Run supported folding models through BioIR's public build_processor
pipeline. The pipeline handles parsing, tokenization, feature generation,
inference, and PDB/CIF writing without an OSS model data module.
The default command runs Boltz-2 on the bundled T1031 sample and prints the structure to stdout:
python examples/folding/run_demo.py > T1031.cif--output-dir writes instead: one <record>.cif (or .pdb) and one
<record>_scores.json per prediction. Without it the structure goes to stdout
and the record id and scores go to stderr, so the redirect above yields a file
with nothing else in it.
Select another model or sample with command-line options:
python examples/folding/run_demo.py \
--model-source openfold3 \
--input examples/data/samples/monomers/T1031.json \
--output-dir output \
--sampling-steps 50--model-source accepts the models the pipeline registry has a factory for
(bionemo_ir/registry.py) — the Boltz, OpenFold and AlphaFold2 variants, named
by their support-matrix keys. --help enumerates those values.
bionemo_ir/hubs/support_matrix.py is a wider list: it names every model the
library can load weights for, including ones with no pipeline factory or with
pipeline stages that still raise NotImplementedError. Those have to be driven
from Python.
The diffusion options (--recycling-steps, --sampling-steps,
--diffusion-samples) apply only to the diffusion models; the rest ignore them.
See python examples/folding/run_demo.py --help for all options.