Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2019 Dec 20;10(1):5811.
doi: 10.1038/s41467-019-13680-7.

The METLIN small molecule dataset for machine learning-based retention time prediction

Xavier Domingo-Almenara  1   2 , Carlos Guijas  3 , Elizabeth Billings  3 , J Rafael Montenegro-Burke  3 , Winnie Uritboonthai  3 , Aries E Aisporna  3 , Emily Chen  4 , H Paul Benton  3 , Gary Siuzdak  5   6
Affiliations

The METLIN small molecule dataset for machine learning-based retention time prediction

Xavier Domingo-Almenara et al. Nat Commun. .

Abstract

Machine learning has been extensively applied in small molecule analysis to predict a wide range of molecular properties and processes including mass spectrometry fragmentation or chromatographic retention time. However, current approaches for retention time prediction lack sufficient accuracy due to limited available experimental data. Here we introduce the METLIN small molecule retention time (SMRT) dataset, an experimentally acquired reverse-phase chromatography retention time dataset covering up to 80,038 small molecules. To demonstrate the utility of this dataset, we deployed a deep learning model for retention time prediction applied to small molecule annotation. Results showed that in 70[Formula: see text] of the cases, the correct molecular identity was ranked among the top 3 candidates based on their predicted retention time. We anticipate that this dataset will enable the community to apply machine learning or first principles strategies to generate better models for retention time prediction.

PubMed Disclaimer

Conflict of interest statement

The authors declare no competing interests.

Figures

Fig. 1
Fig. 1. RT prediction results.
a Composition of the SMRT dataset and structure of the deep-learning model. b Predicted vs experimental RT for the training set and c validation set. Non-retained molecules are indicated (tentatively) in the training set plot. The relative prediction error box plot for the validation set is also shown. The box plot represents median value and interquartile range (25–75% percentiles) excluding outliers.
Fig. 2
Fig. 2. Naïve k-NN approach and DLM approach comparison.
Box plots for the naïve k-NN (N) and DLM (P) relative RT prediction error across similarity thresholds are shown. P values from paired Wilcoxon rank tests are summarized with asterisks (n95 = 648, n90 = 6018, n80 = 14028, n70 = 16974, n50 = 19817) except for the 95% threshold, for which ns (not significant) corresponds to the Wilcoxon rank test whereas the asterisks correspond to the Ansari–Bradley P value. Box plots represent median value and interquartile range (25–75% percentiles) excluding outliers.
Fig. 3
Fig. 3. Prediction performance dependency on the training set.
The prediction error depends on the number and similarity level of molecules in the training set. To assess this dependency, the predicted RT error of molecules in the validation set having at least one similar molecule in the training set (more than 90% of similarity) was compared to molecules in the validation set with no similar molecule in the training set. Box plots represent median value and interquartile range (25–75% percentiles) excluding outliers.
Fig. 4
Fig. 4. RT projection scheme.
a The RT of identified compounds eluting in two chromatographic methods (CMs) can be used to build a projection function (red dashed line). We determined this projection function by comparing predicted RT values (in the CM used to acquire the METLIN’s SMRT dataset) to experimental RT values in a different CM (predicted–experimental projection). b This function enables projecting all the predicted RTs from their specific CM onto other experimental CMs. c By performing an accurate mass search, we can determine all possible known structures that could be attributed to any annotated protonated (M+H) or deprotonated (M−H) ion m/z. Predicted RT allows ranking this putative candidates by prediction error.
Fig. 5
Fig. 5. Predicted–experimental projections results.
Four experimental chromatographic methods (ECM 1 to 4) with five panels each. In the PredRet database, ECMs 1–4 are known as FEM_long, LIFE_old, RIKEN and FEM_orbitrap_plasma, respectively. Panels a, f, k, p show the projection between the experimental and predicted RT. Panels b, g, l, q show relative errors between the experimental and predicted RT. Box plots represent median value and their error bars represent the interquartile range (25–75% percentiles). Panels c, h, m, r show the number of putative candidates before and after filtering by RT error threshold. Panels d, i, n, s show the ROC curves with their respective AUC values. Panels e, j, o, t show the candidate ranking. Ranking panels show the percentage of molecules in which the correct identity is exactly the first, second or third top candidate (y-axis, correct candidate ranking) and there are more than 1, 2 or 3 putative candidates in total (x-axis, total number of candidates). The top-three chart shows the percentage of cases in which the correct identity was ranked among the top three candidates of the total of cases with more than three candidates.
See this image and copyright information in PMC

References

    1. Wu Z, et al. MoleculeNet: a benchmark for molecular machine learning. Chem. Sci. 2018;9:513–530. doi: 10.1039/C7SC02664A. - DOI - PMC - PubMed
    1. Lo Y, Rensi SE, Torng W, Altman RB. Machine learning in chemoinformatics and drug discovery. Drug Discov. Today. 2018;23:1538–1546. doi: 10.1016/j.drudis.2018.05.010. - DOI - PMC - PubMed
    1. Chen H, Engkvist O, Wang Y, Olivecrona M, Blaschke T. The rise of deep learning in drug discovery. Drug Discov. Today. 2018;23:1241–1250. doi: 10.1016/j.drudis.2018.01.039. - DOI - PubMed
    1. Lenselink EB, et al. Beyond the hype: deep neural networks outperform established methods using a ChEMBL bioactivity benchmark set. J Cheminform. 2017;9:45. doi: 10.1186/s13321-017-0232-0. - DOI - PMC - PubMed
    1. Mayr A, Klambauer G, Unterthiner T, Hochreiter S. DeepTox: toxicity prediction using deep learning. Front. Environ. Sci. 2016;3:80. doi: 10.3389/fenvs.2015.00080. - DOI

Publication types

Substances

LinkOut - more resources