HolobiomicsLab retention-time-prediction-optimization
License: restricted — no clear open-source license detected for the underlying tool; verify licensing before commercial use or redistribution.
Summary
Train and optimize graph neural network models to predict liquid chromatography retention times for small molecules using the ABCoRT framework. This skill enables fine-tuning of pre-trained models on domain-specific retention-time datasets via transfer learning, improving prediction accuracy for chemical compound characterization.
When to use
When you have a retention-time dataset (e.g., SMRT or Eawag_XBridgeC18_364) in .xlsx format and need to train or adapt a graph neural network model to predict chromatographic retention times for new compounds. Use this skill when baseline model performance is insufficient and transfer learning from a pre-trained checkpoint can accelerate convergence.
When NOT to use
- Input dataset is not in .xlsx format or lacks expected column structure (SMILES, retention time labels)
- Pre-trained checkpoint is unavailable or corrupted and baseline model training is not feasible
- Retention-time data is from an entirely different chromatographic platform (e.g., GC-MS instead of LC) not covered by pre-training or transfer-learning datasets
Inputs
- Pre-trained model checkpoint (PyTorch .pt or .pth file)
- Retention-time dataset in .xlsx format (e.g., Eawag_XBridgeC18_364.xlsx or SMRT.xlsx)
- SMILES or molecular structure representations (processed via RDKit)
- Command-line arguments specifying dataset path and optional hyperparameters
Outputs
- Fine-tuned model checkpoint saved to disk
- Training logs and metrics (loss, validation accuracy) per epoch
- Inference predictions on test or validation splits
- TorchMetrics monitoring artifacts
How to apply
Execute the training or transfer-learning workflow using Python entry points. For baseline training on the SMRT retention-time dataset, run 'python train_SMRT.py' to instantiate the PyTorch + PyG graph neural network and train from scratch, monitoring loss and validation metrics via TorchMetrics over epochs. For domain adaptation, run 'python train_transfer_FE.py --DataSet .xlsx' to load a pre-trained checkpoint, fine-tune on your target dataset, and save the adapted model. Monitor training metrics (loss, validation accuracy) to detect convergence; save the final model checkpoint and logs upon completion. The choice between baseline and transfer training depends on data availability: transfer learning is preferred when fine-tuning datasets are small relative to the pre-training corpus.
Related tools
- PyTorch (Deep learning framework for graph neural network model training and optimization)
- PyG (PyTorch Geometric) (Graph neural network library for molecular graph construction and message passing)
- RDKit (Cheminformatics toolkit for converting SMILES to molecular graphs and feature extraction)
- Pandas (Data loading and preprocessing for retention-time .xlsx files)
- TorchMetrics (Real-time training metrics monitoring (loss, validation accuracy) during fine-tuning)
- torch-scatter, torch-sparse, torch-cluster (GPU-accelerated graph operations for efficient PyG tensor aggregation and sampling)
Examples
python train_transfer_FE.py --DataSet Eawag_XBridgeC18_364.xlsx
Evaluation signals
- Training loss decreases monotonically or plateaus at a minimum over epochs, indicating model convergence
- Validation metrics (e.g., mean absolute error on held-out retention times) improve or stabilize, showing generalization
- Saved checkpoint file exists and is loadable by PyTorch without schema errors
- Inference predictions on test set fall within the expected retention-time range (e.g., 1–30 min for typical LC) and match literature or experimental references
- TorchMetrics logs record non-NaN values for all tracked metrics at each epoch
Limitations
- Transfer learning performance depends on similarity between pre-training data (SMRT) and target dataset; significant chromatographic differences may limit transfer gains
- Model assumes molecular structures are representable as graphs with RDKit; SMILES parsing failures will cause training to fail
- No explicit hyperparameter tuning guidance provided in README; optimal learning rate, batch size, and epochs must be determined empirically
- Thirteen transfer-learning datasets are mentioned but only one (Eawag_XBridgeC18_364.xlsx) is exemplified; dataset availability and format consistency are not guaranteed
Evidence
- [readme] train_SMRT.py script entry point for baseline model training: "If you want to train the Model. Please command
python train_SMRT.py
```"
- [readme] transfer learning workflow on thirteen datasets with dataset parameter: "If you want to run the transfer learning on thirteen transfer learning data sets, use:
python train_transfer_FE.py --DataSet Eawag_XBridgeC18_364.xlsx
- [intro] graph neural network architecture training on SMRT retention-time dataset: "Execute the training entry point using Python with the command 'python train_SMRT.py', which will instantiate the graph neural network architecture (PyTorch + PyG)"
- [intro] TorchMetrics monitoring during fine-tuning process: "Monitor training metrics via TorchMetrics during fine-tuning. 4. Save the fine-tuned model checkpoint and training logs."
- [intro] RDKit, PyTorch, and PyG as core tools for molecular representation and training: "Execute train_transfer_FE.py with the --DataSet parameter set to Eawag_XBridgeC18_364.xlsx using PyTorch and PyG for graph neural network operations."