model-generalizability-assessment
Summary
Evaluate whether a trained graph neural network model for collision cross section prediction generalizes effectively to held-out test data by computing standard regression metrics (RMSE, MAE, R²) and comparing against baseline values. This skill is essential for validating whether GNN models can reliably predict CCS across diverse molecular datasets beyond their training distribution.
When to use
Apply this skill when you have a pre-trained GNN model for CCS prediction and need to verify that it generalizes to test data that was held out during training. Use it specifically when comparing model performance across different molecular datasets (e.g., METLIN-CCS vs. CCSBase) or when validating that a model trained on one dataset can predict accurately on another held-out dataset.
When NOT to use
- Input data has not been preprocessed or standardized to match the training pipeline (use data preprocessing skill first)
- No held-out test set or pre-defined train/test split exists (use data partitioning skill first)
- Model checkpoint is missing or incompatible with the inference framework
Inputs
- Pre-trained GNN model checkpoint (.h5 format)
- Test dataset in parquet format with columns: smiles, adduct, ccs, coordinates (optional)
- Parameter configuration JSON file specifying model hyperparameters
- Column name mappings for smiles, adduct, ccs, and coordinates
Outputs
- Numerical generalizability metrics (RMSE, MAE, R² values)
- Tabulated results comparing predictions to ground-truth test labels
- Metrics aggregated and compared against reported baseline values
- Output prefix-named prediction files from test set inference
How to apply
Load the pre-trained GNN model checkpoint and the held-out test dataset, ensuring data preprocessing (standardization, coordinate formatting) matches the training pipeline. Partition the data or use the pre-defined test split specified in the repository. Execute model inference on the test set to generate CCS predictions. Compute three generalizability metrics: RMSE (root mean squared error), MAE (mean absolute error), and R² (coefficient of determination) by comparing predictions against ground-truth test labels. Aggregate results into a results table and compare metric values against reported baselines from the paper to verify reproducibility and assess whether generalization performance meets expectations.
Related tools
- PyTorch Geometric (Graph neural network framework for loading, executing, and performing inference with pre-trained GNN model checkpoints)
- enveda/ccs-prediction repository (Source of pre-trained model weights, test datasets, parameter configurations, and training/inference scripts) — https://github.com/enveda/ccs-prediction
- train-test.py script (Python script in repository for training models and generating test set predictions) — https://github.com/enveda/ccs-prediction
Examples
poetry run python scripts/train-test.py --prefix "train-metlin-test-ccsbase" --train-input-file "ccs-prediction/metlin_train_3d.parquet" --test-input-file "ccs-prediction/ccsbase_3d.parquet" --parameter-path "parameter/parameter-train-metlin-test-metlin.json" --model-output-file "model/train-metlin-test-metlin.h5" --coordinates-column-name "coordinates" --coordinates-present --smiles-column-name "smiles" --adduct-column-name "adduct" --ccs-column-name "ccs" --dropout-rate 0.1 --epochs 400
Evaluation signals
- Computed RMSE, MAE, and R² values match or closely approximate the reported baseline metrics in the paper within expected numerical precision
- Test set predictions are generated successfully for all samples without runtime errors or NaN values
- Metrics are aggregated and tabulated in a consistent format enabling direct comparison to Table 1 or equivalent results in the publication
- R² value is positive and typically > 0.7 for well-generalizing models; RMSE and MAE should be in reasonable ranges relative to the CCS value distribution
- Generalizability assessment reveals consistent or declining performance across different training/test dataset combinations (e.g., train-metlin-test-ccsbase vs. train-ccsbase-test-metlin)
Limitations
- Generalizability assessment depends critically on the quality and representation of the held-out test set; if test data is too similar to training data, metrics may overestimate real-world generalization
- Metrics (RMSE, MAE, R²) are aggregated statistics that may mask poor performance on specific molecular subgroups or adduct types; per-group evaluation is recommended for thorough assessment
- Model generalization can degrade significantly when test data comes from a different instrument, acquisition protocol, or chemical space than the training set, limiting transferability across diverse experimental conditions
- The presence of 3D coordinates is optional; if coordinates are not provided, the model generates them, introducing variability that may affect reported metrics
Evidence
- [other] Compute generalizability metrics (RMSE, MAE, R²) comparing predictions against ground-truth test labels: "Compute generalizability metrics (RMSE, MAE, R²) comparing predictions against ground-truth test labels. Aggregate results and tabulate metric values alongside reported baseline values for"
- [other] Execute model inference on the test set using the trained GNN to generate collision cross section predictions: "Execute model inference on the test set using the trained GNN to generate collision cross section predictions."
- [other] Load the CCS dataset and apply any required preprocessing or standardization consistent with training: "Load the CCS dataset and apply any required preprocessing or standardization consistent with training. Partition data into train/validation/test splits or load the held-out test set as specified in"
- [readme] Poetry command to train models with explicit train and test input files: "poetry run python scripts/train-test.py --prefix "train-metlin-test-ccsbase" --train-input-file "ccs-prediction/metlin_train_3d.parquet" --test-input-file "ccs-prediction/ccsbase_3d.parquet""
- [readme] Model evaluation outputs specified with standardized column expectations: "ccs-column-name column name of the ccs, dropout-rate dropout rate of the model, epochs number of epochs to train the model"
1---2name: model-generalizability-assessment3description: Use when you have a pre-trained GNN model for CCS prediction and need to verify that it generalizes to test data that was held out during training. Use it specifically when comparing model performance across different molecular datasets (e.4license: CC-BY-4.05---67# model-generalizability-assessment89## Summary1011Evaluate whether a trained graph neural network model for collision cross section prediction generalizes effectively to held-out test data by computing standard regression metrics (RMSE, MAE, R²) and comparing against baseline values. This skill is essential for validating whether GNN models can reliably predict CCS across diverse molecular datasets beyond their training distribution.1213## When to use1415Apply this skill when you have a pre-trained GNN model for CCS prediction and need to verify that it generalizes to test data that was held out during training. Use it specifically when comparing model performance across different molecular datasets (e.g., METLIN-CCS vs. CCSBase) or when validating that a model trained on one dataset can predict accurately on another held-out dataset.1617## When NOT to use1819- Input data has not been preprocessed or standardized to match the training pipeline (use data preprocessing skill first)20- No held-out test set or pre-defined train/test split exists (use data partitioning skill first)21- Model checkpoint is missing or incompatible with the inference framework2223## Inputs2425- Pre-trained GNN model checkpoint (.h5 format)26- Test dataset in parquet format with columns: smiles, adduct, ccs, coordinates (optional)27- Parameter configuration JSON file specifying model hyperparameters28- Column name mappings for smiles, adduct, ccs, and coordinates2930## Outputs3132- Numerical generalizability metrics (RMSE, MAE, R² values)33- Tabulated results comparing predictions to ground-truth test labels34- Metrics aggregated and compared against reported baseline values35- Output prefix-named prediction files from test set inference3637## How to apply3839Load the pre-trained GNN model checkpoint and the held-out test dataset, ensuring data preprocessing (standardization, coordinate formatting) matches the training pipeline. Partition the data or use the pre-defined test split specified in the repository. Execute model inference on the test set to generate CCS predictions. Compute three generalizability metrics: RMSE (root mean squared error), MAE (mean absolute error), and R² (coefficient of determination) by comparing predictions against ground-truth test labels. Aggregate results into a results table and compare metric values against reported baselines from the paper to verify reproducibility and assess whether generalization performance meets expectations.4041## Related tools4243- **PyTorch Geometric** (Graph neural network framework for loading, executing, and performing inference with pre-trained GNN model checkpoints)44- **enveda/ccs-prediction repository** (Source of pre-trained model weights, test datasets, parameter configurations, and training/inference scripts) — https://github.com/enveda/ccs-prediction45- **train-test.py script** (Python script in repository for training models and generating test set predictions) — https://github.com/enveda/ccs-prediction4647## Examples4849```50poetry run python scripts/train-test.py --prefix "train-metlin-test-ccsbase" --train-input-file "ccs-prediction/metlin_train_3d.parquet" --test-input-file "ccs-prediction/ccsbase_3d.parquet" --parameter-path "parameter/parameter-train-metlin-test-metlin.json" --model-output-file "model/train-metlin-test-metlin.h5" --coordinates-column-name "coordinates" --coordinates-present --smiles-column-name "smiles" --adduct-column-name "adduct" --ccs-column-name "ccs" --dropout-rate 0.1 --epochs 40051```5253## Evaluation signals5455- Computed RMSE, MAE, and R² values match or closely approximate the reported baseline metrics in the paper within expected numerical precision56- Test set predictions are generated successfully for all samples without runtime errors or NaN values57- Metrics are aggregated and tabulated in a consistent format enabling direct comparison to Table 1 or equivalent results in the publication58- R² value is positive and typically > 0.7 for well-generalizing models; RMSE and MAE should be in reasonable ranges relative to the CCS value distribution59- Generalizability assessment reveals consistent or declining performance across different training/test dataset combinations (e.g., train-metlin-test-ccsbase vs. train-ccsbase-test-metlin)6061## Limitations6263- Generalizability assessment depends critically on the quality and representation of the held-out test set; if test data is too similar to training data, metrics may overestimate real-world generalization64- Metrics (RMSE, MAE, R²) are aggregated statistics that may mask poor performance on specific molecular subgroups or adduct types; per-group evaluation is recommended for thorough assessment65- Model generalization can degrade significantly when test data comes from a different instrument, acquisition protocol, or chemical space than the training set, limiting transferability across diverse experimental conditions66- The presence of 3D coordinates is optional; if coordinates are not provided, the model generates them, introducing variability that may affect reported metrics6768## Evidence6970- [other] Compute generalizability metrics (RMSE, MAE, R²) comparing predictions against ground-truth test labels: "Compute generalizability metrics (RMSE, MAE, R²) comparing predictions against ground-truth test labels. Aggregate results and tabulate metric values alongside reported baseline values for"71- [other] Execute model inference on the test set using the trained GNN to generate collision cross section predictions: "Execute model inference on the test set using the trained GNN to generate collision cross section predictions."72- [other] Load the CCS dataset and apply any required preprocessing or standardization consistent with training: "Load the CCS dataset and apply any required preprocessing or standardization consistent with training. Partition data into train/validation/test splits or load the held-out test set as specified in"73- [readme] Poetry command to train models with explicit train and test input files: "poetry run python scripts/train-test.py --prefix "train-metlin-test-ccsbase" --train-input-file "ccs-prediction/metlin_train_3d.parquet" --test-input-file "ccs-prediction/ccsbase_3d.parquet""74- [readme] Model evaluation outputs specified with standardized column expectations: "ccs-column-name column name of the ccs, dropout-rate dropout rate of the model, epochs number of epochs to train the model"
HolobiomicsLab 