PLM-based Linker activity prediction: train and predict with a single config and single entry point.

• Input: Excel with amino acid sequence column and activity column.
• Pipeline: Extract PLM features → optional Ca PCA → K-fold MLP training → optional prediction → log to CSV.
• Output: Models under output/{timestamp}/, predictions as CSV, run log in logs/model_results_log.csv.

• Conda env: Linker (recommended). Activate with conda activate Linker.
• Install: pip install -r requirements.txt
• Set PLM model path in config (e.g. ESM2, Prot-BERT, Prot-T5); see configs/default_unified.json.

# Default: train + predict + log (uses configs/default_unified.json)
python run.py

# Specify config and/or mode
python run.py --config configs/default_unified.json
python run.py --config configs/input.json --mode train
python run.py --config configs/input.json --mode predict
python run.py --mode train_and_predict

• Modes: train, predict, train_and_predict.
• Config: Single JSON with train, predict, log, mode. Example: configs/default_unified.json.

• run.py — Entry script (train / predict / train_and_predict).
• config/ — Config loading and schema (load.py, schema.py).
• pipeline/ — Train, predict, and log helpers.
• main.py — Core run_train, run_predict, train/evaluate/predict logic.
• models.py — K-fold MLP, trainer, evaluator, predictor.
• datasets.py, datasets_ca_pca.py — PLM features, PCA, Ca PCA.
• dataloader.py, features.py — Structure/distance data and features.
• utils.py, infra.py — Helpers and paths.
• configs/ — JSON configs (default: default_unified.json).

• Models: output/{timestamp}/{timestamp}-train-{ac_col}/ (e.g. model_0.pth, ...).
• Features: output/{timestamp}/features/.
• Predictions: output/{timestamp}/predict_{output_name}.csv.
• Log: logs/model_results_log.csv.