AVAIL-omics

Availability-aware multimodal integration for clinical cohorts with missing modalities.

Real multimodal cohorts are incomplete: an fMRI is missing here, a shotgun metagenome there. Complete-case analysis discards subjects; AVAIL-omics drops nobody. Every modality trains its own base learner on all available subjects, a fixed-length meta-feature vector (base probabilities + availability flags) is assembled per subject, and a meta-learner makes the joint decision.

Install

pip install avail-omics           # core: CPU-only, no CUDA needed
pip install "avail-omics[deep]"   # + neural (MLP/MoE/GNN/Transformer) and
                                  #   deep-tabular (TabNet/TabPFN/TabICL) meta-learners
pip install "avail-omics[full]"   # + SHAP-based biomarker interpretation

Quick start (any tabular multimodal dataset)

from avail_omics.generic import run_generic_pipeline

res = run_generic_pipeline(
    data_root="mydata",      # metadata.tsv (SampleID, Group) + features/*.tsv
    out_dir="results",       # one .tsv per modality -> 11 meta-learners, LOO
)

Data layout (one file per modality, rows = samples):

mydata/
  metadata.tsv          # SampleID + Group (+ optional Age, Education, Sex)
  features/
    brain.tsv
    stool16S.tsv
    ...

Or use the CLI:

avail-omics run --data-root mydata --out results
avail-omics robustness --data-root mydata --out robustness \
    --permutations 25 --jobs 4          # full-pipeline label-permutation test
avail-omics sensitivity --data-root mydata --out sensitivity

What is included

• Leak-free protocol: preprocessing, feature selection, imputation and hyperparameters are computed strictly inside each cross-validation fold.
• 16 missing-data strategies + hybrid: complete-case, zero/mean/median/KNN/ MICE/MissForest/SoftImpute imputation, availability-flag encoding, and the hybrid (fold-internal imputation + true availability flags).
• 11 meta-learners: SimpleMean, WSimpleMean, ElasticNetLR, MLP, MoE, GNN, Transformer, XGBoost, TabNet, TabPFN, TabICL.
• Robustness toolkit: synthetic dropout, MCAR/MAR/MNAR simulation, label-permutation test, modality sensitivity/ablation, bootstrap CIs, conformal prediction sets, calibration (isotonic) with Brier reporting, integrity audit.

Evidence summary (primary cohort, n=150 oral-gut-brain)

• HC vs PDD binary AUC 0.978 (MLP, leave-one-out)
• Severity continuum: Spearman ρ = 0.633 vs clinical stage; risk score tracks MMSE (ρ = -0.47) and CDRS (ρ = 0.63) — scores never used in training
• Label-permutation test of the full pipeline: permuted macro-AUC ≈ 0.50 (p = 0.039), ruling out leakage as the source of the separation
• Sustains prediction where complete-case collapses (25-55% dropout)
• External validation on public cohorts: Wallen 2022 PD shotgun (AUC 0.642), HMP2 IBD multi-omics with genuine per-sample missingness (macro-AUC 0.771), preclinical AD (0.542), label-free reconstruction (Vascellari 2020)

Web platform

A full pipeline web app (upload → auto-run → report) accompanies this package; see the project repository.

Citing

Manuscript in preparation. Until the paper is out, please cite the repository and package: avail-omics v1.0.0.

License

MIT