How charlotteinfante/t5pka works
T5pKa predicts molecular microstates with a sequence-to-sequence model and estimates microscopic pKa values with a regression model. The main user workflow runs prediction from SMILES or input files, tokenizes and batches molecular strings, invokes a trained model, and writes predictions; optional ensemble prediction combines model runs. Separate paths support model pretraining and fine-tuning, evaluation, and SHAP-based interpretation. The README establishes the two-stage prediction capability; source excerpts show selected runtime connections, so unsampled internals are represented conservatively.
Main entry points: mol_tokenizers.py Molecule tokenizers · run_prediction.py Prediction runner · run_trainer.py Task fine-tuning · run_pretrain.py Model pretraining
Components 20
Prediction
- CLI dispatcher
t5pka/__main__.py - Prediction runner
t5pka/run_prediction.py - Ensemble prediction
t5pka/ensemble_prediction.py - Prediction files
Modeling
- Molecule tokenizers
t5pka/mol_tokenizers.py - T5 model
t5pka/model.py - Token vocabularies
t5pka/vocab - Training engine
t5pka/trainer.py - Task fine-tuning
t5pka/run_trainer.py - Model pretraining
t5pka/run_pretrain.py - Model checkpoints
Data and evaluation
- Task datasets
t5pka/data_utils.py - Prediction evaluation
t5pka/evaluation.py - Ensemble data splitting
data_scripts/ensemble.py - Legacy data splitting
data_scripts/ensemble_old.py - Seq2seq conversion
data_scripts/seq2seq_to_micropka.py - Molecular data utilities
data_scripts/utils.py
Interpretation
- SHAP interpretation
t5pka/shap_pka.py - SHAP analysis script
data_scripts/shap_pka.py
Other components
- Researcher
Connections 31
- Researcher → CLI dispatcher: starts task
README.md - CLI dispatcher → Prediction runner: dispatches prediction
t5pka/__main__.py - Prediction runner → Molecule tokenizers: uses
t5pka/run_prediction.py - Prediction runner → Task datasets: builds batches
t5pka/run_prediction.py - Prediction runner → T5 model: runs inference
t5pka/run_prediction.py - Prediction runner → Prediction files: writes predictions
README.md - Ensemble prediction → Molecule tokenizers: uses
t5pka/ensemble_prediction.py - Ensemble prediction → Task datasets: builds batches
t5pka/ensemble_prediction.py - Ensemble prediction → T5 model: runs inference
t5pka/ensemble_prediction.py - Ensemble prediction → Prediction files: writes predictions
README.md - Molecule tokenizers → Token vocabularies: loads vocabulary
t5pka/mol_tokenizers.py - Prediction runner → Prediction evaluation: uses metrics
t5pka/run_prediction.py - Prediction files → Prediction evaluation: evaluates
t5pka/evaluation.py - CLI dispatcher → Task fine-tuning: dispatches training
t5pka/__main__.py - Task fine-tuning → Task datasets: loads task data
t5pka/run_trainer.py - Task fine-tuning → Molecule tokenizers: constructs tokenizer
t5pka/run_trainer.py - Task fine-tuning → T5 model: configures model
t5pka/run_trainer.py - Task fine-tuning → Training engine: trains model
t5pka/run_trainer.py - Task fine-tuning → Model checkpoints: saves checkpoints
t5pka/run_trainer.py - CLI dispatcher → Model pretraining: dispatches pretraining
t5pka/__main__.py - Model pretraining → Task datasets: loads pretrain data
t5pka/run_pretrain.py - Model pretraining → Molecule tokenizers: constructs tokenizer
t5pka/run_pretrain.py - Model pretraining → T5 model: configures model
t5pka/run_pretrain.py - Model pretraining → Training engine: trains model
t5pka/run_pretrain.py - Model pretraining → Model checkpoints: saves checkpoints
t5pka/run_pretrain.py - Ensemble data splitting → Molecular data utilities: uses utilities
data_scripts/ensemble.py - Legacy data splitting → Molecular data utilities: uses utilities
data_scripts/ensemble_old.py - SHAP interpretation → T5 model: explains predictions
t5pka/shap_pka.py - SHAP interpretation → Molecule tokenizers: uses
t5pka/shap_pka.py - SHAP analysis script → T5 model: uses model
data_scripts/shap_pka.py - SHAP analysis script → Molecule tokenizers: uses
data_scripts/shap_pka.py