Taxonomic classification, that is, the assignment to biological clades with shared ancestry, is a common task in genetics, mainly based on a genome similarity search of large genome databases. The classification quality depends heavily on the database, since representative relatives must be present. Many genomic sequences cannot be classified at all or only with a high misclassification rate. Here we present BERTax, a deep neural network program based on natural language processing to precisely classify the superkingdom and phylum of DNA sequences taxonomically without the need for a known representative relative from a database. We show BERTax to be at least on par with the state-of-the-art approaches when taxonomically similar species are part of the training data. For novel organisms, however, BERTax clearly outperforms any existing approach. Finally, we show that BERTax can also be combined with database approaches to further increase the prediction quality in almost all cases. Since BERTax is not based on similar entries in databases, it allows precise taxonomic classification of a broader range of genomic sequences, thus increasing the overall information gain.

Small molecule machine learning aims to predict chemical, biochemical, or biological properties from molecular structures, with applications such as toxicity prediction, ligand binding, and pharmacokinetics. A recent trend is developing end-to-end models that avoid explicit domain knowledge. These models assume no coverage bias in training and evaluation data, meaning the data are representative of the true distribution. However, the domain of applicability is rarely considered in such models. Here, we investigate how well large-scale datasets cover the space of known biomolecular structures. For doing so, we propose a distance measure based on solving the Maximum Common Edge Subgraph (MCES) problem, which aligns well with chemical similarity. Although this method is computationally hard, we introduce an efficient approach combining Integer Linear Programming and heuristic bounds. Our findings reveal that many widely-used datasets lack uniform coverage of biomolecular structures, limiting the predictive power of models trained on them. We propose two additional methods to assess whether training datasets diverge from known molecular distributions, potentially guiding future dataset creation to improve model performance.

Metabolites provide a direct functional signature of cellular state. Untargeted metabolomics usually relies on mass spectrometry, a technology capable of detecting thousands of compounds in a biological sample. Metabolite annotation is executed using tandem mass spectrometry. Spectral library search is far from comprehensive, and numerous compounds remain unannotated. So-called in silico methods allow us to overcome the restrictions of spectral libraries, by searching in much larger molecular structure databases. Yet, after more than a decade of method development, in silico methods still do not reach the correct annotation rates that users would wish for. Here, we present a novel computational method called Mad Hatter for this task. Mad Hatter combines CSI:FingerID results with information from the searched structure database via a metascore. Compound information includes the melting point, and the number of words in the compound description starting with the letter ‘u’. We then show that Mad Hatter reaches a stunning 97.6% correct annotations when searching PubChem, one of the largest and most comprehensive molecular structure databases. Unfortunately, Mad Hatter is not a real method. Rather, we developed Mad Hatter solely for the purpose of demonstrating common issues in computational method development and evaluation. We explain what evaluation glitches were necessary for Mad Hatter to reach this annotation level, what is wrong with similar metascores in general, and why metascores may screw up not only method evaluations but also the analysis of biological experiments. This paper may serve as an example of problems in the development and evaluation of machine learning models for metabolite annotation.

Small molecule machine learning tries to predict chemical, biochemical or biological properties from the structure of a molecule. Applications include prediction of toxicity, ligand binding or retention time. A recent trend is to develop end-to-end models that avoid the explicit integration of domain knowledge via inductive bias. A central assumption in doing so, is that there is no coverage bias in the training and evaluation data, meaning that these data are a representative subset of the true distribution we want to learn. Usually, the domain of applicability is neither considered nor analyzed for such large-scale end-to-end models. Here, we investigate how well certain large-scale datasets from the field cover the space of all known biomolecular structures. Investigation of coverage requires a sensible distance measure between molecular structures. We use a well-known distance measure based on solving the Maximum Common Edge Subgraph (MCES) problem, which agrees well with the chemical intuition of similarity between compounds. Unfortunately, this computational problem is provably hard, severely restricting the use of the corresponding distance measure in large-scale studies. We introduce an exact approach that combines Integer Linear Programming and intricate heuristic bounds to ensure efficient computations and dependable results. We find that several large-scale datasets frequently used in this domain of machine learning are far from a uniform coverage of known biomolecular structures. This severely confines the predictive power of models trained on this data. Next, we propose two further approaches to check if a training dataset differs substantially from the distribution of known biomolecular structures. On the positive side, our methods may allow creators of large-scale datasets to identify regions in molecular structure space where it is advisable to provide additional training data.

Liquid chromatography is frequently employed for the separation of metabolites and other small molecules. Prediction of retention times via machine learning methods can assist compound annotation. Yet, transferable predictions are intrinsically complicated for novel compounds and novel chromatographic conditions because retention times depend both on compound structure and the employed chromatographic system. We present RepoRT, the first repository for retention time data. RepoRT presently contains 373 datasets, 8809 unique compounds, and 88,325 retention time entries measured on 49 different chromatographic columns using varying eluents, flow rates, and temperatures. We put particular effort on making RepoRT “machine learning-ready”: We performed an extensive manual curation, cleaning and completion of the available data; we developed automated methods for data validation during upload; we collected more than 45,000 different columns of different vendors with their lengths, particle sizes, inner diameters, and pore sizes to create a database with normalized column names; and, we ensured that features required for transferable predictions, such as parameters numerically describing the selectivity of chromatographic columns, are readily available. For version control and reproducible research, RepoRT is hosted on GitHub.