eSPred: Explainable scRNA-seq Prediction via Customized Foundation Models and Pathway-Aware Fine-tuning
eSPred is a framework that enhances foundation models for single-cell RNA sequencing (scRNA-seq) analysis through cell-type-informed input strategies and biological pathway integration to achieve superior subject-level predictions and interpretability.
eSPred integrates three key components to bridge cellular heterogeneity with clinical outcome prediction:
- Cell-type-aware Grouping: A customized pre-training strategy that groups cells of the same type to generate more informative cell embeddings
- Pathway-guided Decoder: A biologically-informed neural architecture that integrates Reactome pathway knowledge during fine-tuning
- Hierarchical Classification: A two-step approach that aggregates cell-level predictions to subject-level outcomes
## Option 1: Install via requirements.txt
pip install -r requirements.txt
## Option 2: Follow scGPT installation instructions
See scGPT.README.md for detailed instructionsThis step enhances the foundation model by incorporating cell-type information through strategic grouping.
python pretrain.py \
--data-source $DATASET \
--save-dir ./save/eval-$(date +%b%d-%H-%M-%Y) \
--load-model $CHECKPOINT \
--max-seq-len $MAX_LENGTH \
--batch-size $per_proc_batch_size \
--eval-batch-size $(($per_proc_batch_size * 2)) \
--epochs 100 \
--log-interval $LOG_INTERVAL \
--trunc-by-sample \
--no-cls \
--no-cce \
--fp16This step integrates biological pathway knowledge into the model during fine-tuning.
python finetune.py \
--dataset_name covid \
--load_model /path/to/pretrained_model \
--output_dir ./results/covid \
--epochs 50 \
--batch_size 32 \
--lr 1e-4 \
--layer_size 128 \
--n_hidden 2 \
--dropout 0.2 \
--freeze