somascan-ad-regression-tabpfn-v2

SomaScan → Alzheimer's-phenotype models — regression (distilled TabM students, TabPFN v2)

69 lightweight, standalone models that predict continuous Alzheimer's-disease–related phenotypes (cognition, neuropathology burden, motor/functional measures, demographics, a genetic risk score, and longitudinal change) from SomaScan plasma proteomics.

These accompany the manuscript "Fusing depth and breadth: Plasma proteomics scales deep Alzheimer's phenotyping to global cohorts." Discrete phenotypes (e.g. medication use, APOE genotype) live in the companion repo somascan-ad-classification-tabpfn-v2.

• Task: 69 regression phenotypes.
• Input: SomaScan 7K (v4.1) plasma proteomics, ANML-normalized, as natural-log RFU (ln(RFU); not log2 or log10) — aptamers named by SeqId (e.g. seq.10001.7); see input_schema.csv (7,287 aptamers). Each model uses its own ~500 selected aptamers.
• Artifact: one meta/<id>.json + one TabM .pt (~22 MB) per phenotype.
• Pipeline: TabPFN_pipeline (not needed at inference).

Quickstart

pip install torch tabm rtdl_num_embeddings numpy pandas

# predict ALL 69 phenotypes (wide output: sample_id + one column per phenotype)
python predict.py --input plasma_somascan.tsv --output predictions.csv

# specific phenotypes (name or target_id; comma-separated)
python predict.py --input plasma.tsv --output out.csv --target age_at_visit,cogn_global

# list what's available
python predict.py --list

predict.py depends on public packages only. Input is a CSV/TSV with a sample-id column (sample_id / SAMPLE.ID, optional) plus SomaScan aptamer columns matched by name; column order is irrelevant and extra columns are ignored. The models were trained on complete data and ship no imputer — provide complete aptamer values (NaN cells are rejected).

Repo layout

How the models were built

1. Feature pre-filter (capacity reduction, not selection): CatBoost ranks aptamers and the top 500 are passed to the teacher — solely to fit TabPFN's input limit (matches the manuscript: --method CatBoost, --selected_k 500).
2. Teacher: TabPFN v2 regressor (tabpfn-v2-regressor.ckpt), a foundation model.
3. Distillation → student: a TabM network (parameter-efficient MLP ensemble + piecewise-linear embeddings) is trained to reproduce the teacher's predictions on a GMM-augmented transfer set (gmm:10000), with a 5-quantile pinball loss; the released point estimate is the quantile-integral mean.
4. Deployable model: the teacher is dropped; only the student + preprocessing metadata ship. No feature scaling (models trained on unscaled data).

Performance (held-out cross-validation)

Evaluated on the manuscript's exact per-phenotype, person-grouped 10-fold partition (recovered and re-applied), so these numbers are directly comparable to the paper. The student is distilled separately within each fold.

Per-phenotype numbers are in parity_table.csv.

Intended use & limitations

• Research use only. Not a diagnostic device; predictions are population-level estimates.
• Platform-specific: trained on SomaScan aptamer measurements; not transferable to other proteomic platforms without recalibration.

License

Student weights are distilled derivatives of a TabPFN v2 teacher. The teacher (Prior-Labs/TabPFN-v2-reg) is released under the Prior Labs License — Apache 2.0 with an additional attribution requirement (https://priorlabs.ai/tabpfn-license/). As an Apache-2.0-based license, these derivative students may be redistributed and used provided the Prior Labs / TabPFN attribution is preserved.

Citation

If you use these models, cite the manuscript (citation TBD), TabPFN (Hollmann et al., Nature 2025), and TabM (Gorishniy et al., 2024).