AVAIL Research Platform

Research software · multimodal integration

Availability-aware multimodal analytics
for incomplete clinical cohorts

Keep every subject, model every available modality, and report missingness instead of hiding it. This is the research platform for the open-source avail-omics package.

Research use onlyNot a medical deviceNo account requiredNo public uploadsOpen source

Private by designPublic site accepts no data; demo is illustrative.

No public computeFull analysis runs locally, in Docker, Colab or your GPU worker.

Inspectable pipelineOpen-source package and validation notes; no hidden model.

Runs where you decideThe Python package and Docker image are the compute layer.

Public uploads are disabled. This site is a static, privacy-first showcase. The figure and report below are from a reference run; they are not generated from your data.

Reference 11-model benchmark

Aggregate output from a reference AVAIL run. Research use only; not a clinical result.

Example AI advisor report

Download the illustrative 11-model advisor report as PDF.

Open PDF report

What the full pipeline returns

Meta-learner table, availability/coverage figures, per-class AUC, LOO predictions, and a reproducible results ZIP. The public site does not execute your data.

Run it yourself

Documentation

What this platform does

AVAIL Research Platform is a demo interface for avail-omics. It accepts a ZIP with the generic AVAIL layout, validates it, runs three fast meta-learners, and returns a results ZIP plus a summary table.

For the full 11-model zoo, missing-data strategy benchmark, sensitivity, robustness and R export, install the package and run locally:

pip install "avail-omics[deep]"
avail-omics run-generic --data-root mydata --out results

Input layout

mydata.zip
  metadata.tsv          # required: SampleID, Group
  features/
    modality_a.tsv      # rows = samples, first column = SampleID
    modality_b.tsv
    ...
  compositional.txt     # optional

metadata.tsv may also contain Age, Education, Sex.
Each feature table must have a unique SampleID column and numeric features.
Missing modalities are represented by absent rows in that modality's feature file; do not write all-NaN rows.
At least two classes are required. The demo uses the first three most frequent classes; the full package accepts explicit --classes.
Raw DICOM/NIfTI is not accepted by the demo; use the package's imaging adapter locally.

Output

meta_benchmark_results.tsv
per_class_auc.tsv, loo_predictions.tsv, subject_classes.tsv
availability / coverage figures
a ZIP containing the run directory

Imageomics and radiomics

Imaging can be used as another modality: connectivity, optional radiomics and clinical/cognitive blocks are assembled into the standard features/*.tsv layout, then passed to run-generic. See IMAGEOMICS.md.

AI advisor and results agent

The package also ships an optional, bring-your-own-key AI layer:

avail-omics report --out results
avail-omics chat --out results
avail-omics agent --out results

report writes llm_advisor.json and llm_advisor_report.md.
chat / agent answers questions grounded in the run directory.
advise / hyper suggest models or hyperparameters.
The public demo does not call an LLM; this layer runs locally with your own key.

See DR_AVAIL.md for commands, providers, privacy and limits.

Data handling

The public demo is for de-identified or synthetic data only. Uploaded files are processed in a temporary directory and deleted after the job. For identifiable data, use the local package, Docker, or your own worker. See Private deployment.

Tutorial

1. Prepare a dataset

Create a ZIP with this structure:

mydata/
  metadata.tsv
  features/
    brain.tsv
    stool16S.tsv

Example metadata.tsv:

SampleID    Group   Age Sex
S001    HC  62  F
S002    MCI 71  M
S003    PDD 68  F

Example features/brain.tsv:

SampleID    FC_A    FC_B    FC_C
S001    0.41    0.22    0.11
S002    0.37    0.19    0.14
S003    0.29    0.31    0.08

A subject absent from features/stool16S.tsv is treated as missing that modality.

2. Run the public demo

Open the Analyze tab.
Confirm that the data are synthetic or de-identified.
Upload the ZIP.
Click Run quick benchmark.
Inspect the table and download the results ZIP.

The public demo uses three fast meta-learners; it is not the full 11-model run.

3. Move to the full pipeline

pip install "avail-omics[deep]"
avail-omics run-generic --data-root mydata --out results
avail-omics robustness --data-root mydata --out robustness
avail-omics sensitivity --data-root mydata --out sensitivity
avail-omics compare --entry "Cohort A=results" --out comparison

4. Synthetic demo

Use Run synthetic demo to see the output format without uploading anything. The cohort is generated locally and contains three modalities with deliberate missingness.

Dr. AVAIL — results-grounded agent

Dr. AVAIL is the package's optional assistant layer. It reads an existing results directory and helps interpret the run; it does not train a new model and it is not available as a public chat service.

What it does

recommends a meta-learner/strategy from the run tables
flags caveats (class imbalance, leakage risk, degenerate learners)
answers questions grounded in the run directory
grounds biological statements in retrieved PubMed records

Commands

avail-omics report --out results     # advisor report
avail-omics chat   --out results     # interactive agent
avail-omics agent  --out results     # alias

Example questions:

“Which meta-learner is most stable across subgroups?”
“Which caveats belong in the limitations section?”
“Does the evidence support a molecular interpretation?”
“Which model should I use for a new cohort?”

Bring your own key

avail-omics set-key
# or
export AVAIL_LLM_API_KEY=...

Providers: DeepSeek, OpenAI, Groq, Gemini, OpenRouter.

Where it runs

Dr. AVAIL runs locally, with your own key. The public site does not expose a live chat and does not call an LLM. Raw data stay on your machine; only derived summaries and prompts are sent to the provider.

Example report

Download the illustrative advisor report (PDF). It comes from a reference run, not from your data.

Limits

Not a clinical decision tool.
Not a substitute for expert review.
Citation grounding is retrieval-based; verify citations before submission.
No key: the core pipeline runs normally; AI sections are skipped.

Imageomics and radiomics

AVAIL treats imaging as another tabular modality: raw NIfTI/DICOM is converted into reproducible feature blocks, then passed to the same availability-aware pipeline as omics tables.

Feature blocks

Connectivity — Schaefer-100 (17 networks) functional connectivity, full and partial Fisher-z summaries.
Radiomics — optional PyRadiomics features from T1 (whole-brain and per-network) and, when requested, ALFF/fALFF/ReHo maps.
Clinical/cognitive — optional tabular blocks assembled alongside imaging.

Raw imaging is not accepted by the public demo. Imageomics runs locally or on a user-owned GPU/CPU worker because the preprocessing is heavy and the data are usually governed.

Build an imaging dataset locally

The imaging path is documented in the package repository:

AFNI preprocessing: tools/rsfmri_process_subject.sh
Connectivity extraction: rsfmri_extract_fc.py
Radiomics extraction: radiomics_extract.py
Dataset assembly: build_imaging_dataset.py

The result is the standard AVAIL layout:

mydata/
  metadata.tsv
  features/
    img_full.tsv
    img_partial.tsv
    clin_hy.tsv
    clin_cog.tsv
    clin_tests.tsv
    radiomics.tsv

Then run the generic pipeline:

avail-omics run-generic --data-root mydata --out results

Optional radiomics command

If the imaging features are already assembled, the package can generate the radiomics block:

avail-omics radiomics \
  --data-root mydata \
  --out radiomics \
  --mode both

--mode wb gives whole-brain features; --mode both adds the 17-network features. Functional maps can be added with --maps alff,falff,reho when the preprocessed BOLD images are available.

Missing imaging

A subject with no imaging is simply absent from the imaging feature file. AVAIL keeps that subject if another modality is present and represents the absence with availability flags. Partial imaging blocks are handled the same way.

Scope and limits

Radiomics is an engineering extension, not a validated clinical assay.
Preprocessing choices (atlas, smoothing, band-pass) affect the features.
The public site does not upload or process imaging data.
For real cohorts, use local infrastructure with the appropriate ethics and data-use approvals.

Local / Docker

pip install "avail-omics[deep]"
avail-omics run-generic --data-root mydata --out results
avail-omics figures --out results

Docker version:

docker build -t avail-demo .
docker run --rm -p 7860:7860 avail-demo

Colab

Download the notebook and upload it to Google Colab. It installs the package, accepts a ZIP and returns results.

Download Colab notebook

Open Google Colab

Colab runs in your own Google account; the public demo server is never used for your data.

Privacy

Do not upload identifiable patient data to any public site. For governed data, use the package, local Docker, or your own GPU worker.

Download

Python package

pip install --upgrade avail-omics
pip install "avail-omics[deep]"

PyPI: https://pypi.org/project/avail-omics/
Source: https://github.com/busdel/avail-cloud-runner
Gallery: https://busdel.github.io/avail-cloud-runner/viz/

Docker

docker build -t avail-demo .
docker run --rm -p 7860:7860 avail-demo

Then open http://localhost:7860.

Private deployment

For identifiable or governed data, see PRIVATE_DEPLOYMENT.md.

About AVAIL Research Platform

AVAIL Research Platform is the web-facing demo for avail-omics, an availability-aware multimodal integration framework for research cohorts with missing modalities. Each modality trains a base learner on the subjects who have it; a meta-learner combines the base probabilities with explicit availability flags, so subjects are not discarded or imputed away.

Research use only. Not a medical device.

Trust and data handling

No account required. The public demo does not create user profiles.
Temporary processing. Uploaded ZIPs are extracted into a per-job temporary directory and deleted after the run.
No raw-data logging. Application logs record job status, not feature values or sample identifiers.
No LLM by default. The demo does not send data to an external LLM.
Open source. The package and validation code are public; you can inspect and reproduce the pipeline.
Private path. For identifiable or governed data, run the package, Docker, or your own GPU worker locally. See Private deployment.

What the public demo does not do

It does not run the full 11-model benchmark.
It does not make diagnostic or clinical claims.
It does not accept identifiable patient data.
It does not replace local analysis or the published validation protocol.

Availability-aware multimodal analyticsfor incomplete clinical cohorts

Reference 11-model benchmark

Example AI advisor report

What the full pipeline returns

Documentation

What this platform does

Input layout

Output

Imageomics and radiomics

AI advisor and results agent

Data handling

Tutorial

1. Prepare a dataset

2. Run the public demo

3. Move to the full pipeline

4. Synthetic demo

Dr. AVAIL — results-grounded agent

What it does

Commands

Bring your own key

Where it runs

Example report

Limits

Imageomics and radiomics

Feature blocks

Build an imaging dataset locally

Optional radiomics command

Missing imaging

Scope and limits

Local / Docker

Colab

Privacy

Download

Python package

Docker

Private deployment

About AVAIL Research Platform

Trust and data handling

What the public demo does not do

Links

Availability-aware multimodal analytics
for incomplete clinical cohorts