Sebastian Böcker

Liquid chromatography is a predominant technology for the separation of small molecules. Hundreds of retention time prediction models have been published over the past decades, yet truly transferable prediction (requiring no training data from the target chromatographic system) remains an open challenge. Unfortunately, retention times may change massively, even for nominally identical chromatographic conditions. Retention order is considerably more conserved; but even retention order may change if chromatographic conditions vary. Here we present 2-step, a two-step method for the prediction of retention times in reversed-phase chromatography. In the first step, a machine-learning model predicts a retention order index, taking into account chromatographic conditions. In the second step, we map predicted indices to absolute retention times. Disentangling these two tasks finally enables transferable retention time prediction across chromatographic conditions and compound classes, without requiring any target-system training data. Our 2-step method outperforms existing methods that were trained on the target dataset. Finally, we systematically study what chromatographic conditions result in notable changes of retention order.

Affinity-selection platforms are powerful tools in early drug discovery, but current technologies – most notably DNA-encoded libraries (DELs) – are limited by synthesis complexity and incompatibility with nucleic acid-binding targets. We present a barcode-free self-encoded library (SEL) platform that enables direct screening of over half a million small molecules in a single experiment. SELs combine tandem mass spectrometry with custom software for automated structure annotation, eliminating the need for external tags for the identification of screening hits. We develop efficient, high-diversity synthesis protocols for a broad range of chemical scaffolds and benchmark the platform in affinity selections against carbonic anhydrase IX, identifying multiple nanomolar binders. We further apply SELs to flap endonuclease 1 (FEN1) – a disease related DNA-processing enzyme inaccessible to DELs – and discover potent inhibitors. Taken together, screening barcode-free libraries of this scale all at once represents an important development, enables access to novel target classes, and promises substantial impact on both academic and industrial early drug discovery.

Hundreds of models for the prediction of small molecule retention times were published during the last decades. Our goal is the transferable prediction of retention times: Our method should predict retention times for a target dataset, without the need of training data from that chromatographic system. Unfortunately, retention times may change massively, even for nominally identical chromatographic conditions. Retention order is much better retained, yet even the retention order of compounds may change if chromatographic conditions vary. We present a machine learning model that can predict retention order or, more precisely, a retention order index, taking into account chromatographic conditions. We show how to map predicted retention order indices to retention times. Disentangling these two task finally enables transferable retention time prediction across chromatographic conditions and compound classes. Our 2-step method outperforms existing methods that were trained on the target dataset. Finally, we systematically study what chromatographic conditions result in notable changes of retention order.

Metabolite identification in non-targeted mass spectrometry-based metabolomics remains a major challenge due to limited spectral library coverage and difficulties in predicting metabolite fragmentation patterns. Here, we introduce Multiplexed Chemical Metabolomics (MCheM), which employs orthogonal post-column derivatization reactions integrated into a unified mass spectrometry data framework. MCheM generates orthogonal structural information that substantially improves metabolite annotation through in silico spectrum matching and open-modification searches, offering a powerful new toolbox for the structure elucidation of unknown metabolites at scale.

Affinity-selection platforms are powerful tools in early drug discovery, but current technologies – most notably DNA-encoded libraries (DELs) – are limited by synthesis complexity and incompatibility with nucleic acid-binding targets. We present a barcode-free self-encoded library (SEL) platform that enables direct screening of ~1 million small molecules in a single experiment. SELs combine tandem mass spectrometry with custom software for automated structure annotation, eliminating the need for external tags for the identification of screening hits. We developed efficient, high-diversity synthesis protocols for a broad range of chemical scaffolds and benchmarked the platform in affinity selections against carbonic anhydrase IX, identifying multiple nanomolar binders. We further applied SELs to flap endonuclease 1 (FEN1) – a disease related DNA-processing enzyme inaccessible to DELs – and discovered potent inhibitors. Taken together, screening barcode-free libraries of this scale is unprecedented, enables access to novel target classes, and promises substantial impact on both academic and industrial early drug discovery.

Untargeted metabolomics can comprehensively map the chemical space of a biome, but is limited by low annotation rates (<10%). We used chemistry-based vectors, consisting of molecular fingerprints or chemical compound classes, predicted from mass spectrometry data, to characterize compounds and samples. These chemical characteristics vectors (CCVs) estimate the fraction of compounds with specific chemical properties in a sample. Unlike the aligned MS1 data with intensity information, CCVs incorporate actual chemical properties of compounds, offering deeper insights into sample comparisons. Thus, we identified key compound classes differentiating biomes, such as ethers which are enriched in environmental biomes, while steroids enriched in animal host-related biomes. In biomes with greater variability, CCVs revealed key clustering compound classes, such as organonitrogen compounds in animal distal gut and lipids in animal secretions. CCVs thus enhance the interpretation of untargeted metabolomic data, providing a quantifiable and generalizable understanding of the chemical space of natural biomes.

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.

Theinability to identify the structures of most metabolites detectedin environmental or biological samples limits the utility of nontargetedmetabolomics. The most widely used analytical approaches combine massspectrometry and machine learning methods to rank candidate structurescontained in large chemical databases. Given the large chemical spacetypically searched, the use of additional orthogonal data may improvethe identification rates and reliability. Here, we present resultsof combining experimental and computational mass and IR spectral datafor high-throughput nontargeted chemical structure identification.Experimental MS/MS and gas-phase IR data for 148 test compounds wereobtained from NIST. Candidate structures for each of the test compoundswere obtained from PubChem (mean = 4444 candidate structures per testcompound). Our workflow used CSI:FingerID to initially score and rankthe candidate structures. The top 1000 ranked candidates were subsequentlyused for IR spectra prediction, scoring, and ranking using densityfunctional theory (DFT-IR). Final ranking of the candidates was basedon a composite score calculated as the average of the CSI:FingerIDand DFT-IR rankings. This approach resulted in the correct identificationof 88 of the 148 test compounds (59%). 129 of the 148 test compounds(87%) were ranked within the top 20 candidates. These identificationrates are the highest yet reported when candidate structures are usedfrom PubChem. Combining experimental and computational MS/MS and IRspectral data is a potentially powerful option for prioritizing candidatesfor final structure verification.