biopipelines

name

biopipelines

description

Author and run BioPipelines (locbp-uzh, CSBJ 2026) computational protein & ligand design pipelines on a GPU. BioPipelines gives ~76 comp-bio tools one common Pipeline/Resources API: structure prediction and docking (AlphaFold, Boltz2, ESMFold2, DiffDock, GNINA, NeuralPLexer, DynamicBind), de novo generation (RFdiffusion 1/2/3, RFdiffusionAllAtom, BoltzGen, PocketGen, HBDesigner), inverse folding / sequence design (ProteinMPNN, LigandMPNN, LASErMPNN, Frame2Seq, ThermoMPNN), MD (OpenMM), pocket detection (FPocket, P2Rank, AF2BIND), cheminformatics (RDKit, OpenBabel, ADMET-AI), and interaction/stability analysis (PLIP, ProLIF, PoseBusters, Prodigy). Use for protein binder/enzyme design, inverse folding, ligand docking, compound-library (incl. covalent) screening, and codon optimization. Load when the user wants a BioPipelines workflow, names one of these tools, or wants to run such a design/screening pipeline on GPU.

BioPipelines

BioPipelines (Quargnali & Rivera-Fuentes, LOC-BP UZH; CSBJ 10.34133/csbj.0129) is a Python framework that puts ~76 protein/ligand-modeling tools behind one declarative Pipeline API. You describe a workflow as a chain of tools; the framework generates and runs the per-tool scripts, tracks IDs through typed data streams, and materializes outputs.

This skill ships inside the repo so that adding it to any agent is one step (import the skill from this repository) and it tracks the framework as it evolves. It covers using the framework and running it on any single-node GPU backend (a managed-container provider like Modal/RunPod, a plain Docker GPU box, or an interactive Slurm GPU shell) via the generic container config variant.

Absorb, don't recite

The llm/ folder in this repo is the authoritative, author-maintained agent contract. Read those first; they override anything paraphrased here. Do not duplicate their content into your own notes — point at them:

• llm/pipelines.md — how to author a Pipeline, the tool/data-stream model, ID tracking.
• llm/development.md — conventions, gotchas, how tools are structured.
• llm/cluster.md, llm/colab.md — the two backends the repo ships natively.
• docs/tool_index.md, docs/tool_reference.md — the full tool catalog and per-tool signatures.
• references/container_backend.md (in this skill) — the generic single-node GPU backend.

The API in one screen

from biopipelines import Pipeline, Resources, Sequence, Ligand, Boltz2

with Pipeline("Project", "job_name", description="..."):
    Resources(gpu="A100", memory="64GB", time="6:00:00", cpus=8)
    prot = Sequence("MSEQ...", type="protein")           # or Sequence.from_uniprot("Q15436")
    lig  = Ligand(smiles="C[N+]1=C(...)...")
    Boltz2(prot, lig, affinity=True, output_format="mmcif")

Resources(...) is set once per pipeline. Tool.install() is called inside the with Pipeline(...) block (a bare Boltz2.install() at module scope is a silent no-op) — or just use bp-warm (below), which wraps it for you.

Running on any single-node GPU host (the container backend)

The repo ships a generic single-node variant so you do not hand-author a config per provider. Set three env vars and point one config line at your persistent mount:

export BIOPIPELINES_CONFIG_VARIANT=container   # select config.container.yaml
export BIOPIPELINES_OTF=1                       # run tools inline (no scheduler)
export BIOPIPELINES_LOCAL_OUTPUT=0              # honor configured output dir, NOT cwd

In config.container.yaml edit only folders.base.root: to your persistent mount (default /workspace); every other path — env root, weight caches, outputs — derives from it. First, warm the tools you need onto that mount once:

bp-warm Boltz2 ProteinMPNN        # builds per-tool micromamba envs + downloads weights
python my_pipeline.py             # subsequent runs reuse the warm env + cached weights

Dockerfile.container in the repo root builds a ready image (CUDA 12.4 + micromamba + pip install -e '.[colab]', with build-essential/gcc present — several tools JIT-compile a CUDA/C helper at import and fail without it).

Why BIOPIPELINES_LOCAL_OUTPUT=0 matters: with OTF on a non-Colab scheduler the framework otherwise diverts output to the ephemeral ./outputs (cwd), silently overriding your configured biopipelines_output. On a container that directory is lost at teardown. Setting it to 0 routes results to the persistent mount.

Reporting back

Save the structure (.cif/.pdb), the confidence/affinity JSON, and a summary figure as artifacts. For a co-fold, report pTM, ipTM/ligand-ipTM, complex pLDDT, and (if affinity=True) affinity_probability_binary. Note that Boltz2 has no covalent-mechanism knowledge: an unconstrained co-fold of a covalent ligand finds a non-covalent pocket, not the reactive residue — use Boltz2's covalent_linkage constraint when the mechanism is covalent.