FLARE: Fine-grained alignment between spectra and molecular structure improves metabolite annotation. FLARE learns fine-grained correspondences between spectral features and molecular substructures, enabling more precise identification of candidate molecules. By modeling detailed structure–spectra relationships, FLARE improves ranking accuracy and provides interpretable connections between spectral evidence and molecular structure.

GIF: Simulating fragmentation processes using general intelligence models. GIF models fragmentation as a reasoning process, generating peak-labeled spectra from molecular structure using large language model-based reasoning. This approach enables simulation of fragmentation pathways, improves understanding of spectra–structure relationships, and provides a foundation for interpretable metabolite annotation.

Fine-grained structural classification connects biosynthetic gene clusters to molecular function. This work introduces methods to predict detailed structural classes of natural products directly from biosynthetic gene cluster sequences. Fine-grained classification enables linking genomic information to molecular structure, improving discovery, annotation, and functional interpretation of microbial secondary metabolism.

MVP: Learning from all views to improve metabolite annotation with multiview contrastive learning. MVP leverages multiple complementary “views” of each metabolite (e.g., spectra-derived signals and molecular representations) and learns to align them in a shared embedding space, and produce improved candidate ranking.

OMG: De Novo molecular generation from spectra via transfer learning and curriculum learning. We can use state-of-the-art generators based on molecular formula to generate candidates, and then use state-of-the-art rankers to rank.

MADGEN: Generating de novo molecules from partial structures, guided by spectra. Starting from scratch is hard. We start from a scaffold and complete a molecular graph guided by the query spectrum.

Metabolite annotation using a novel paradigm. Avoid the explicit construction of spectra and molecules. Instead, compare query spectrum against candidates in the embedding space to identify the best matching molecule. Regulization with candidate molecules shows great value.

Benchmarking for metabolite annotation: MassSpecGym, a community effort to standarize the evaluation of mass spec annotation techniques.

Learning to rank for metabolite annotation for metabolomics. Ensemble Spectral Prediction (ESP) impelments a novel pipleine that includes: molecular representation learning, spectral prediction with peak co-dependency analysis, and rank-based learning. Performance is improved up to 41% over MLPs.

Separate normalization of normal and [CLS] tokens in self-supervised transformers. Transformer models typically utilize a single normalization layer for both the class token [CLS] and normal tokens. We show a 2.7% performance improvement in image, natural, and graph tasks when utilizing separate normalization layers.

Enzyme-substrate interaction prediction using contrastive multiview coding (CMC).
Two enzymes and a molecule; two molecules and an enzyme.
These are two different views of an interaction. We stratify data and use CMC to outperform all known interaction prediction methods.

What's new in computational methods for Metabolomics?
Review article on novel methods for mass spec analysis with colleauges from the 2022 Dagstuhl seminar, “Computational Metabolomics: From Spectra to Knowledge”.

Predicting site-of-metabolism for enzymatic reactions. Molecules are graphs. We can use GNNs to build SOM predictors for all classes of enzymes! Results demonstrate benefits to enhancing rule-based biotransformation prediction methods.

Recommender Systems to recommend enzymes to molecules and vice versa. Boost-RS is a general recommender system that utilizes relational and group auxiliary data to boost learned representations. Boost-RS's versatility is demonstrated for matching molecules and enzymes.

Graph Matching and molecular alignment. A novel deep learning method for aligning two graphs, with application to molecular matching. Useful for knowing where two molecules match and differ.

Enzyme Classification on Molecules. A new method for training enzyme-specific predictors that take as input a given query substrate molecule and return whether the enzyme would act on that substrate or not.

Metabolic Disruption analysis in engineered hosts. Engineering cellular hosts may result in unexpected and undesirable byproducts due to promiscuous interactions of native/heterelogous enzymes and molecules. These effects are not only disruptive to the host metabolism but also to the intended end-objective of high yield. How do you analyze such disruptions? MDFlow is the answer!

Enzymatic Link Prediction. Learning graph representations of biochemical networks and its application to predicting enzymatic links between two molecules.

Probabilistic Analysis of Pathway Activities and Metabolite Annotations. Using inference, we learn the likelihood of metabolic pathways being responsible for presence of metabolomics measurements, and the likelihood of annotations.

Extended Metabolic Models (EMMs). This workflow utilizes our tool PROXIMAL to create Extended Metabolic Models (EMMs) that contain not only canonical substrates and products of enzymes already cataloged for an organism, but also metabolites that can form due to substrate promiscuity. We created an EMM for E. coli and to analyze metabolomics data for CHO and murine cecal microbiota.

Machine learning analysis pipeline for antibody sequences. Important features in antibody sequences are identified relative to a reference set using machine learning and statistical analysis.

ProPASS. A workflow that links synthesis pathway construction with the exploration of available enzyme sequences that are predicted soluble in the host. ProSol DB is a database cataloging the predicted solubility of over 250,000 reviewed enzyme sequences from UNIPROT.

gEFM. A method for computing elementary flux modes.

PROXIMAL. A method to derive biotransformation operators from a reaction and to apply them to a target molecule. Operators can be applied to predict xenobiotic products (as suggested in the paper), or applied more generically to identify derivative metabolites for a specific query molecule and a specified enzyme.