---
license: apache-2.0
tags:
  - biology
  - proteomics
  - aging
  - aging-clock
  - tabular
  - tabm
  - distillation
  - somascan
pipeline_tag: tabular-regression
---

# SomaScan Inflammation Aging Clock (85 proteins)

A lightweight **plasma-protein aging clock** that predicts chronological age from
**85 unique inflammatory proteins measured by SomaScan** (125 aptamers /
SomaScan features). The model is a **TabM† student** distilled from a **TabPFN
v2** teacher, so it runs at inference **without any TabPFN dependency** (small,
DUA-friendly artifacts).

Trained on the **ROSMAP** cohort, panel-matched to the Olink Inflammation
companion model.

| Cohort | Platform | Aptamers (features) | Unique proteins | N (persons) | Teacher R² | Student R² | Gap (mean ± sd) |
|---|---|---|---|---|---|---|---|
| ROSMAP | SomaScan | 125 | 85 | 1,611 (1,313) | 0.304 | 0.302 | 0.002 ± 0.006 |

*(10-fold person-grouped CV, R². Student fidelity-RMSE ≈ 0.070 in scaled-y.)* The
TabM† student reproduces its TabPFN v2 teacher to within ~0.002 R² — inside
fold-to-fold noise.

> **Note on counts:** SomaScan measures some proteins with more than one aptamer.
> The model consumes **125 aptamer-level features** (the `meta.json` →
> `feature_name` list), which map to **85 unique proteins**.

## Files

```
predict.py                       standalone inference script
meta.json                        feature order, y-scaler, arch config
models/
  T0001_model.pkl                slim model (~4 KB)
  T0001_student_tabm.pt          TabM† weights (~8 MB) + quantile bins
results/, *_results.csv          aggregate CV metrics summaries
```

Target `T0001` = `age_at_visit` (chronological age at the visit).

## Usage

`predict.py` runs the model with **only public dependencies**:

```bash
pip install torch tabm rtdl_num_embeddings numpy pandas

python predict.py --input proteins.csv --output ages.csv
```

`--input` is a CSV/TSV with one row per sample and one column per aptamer feature,
named exactly as in `meta.json` → `feature_name` (125 SomaScan features). Column
order does not matter; an optional `sample_id` column is carried through. Output
is `sample_id, predicted_age`. This model expects complete values (no imputer).
The script errors if any required feature is missing and warns about unused
input columns.

## Method

- **Teacher:** TabPFN v2.
- **Student:** TabM† distilled on a GMM-augmented transfer set (target 10,000
  rows) with a 5-quantile pinball regression loss.
- **CV:** 10-fold person-grouped (verified no person spans >1 fold).

## Inference contract

Order the features by `meta.json → feature_name`; rebuild the TabM model with
piecewise-linear numeric embeddings and load the weights; average the 5-quantile
(trapezoidal-mean) point estimate over the 32 ensemble members; inverse the
y-scaler to recover age.

## Citation

Inflammatory aging clock for plasma inflammatory proteins (ROSMAP). Manuscript in
preparation. See also the Olink companion model:
[`inflammatory-aging-clock/olink-inflammation-92`](https://huggingface.co/inflammatory-aging-clock/olink-inflammation-92).