Few shot learning calibration design

few-shot-learning-calibration-design

License: restricted — no clear open-source license detected for the underlying tool; verify licensing before commercial use or redistribution. <!-- asb-license-banner -->

Summary

A Bayesian meta-learning approach that projects retention times between chromatographic methods using minimal calibration data (≥10 molecules), enabling rapid method transfer without extensive experimental validation. This skill applies Gaussian Process meta-learning to learn transferable RT mappings from small reference molecule sets, achieving competitive accuracy against traditional regression baselines.

When to use

You have experimental retention times measured on a source chromatographic method and want to predict RTs on a target chromatographic method, but possess only a small set (10–100) of molecules with ground-truth measurements on both methods. This situation arises during analytical method harmonization, platform migration, or when establishing calibration for newly acquired instruments where full re-validation is impractical.

When NOT to use

• Calibration set is smaller than 10 molecules; meta-learning requires sufficient diversity to learn the projection manifold.
• Source and target chromatographic methods operate under fundamentally different physical principles (e.g., hydrophilic interaction vs. reversed-phase); molecular feature space may not bridge the gap.
• You have access to a large training dataset (>1000 molecules) with full experimental coverage; standard supervised regression is more efficient and interpretable than meta-learning.

Inputs

• Calibration molecule set with known RTs on both source and target chromatographic methods (≥10 molecules)
• SMILES strings or structural files for calibration and test molecules
• Molecular descriptors and fingerprints (MACCS166, Extended Connectivity, Path Fingerprints) generated via alvaDesc

Outputs

• Trained Bayesian meta-learned Gaussian Process projection model
• Predicted retention times for test molecules on target chromatographic method
• Mean absolute error (MAE) and median absolute error (MedAE) metrics
• Uncertainty estimates for projected retention times

How to apply

First, generate standardized molecular descriptors and fingerprints (MACCS166, Extended Connectivity, Path Fingerprints) for the small calibration molecule set using alvaDesc software. Train a Bayesian Gaussian Process meta-learner on this calibration set to model the RT transformation function between source and target chromatographic methods. The meta-learner learns a distribution over projection functions rather than a single point estimate, enabling uncertainty quantification. Apply the trained meta-learner to project RTs for uncharacterized test molecules. Evaluate performance by computing mean absolute error (MAE) and median absolute error (MedAE) between predicted and ground-truth experimental RTs on a held-out validation set, comparing against linear regression and standard machine learning regressors (random forest, gradient boosting) as baselines. The meta-learning framework leverages the structure of the calibration molecules to generalize effectively despite the small sample size.

Related tools

• alvaDesc (Generates 5,666 molecular descriptors and 2,214 fingerprints (MACCS166, Extended Connectivity, Path Fingerprints) for calibration and test molecules prior to meta-learning training) — https://www.alvascience.com/alvadesc/
• cmmrt (Implements the Bayesian meta-learning approach, Gaussian Process projection training, and RT prediction pipelines for chromatographic method transfer) — https://github.com/constantino-garcia/cmmrt

Examples

make test_projections  # Tests meta-trained GP projections on reference chromatographic methods using PredRet database; outputs CSV and figure summaries to results/projection/

Evaluation signals

• MAE and MedAE of projected RTs fall within the range of published baselines (39.2±1.2 s and 17.2±0.9 s for best-performing DNN), confirming competitive accuracy.
• Projected RTs exhibit correlation with ground-truth experimental values on held-out test molecules; residuals should be approximately normally distributed with mean near zero.
• Meta-learned uncertainty estimates (posterior variance) are well-calibrated: confidence intervals should contain the true RT with the stated coverage probability.
• Performance remains stable across different random seeds and cross-validation folds, indicating the meta-learner has not overfit the small calibration set.
• Error rates do not degrade significantly when the number of calibration molecules is reduced from the full set to the minimum (≥10), validating the few-shot learning property.

Limitations

• Bayesian meta-learning requires careful tuning of hyperparameters (kernel choice, prior specification, meta-training iterations); poorly tuned models may underperform simpler baselines on small calibration sets.
• The approach assumes the source and target chromatographic methods operate under related physical principles such that a smooth RT transformation exists; if methods differ fundamentally, learned projections may be unreliable.
• Fingerprint-based features (MACCS166, ECFP, PFP) capture 2D molecular structure only; they may not capture method-specific effects (e.g., surface silanols, column aging, buffer pH) that influence RT in practice.
• Generalization beyond the chemical space represented in the calibration set is limited; molecules with unprecedented structural features may have high prediction uncertainty.
• The code repository notes that the main branch is under active development for integration into the CEU Mass Mediator platform; users should refer to the 'paper' branch to reproduce published results.

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"
• [intro] Train machine learning regressors on retention time data using experimental RTs from the METLIN small molecule dataset: "We have trained state-of-the-art machine learning regressors using the 80,038 experimental RTs from the METLIN small molecule dataset (SMRT)"
• [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"
• [readme] Meta-train a Gaussian Process (GP) for computing projections between CMs using the PredRet database. The resulting GP is stored in the saved_models folder.: "Meta-train a Gaussian Process (GP) for computing projections between CMs using the PredRet database for a few epochs. The resulting GP is stored in the saved_models folder."