Multi-modal Modeling

MESA Modeling

CFTK includes MESA-style multimodal modeling commands.

Install the analysis dependencies, including mesa-cfdna, before using these commands:

python -m pip install ".[analysis]"

Run modality performance screening:

cftk --config cftk_init.json mesa --performance

Run model construction:

cftk --config cftk_init.json mesa --mesa-model

Run leave-one-out cross-validation and plots:

cftk --config cftk_init.json mesa --loocv

Commonly used together:

cftk --config cftk_init.json mesa --performance --mesa-model --loocv

Expected Outputs

MESA writes a compact set of tables, a serialized model, and LOOCV figures:

results/5_mesa/
|-- label.tsv
|-- modality_performance.tsv
|-- MESA_model.pkl
|-- loocv_predictions.tsv
|-- mesa_roc.png / mesa_roc.pdf
|-- mesa_heatmap.png / mesa_heatmap.pdf
`-- mesa_spearman.png / mesa_spearman.pdf

The observed visual below combines those plot families with the prediction table for five controls and five sALS samples. It is a technical workflow example only. In particular, perfect-looking internal screening values can occur in a ten-sample run and do not estimate performance for a cohort or a clinical assay. Inspect loocv_predictions.tsv, the model settings, and the provenance manifest before interpreting any result.

Observed MESA output from five controls and five sALS samples. The internal screening bars and LOOCV curves are descriptive artifacts from this ten-sample technical run, not biological or clinical validation and not a recommended acceptance threshold. Download the sanitized aggregate metadata: JSON.