BioEmu (Microsoft)

How SubSeq Runs BioEmu

• Guided mode is one sequence-to-ensemble workflow: precompute embeddings from A3M, sample from the warmed cache, optionally apply physical steering, and optionally reconstruct all-heavy-atom side chains in the same job.
• Advanced args use one BioEmu command per line, similar to GROMACS.
• Supported commands are python -m bioemu.subseq_precompute_embeds ..., python -m bioemu.sample ..., and python -m bioemu.sidechain_relax ....
• Arbitrary Python scripts, python -c, and unsupported bioemu.* modules are rejected.
• Shell operators, comments, quotes, and command chaining are rejected in advanced BioEmu command lines.
• Runtime dependencies, model assets, and caches are managed by SubSeq.
• Sampling paths are enforced: --output_dir=/outputs, --cache_embeds_dir=/outputs/.bioemu_embeds_cache, --cache_so3_dir=/ref/.bioemu_so3_cache.
• This deployment rejects direct sequence strings and FASTA input at submit-time; --sequence must point to an .a3m file.

Sampling

Generate backbone-frame equilibrium ensemble outputs from an A3M alignment:

python -m bioemu.subseq_precompute_embeds --sequence=/inputs/sequence.a3m --cache_embeds_dir=/outputs/.bioemu_embeds_cache

python -m bioemu.sample --sequence=/inputs/sequence.a3m --num_samples=100 --batch_size_100=10 --filter_samples=True --base_seed=42

• --sequence: path to an .a3m file.
• --num_samples: number of sampled structures to generate.
• Guided physical steering is exposed as a strength parameter, not as a separate mode.
• Guided atom detail chooses between native backbone/topology trajectory outputs and all-heavy-atom side-chain outputs.
• --denoiser_config and --steering_config must point to mounted config files under /inputs, /aux, or /ref.
• User-supplied --ckpt_path, --model_config_path, and --model_name are ignored by SubSeq.

Side-chain Reconstruction

Reconstruct all-heavy-atom side chains from BioEmu topology.pdb and samples.xtc outputs. Guided full-atom jobs run this after sampling using the files just written under /outputs:

python -m bioemu.sidechain_relax --pdb-path /outputs/topology.pdb --xtc-path /outputs/samples.xtc --outpath /outputs --prefix samples --no-md-equil

• --outpath is normalized to /outputs.
• Without --no-md-equil, BioEmu can run OpenMM local minimization/equilibration after side-chain reconstruction.
• Outputs include files such as samples_sidechain_rec.pdb, samples_sidechain_rec.xtc, samples_md_equil.pdb, and samples_md_equil.xtc.

This BioEmu module does not select representative centroid frames from the trajectory; use a trajectory postprocessor for clustering/extracting representative PDBs.

Outputs and Caches

• Sampling writes topology.pdb, samples.xtc, sequence.fasta, and batch_*.npz under /outputs.
• Embedding cache is written to /outputs/.bioemu_embeds_cache.
• SO(3) lookup cache is written to /ref/.bioemu_so3_cache.

Offline/Reproducible Use

• Prefer A3M input when you want to avoid remote MSA services.
• Side-chain reconstruction uses image-bundled OpenMM and HPacker so jobs do not install dependencies at runtime.