Metabolite candidate ranking by confidence

metabolite-candidate-ranking-by-confidence

Summary

Rank and filter metabolite annotation candidates by computing retention time (RT) prediction confidence intervals and comparing observed versus predicted RT values using a Bayesian meta-learned approach. This skill integrates predicted RTs from a deep neural network with meta-learned chromatographic method projections to produce scored candidate lists suitable for prioritizing follow-up identification work.

When to use

You have a set of candidate metabolites for an unknown compound detected in a liquid chromatography–mass spectrometry (LC-MS) experiment, predicted RTs from a trained DNN model, and access to calibration molecules (minimum 10) that connect your observed chromatographic method to a source method with known RT predictions. Use this skill to rank candidates by confidence and filter out those whose observed RT falls outside the projected uncertainty bounds.

When NOT to use

• Calibration molecule set is smaller than 10 molecules; meta-learned RT projection requires minimum 10 molecules to achieve competitive error rates.
• No trained DNN model is available for the target LC-MS method; you must first train or source a pre-trained predictor on a representative dataset (e.g., METLIN SMRT with 80,038 experimental RTs).
• Observed retention time is not measured on the same chromatographic method as the candidate predictions; RT projections are method-specific and cannot reliably extrapolate across incompatible gradient systems or stationary phases.

Inputs

• molecular structures (SMILES or SDF format) of candidate metabolites
• predicted retention times from pre-trained DNN model for candidates
• observed retention time from LC-MS experiment for unknown compound
• calibration molecule set with experimental RTs in both source and target chromatographic methods (≥10 molecules)
• meta-learned Gaussian Process model trained on PredRet database or local reference CMs

Outputs

• ranked candidate metabolite table with annotation scores
• RT prediction errors and confidence intervals for each candidate
• filtered candidate list (candidates within projected RT uncertainty bounds)
• metadata on projection quality (number of calibration molecules used, GP posterior variance)

How to apply

First, load molecular structures and their DNN-predicted retention times (mean absolute error 39.2±1.2 s) for all candidate metabolites. Apply meta-learned Bayesian RT projection to map retention times between source and target chromatographic methods using your calibration molecules. For each candidate, compute a RT prediction confidence interval by fitting a Gaussian Process prior trained on the calibration set. Score each candidate by testing whether the observed RT from your LC-MS experiment falls within the projected uncertainty bounds. Rank annotation candidates in descending order by confidence score (highest confidence first), filtering out candidates where observed RT falls outside the bounds. Output a ranked candidate table with annotation scores, RT prediction errors, and confidence intervals to guide manual validation or downstream MS/MS matching.

Related tools

• alvaDesc (Generate molecular descriptors (5,666) and fingerprints (2,214 including MACCS166, Extended Connectivity, Path Fingerprints) used as input features for the DNN retention time predictor) — https://www.alvascience.com/alvadesc/
• cmmrt (Python package implementing the DNN retention time predictor, meta-learned Bayesian RT projection, and candidate ranking workflow; includes pre-trained models and Jupyter notebooks for reproducible integration into metabolite annotation pipelines) — https://github.com/constantino-garcia/cmmrt

Examples

python cmmrt/rt/predict.py --candidates candidates.csv --observed_rt 45.3 --calibration_mols calibration.csv --model saved_models/dnn_blender.pkl --meta_gp saved_models/meta_gp.pkl --output ranked_candidates.csv

Evaluation signals

• Ranked candidates have RT prediction errors within reported DNN performance envelope (mean absolute error 39.2±1.2 s, median absolute error 17.2 ± 0.9 s).
• Confidence intervals for RT projections narrow with increasing number of calibration molecules; variance should decrease monotonically from 10 to ~50 calibration molecules.
• Observed RT of true metabolite falls within top-ranked candidate's projected confidence interval; false positive candidates are filtered (observed RT outside bounds).
• Ranking order is consistent with descending Gaussian Process posterior variance or annotation score; ties or inversions indicate potential GP fitting issues or edge cases near projection boundaries.
• Cross-validation on held-out chromatographic methods (e.g., FEM long, LIFE old, FEM orbitrap plasma, RIKEN) shows that meta-training on other CMs produces error rates comparable to single-method models.

Limitations

• Requires a pre-trained DNN model and meta-learned Gaussian Process; the approach is not applicable to entirely novel chromatographic methods without calibration data or transfer learning.
• Minimum 10 calibration molecules required for competitive RT projection error rates; sparse calibration scenarios may yield wide confidence intervals that fail to filter false candidates effectively.
• RT prediction accuracy (MAE 39.2 s) means some candidates will have overlapping projected RT windows; this skill ranks but cannot definitively disambiguate candidates with similar predicted retention times without additional orthogonal information (e.g., MS/MS spectra).
• Molecular fingerprints (MACCS166, ECFP, PFP) outperform descriptor-only models; if fingerprints cannot be computed (e.g., due to licensing constraints on alvaDesc), performance degrades substantially.
• Method-specificity of projections means the skill cannot be applied across incompatible chromatographic systems; source and target methods must share sufficient physicochemical correlation for the GP prior to generalize.

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"
• [other] A Bayesian meta-learning approach enables RT projection between chromatographic methods from as few as 10 molecules while obtaining competitive error rates compared with previous approaches.: "A Bayesian meta-learning approach enables RT projection between chromatographic methods from as few as 10 molecules while obtaining competitive error rates"
• [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)."
• [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: "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"
• [readme] We illustrate how the proposed DNN+meta-learned projections can be integrated into a metabolite annotation workflow.: "We illustrate how the proposed DNN+meta-learned projections can be integrated into a metabolite annotation workflow."