FateFormer

Multimodal transformer models for cell fate prediction (reprogramming vs dead-end) from:

• RNA-seq
• ATAC-seq
• Metabolic flux

Prepare data

Download from Zenodo and place files in datasets/:

• clones.csv
• all_atac_d3_motif.h5ad
• flux_labelled.csv
• all_rna_d3_labelled.h5ad
• all_rna_d3_unlabelled.h5ad

Training / fine-tuning

Use notebooks:

• Model_RNA.ipynb (train on RNA only)
• Model_ATAC.ipynb (train on ATAC only)
• Model_Flux.ipynb (train on flux only)
• Model_Multimodal.ipynb (train multimodal model)

Full benchmark

python model_analysis.py

Default behavior:

• creates 4 models: RNA, ATAC, Flux, Multimodal
• uses 5-fold CV
• uses 5 seeds ([0, 6, 42, 123, 1000])
• Total runs: 100
• Writes outputs to: analysis docs/metrics/

Outputs:

• analysis docs/metrics/models/ - trained checkpoints per fold/seed/model
• analysis docs/metrics/metrics/ - CSV metric summaries
• analysis docs/metrics/fold_results/ - per-fold serialized results (.pkl)

Plotting and analysis

Open Plots.ipynb after model_analysis.py completes.