync \
-i genome.fna \
-o results/tmvec \
--config config/orchestration.yaml \
--device cuda \
--tmvec \
--tmvec-gpu
Plain --tmvec runs on the configured CPU or CUDA device. --tmvec-gpu
selects CUDA and requires it. An enabled TMVec2 run stops before analysis when
its runtime or resources do not pass validation.
Boltz and Foldseek¶
Install the isolated Boltz runtime:
Set phase3.viral_structure_db to a Foldseek database prefix and set
phase3.boltz_use_msa_server: true. ViroSync writes sequence-only Boltz inputs
and does not write local multiple-sequence alignments.
Check the runtime and database:
Enable the layer with --boltz --no-skip-structural. The layer needs the
runtime, database, and MSA-server setting. ViroSync disables it when one of
these inputs is absent.
InterProScan¶
Point phase3.interproscan_dir at an InterProScan installation that contains
an executable interproscan.sh, then check it:
For an InterProScan archive, pass both --interproscan-url and
--interproscan-resource-sha256. Setup checks the archive digest and preserves
the executable bit on interproscan.sh.
Enable annotation with --interproscan. ViroSync disables the layer when the
configured program is absent.
Check all optional layers¶
This check does not download models or databases. It exits with status 1 and lists each missing requirement when the optional setup is incomplete.