Retention time prediction scoring

Retention Time Prediction Scoring

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

Summary

Score and rank metabolite annotation candidates by comparing observed chromatographic retention times against DNN-predicted values with confidence intervals, filtering candidates whose observed RT falls within projected uncertainty bounds. This probabilistic scoring integrates predicted RTs (MAE 39.2±1.2 s) with meta-learned RT projections between chromatographic methods to increase annotation specificity.

When to use

You have a list of candidate metabolites for an unknown compound (from mass-to-structure search or library matching), experimental retention time(s) from one or more chromatographic methods, and access to a trained DNN RT predictor and meta-learned RT projection model. Use this skill to rank candidates by how well their predicted RTs match observed RTs, prioritizing high-confidence matches and discarding outliers.

When NOT to use

• Input is a single metabolite without alternatives to rank or score—use RT prediction alone instead.
• Observed retention times are from an unmapped chromatographic method and you have <10 calibration molecules—the meta-learned projection will not be reliable; collect more calibration data or use a baseline linear regression.
• The DNN model has not been retrained or validated on your specific chromatographic method or ionization mode—model generalization is limited; retrain on representative data from your platform.

Inputs

• List of candidate metabolites with molecular structures (SMILES, SDF, or mol format)
• Observed retention time value(s) from liquid chromatography (numeric, in seconds)
• Pre-trained DNN retention time prediction model (serialized neural network weights)
• Meta-learned Bayesian GP prior for RT projection (if projecting between chromatographic methods)
• Calibration molecule set with known RTs in both source and target chromatographic methods (≥10 molecules, recommended for projection)

Outputs

• Ranked annotation candidate table (CSV or similar format)
• Per-candidate RT prediction confidence intervals (lower and upper bounds)
• Per-candidate annotation scores (e.g., 0–1 scale based on RT agreement)
• Per-candidate RT prediction errors (MAE, median absolute error in seconds)
• Filtered candidate subset (candidates with observed RT outside confidence interval removed or flagged)

How to apply

Load the pre-trained DNN model and compute predicted retention times for each candidate metabolite using its molecular structure (fingerprints: MACCS166, Extended Connectivity, Path Fingerprints). For each candidate, compute the RT prediction confidence interval (uncertainty bounds). If projecting between chromatographic methods, apply the meta-learned Bayesian GP prior (trained on ≥10 calibration molecules from the source method) to map observed RTs to the target method's scale. Score each candidate by checking whether the observed RT falls within the predicted confidence interval; candidates within bounds receive high scores (ranked first), while those outside bounds are filtered or ranked lower. Output a ranked table of annotation candidates sorted by descending score (highest confidence first), with annotation scores and RT prediction errors (MAE, MedAE) for each candidate.

Related tools

• alvaDesc (Generates MACCS166, Extended Connectivity, and Path fingerprints required as input features to the DNN retention time predictor) — https://www.alvascience.com/alvadesc/
• cmmrt (CMM-RT) (Reference implementation of the DNN RT predictor, meta-learned Bayesian GP for RT projection, and scoring workflow from the paper) — 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.rt import predict_and_score; candidates = predict_and_score(structures=['C1=CC=C(C=C1)C(=O)O', ...], observed_rt=[125.4, 128.1], model_path='saved_models/dnn.pkl')"

Evaluation signals

• Ranked candidates with observed RT within predicted confidence interval receive higher scores than those outside bounds.
• Mean absolute error (MAE) of projected RTs matches paper benchmark (39.2±1.2 s) when evaluated on held-out test set.
• Median absolute error (MedAE) of projected RTs is ≤17.2±0.9 s on METLIN SMRT dataset or comparable external validation set.
• When meta-learning on ≥10 calibration molecules, RT projection error remains competitive with baseline methods (linear regression, standard ML regressors).
• Filtered candidates (removed because observed RT fell outside confidence interval) are confirmed to be true negatives or low-probability matches in independent validation.

Limitations

• DNN generalization depends on feature representation (fingerprints vs. descriptors); fingerprints outperform descriptors alone, but cross-method and cross-platform transfer requires retraining or fine-tuning.
• Meta-learned RT projection is most reliable with ≥10 calibration molecules; fewer molecules may yield unreliable uncertainty estimates.
• Confidence intervals assume the DNN uncertainty estimates are well-calibrated; calibration should be validated on the target chromatographic method before filtering or ranking in production.
• The method requires alvaDesc (proprietary, under license) for fingerprint generation; open-source alternatives may degrade performance.
• Both retained and unretained molecules are included in the METLIN SMRT training set; RT predictions for unretained compounds may have higher error or different uncertainty properties.

Evidence

• [other] Score candidate metabolites by computing RT prediction confidence intervals and comparing observed versus predicted RT values, filtering candidates where observed RT falls within the projected uncertainty bounds.: "Score candidate metabolites by computing RT prediction confidence intervals and comparing observed versus predicted RT values, filtering candidates where observed RT falls within the projected"
• [readme] A heavily regularized DNN trained with cosine annealing warm restarts and stochastic weight averaging achieved 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] A Bayesian meta-learning approach enables RT projection between chromatographic methods using as few as 10 molecules with competitive error rates: "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."
• [readme] 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."
• [readme] Results suggest that fingerprints tend to perform better than descriptors alone or in combination: "Results suggest that fingerprints tend to perform better."
• [other] Load molecular structures and their corresponding predicted retention times from the pre-trained DNN model (mean absolute error 39.2±1.2 s): "Load molecular structures and their corresponding predicted retention times from the pre-trained DNN model (mean absolute error 39.2±1.2 s)."