ai4science-healthcare

name	ai4science-healthcare
description	Applies to healthcare and life sciences ML recipes under healthcare/models/ in AI4Science Studio, with privacy and non-clinical-use guardrails.

Healthcare & Life Sciences (HCLS) domain

Scope

The healthcare/ domain holds recipes for healthcare and life sciences open models (imaging, clinical NLP, genomics helpers, drug discovery, molecular design, etc.) on Hugging Face. All content is for research and engineering support, not patient care decisions.

Layout

• Models: healthcare/models/<model-slug>/
• Slug rules: healthcare/models/README.md
• Structural template reference: _template/ (repo root)

Mandatory guardrails

• No PHI: Do not add patient-identifiable information, hospital identifiers, free-text notes from real charts, or internal EHR exports to the repository.
• Not medical advice: Documentation and scripts must not present outputs as diagnosis, treatment, or clinical guidance.
• Intended use: Copy or summarize the model card's intended use and limitations into the model README.md.

Agent guidance

• Prefer public benchmarks (e.g. de-identified challenge datasets) in examples.
• When discussing evaluation, distinguish offline metrics from regulatory or deployment readiness.
• If a user asks for real-patient pipelines, redirect to institutional compliance (IRB, BAAs, local policy) outside this repo's scope.

AMD/HPC patterns for healthcare models

GP-MoLFormer

• Apptainer SLURM path: sbatch_inference_amd.sh clones IBM/gp-molformer inside the container on first run, installs deps via pip install --target /opt/gpmol-pkgs, then calls run_generation.sh. Internet access from compute nodes required for first run. Subsequent runs reuse the clone from GPMOL_WORK_DIR.
• Upstream uses environment.yml (conda), not requirements.txt. The sbatch script translates conda deps to pip equivalents. Key mappings: pytorch is already in SIF (skip), rdkit becomes rdkit-pypi, pytorch-cuda is skipped (ROCm).
• No overlay needed: deps are lightweight; runtime install (~1 min) is acceptable. Torch is stripped from --target dir after install to keep the SIF's ROCm torch.
• fast_transformers unavailable on ROCm: the code falls back to standard PyTorch transformers automatically; no user action needed.
• Key env vars: GPMOL_SIF (required), GPMOL_WORK_DIR (default: examples dir), SCAFFOLD (empty = unconditional), NUM_BATCHES (default 1 = 1000 molecules), OUTPUT_FILE.
• ROCm version compat: The ROCm 7.0 SIF (py3.10) fails silently on hosts with newer amdgpu drivers (e.g. kernel module 6.16.6 on MI300A clusters) — torch.cuda.is_available() returns False and generation falls back to CPU (~10-20x slower). Use the ROCm 7.2.2 SIF (py3.12) on such clusters, or set HSA_OVERRIDE_GFX_VERSION=9.4.2 as a workaround. The sbatch script includes a GPU detection check that warns when this happens.
• rdkit-pypi has no py3.12 wheel: When using the ROCm 7.2.2 SIF (py3.12), the RDKit validity check is gracefully skipped. Molecules are still generated correctly.
• Validated results:
  • MI300X: 996/1000 valid (99.6%), ~41s wall time (ROCm 7.0 SIF)
  • MI300A: 1000 molecules generated, ~18s wall time (ROCm 7.2.2 SIF)

General healthcare model pattern

• Use sbatch_<task>_amd.sh naming (not _mi300x.sh); same script covers MI250X, MI300X, MI350X.
• Always include the research/engineering-only disclaimer both in the recipe README and in the sbatch script header comment.