Chromatographic Method Transfer Learning

chromatographic-method-transfer-learning

Summary

Apply Bayesian meta-learning to project retention times (RTs) across chromatographic methods using a minimal calibration set (≥10 molecules), enabling rapid method transfer without full retraining. This skill bridges different LC-MS platforms by learning a probabilistic RT projection mapping from a small set of known compounds.

When to use

You have experimental RT measurements from a source chromatographic method and need to predict RTs for the same molecules on a target chromatographic method, but lack a large calibration dataset (typical scenario: 10–100 molecules with ground truth RTs on both methods). Use this skill when complete retraining on the target method is impractical or when you need to integrate a new instrument into an existing metabolomics workflow.

When NOT to use

• Calibration dataset is already large (>500 molecules): standard supervised retraining on the target method will be more stable and accurate than meta-learning.
• Source and target methods are fundamentally incompatible (e.g., completely different separation chemistry or detector type): RT projections may not preserve ranking or linearity.
• Only a single molecule is available for calibration: meta-learning requires ≥10 molecules to learn a meaningful prior.

Inputs

• Calibration molecule set with SMILES or chemical structure files (SDF, mol, mol2, hin)
• Experimental retention times for calibration molecules on both source and target chromatographic methods
• Molecular descriptor/fingerprint vectors (MACCS166, Extended Connectivity, Path Fingerprints) generated by alvaDesc
• Test molecule set with known RTs on source method (to be projected to target method)

Outputs

• Trained Bayesian meta-learned Gaussian Process prior (saved model object)
• Projected retention times for test molecules on the target chromatographic method
• Mean absolute error (MAE) and median absolute error (MedAE) metrics comparing projected vs. experimental RTs
• Performance comparison table against baseline methods (linear regression, standard regressors)

How to apply

First, generate molecular descriptors and fingerprints (MACCS166, Extended Connectivity, Path Fingerprints) for all calibration molecules using alvaDesc software. Train a Bayesian meta-learning Gaussian Process on this small set (≥10 molecules) to learn the RT projection mapping between source and target chromatographic methods; the meta-learner infers a prior from the calibration data rather than starting from scratch. Apply the trained meta-learner to project RTs for test molecules onto the target method. Evaluate performance by calculating mean absolute error (MAE) and median absolute error (MedAE) of projected RTs against ground-truth experimental values on held-out test molecules. Compare error rates (target: ≤40 s MAE, ≤20 s MedAE based on the DNN baseline) against traditional linear regression or standard ML regressors to confirm the meta-learned projection is competitive.

Related tools

• alvaDesc (Generates 5,666 molecular descriptors and 2,214 fingerprints (MACCS166, Extended Connectivity, Path Fingerprints) required as input features for the meta-learning model) — https://www.alvascience.com/alvadesc/
• cmmrt (CMM-RT) (Reference implementation of Bayesian meta-learning for RT projection; provides notebooks, data processing pipelines, and trained models for fingerprint generation and GP meta-training) — https://github.com/constantino-garcia/cmmrt

Examples

python cmmrt/rt/train_model.py --storage sqlite:///results/optuna/train.db --save_to saved_models; python -c "from cmmrt.projections import meta_train_gp; gp = meta_train_gp(calibration_fingerprints, calibration_rts_source, calibration_rts_target, n_epochs=100); projected_rts = gp.predict(test_fingerprints)"

Evaluation signals

• MAE and MedAE of projected RTs should be ≤40 s and ≤20 s respectively (competitive with DNN baseline 39.2±1.2 s and 17.2±0.9 s).
• Projection error should not degrade substantially when calibration set size is exactly 10 molecules (paper's minimum threshold).
• Performance on test molecules should not be significantly worse than baselines (linear regression, standard ML regressors) measured on the same held-out test set.
• Predicted RT ranks for test molecules should preserve compound ordering relative to the target chromatographic method (ordinal correlation).
• Meta-trained model should generalize across different target chromatographic methods without retraining (transfer learning invariant).

Limitations

• Requires alvaDesc software (commercial license) to generate fingerprints; no open-source fingerprint generation substitute is mentioned in the repository.
• Performance depends critically on calibration molecule representativeness: if calibration set is biased toward short or long retention times, projection accuracy may degrade for underrepresented chemical space.
• Method assumes RT correlation is learnable from a small calibration set; highly non-linear or noisy RT relationships between methods may exceed meta-learning's capacity.
• Code repository is marked as under active development; reproducibility depends on using the paper branch as noted in the README, not the main development branch.

Evidence

• [intro] A novel Bayesian meta-learning approach is proposed for RT projection between CMs from as few as 10 molecules while still obtaining competitive error rates compared with previous approaches.: "A novel Bayesian meta-learning approach is proposed for RT projection between CMs from as few as 10 molecules while still obtaining competitive error rates compared with previous approaches."
• [intro] 5,666 molecular descriptors and 2,214 fingerprints (MACCS166, Extended Connectivity, and Path Fingerprints fingerprints) were generated with the alvaDesc software: "5,666 molecular descriptors and 2,214 fingerprints (MACCS166, Extended Connectivity, and Path Fingerprints fingerprints) were generated with the alvaDesc software"
• [intro] The best results were obtained by a heavily regularized DNN trained with cosine annealing warm restarts and stochastic weight averaging, achieving a mean and median absolute errors of 39.2±1.2 s and 17.2 ± 0.9 s, respectively.: "The best results were obtained by a heavily regularized DNN trained with cosine annealing warm restarts and stochastic weight averaging, achieving a mean and median absolute errors of 39.2±1.2 s and"
• [readme] To train your own model or to predict the RT of your own set of compounds it is necessary to generate the fingerprints using alvaDesc software (under license, check alvadesc software).: "To train your own model or to predict the RT of your own set of compounds it is necessary to generate the fingerprints using alvaDesc software (under license, check alvadesc software)."
• [readme] Note that, to integrate the proposal into the CEU Mass Mediator platform, the code in this repository will continue to be developed. Hence, branch paper should be used as reference for reproducing the results of the paper.: "Note that, to integrate the proposal into the CEU Mass Mediator platform, the code in this repository will continue to be developed. Hence, branch paper should be used as reference for reproducing"