AI for DNA, Chemicals and Microbiome

The vast majority of tandem mass spectra acquired in metabolomics experiments remain unannotated. We present DreaMS, a transformer-based neural network pre-trained in a self-supervised way on millions of unannotated mass spectra from public repositories. Through masked peak prediction and retention order learning, DreaMS acquires rich representations of molecular structures without relying on annotated spectral libraries or domain expertise. We show how fine-tuning DreaMS yields state-of-the-art performance across diverse tasks including spectral similarity, molecular fingerprint prediction, and fluorine detection. We then present five new methods built on DreaMS — targeting the discovery of plant fluorinated natural products, environmental PFAS screening, novel metabolite prioritization, billion-scale spectral search, and general-purpose structure identification. Finally, we introduce DreaMS Atlas, a molecular network of 200 million mass spectra connecting compounds across thousands of independent studies.

TBA

Proteomic profiles offer powerful predictors of disease risk but remain costly and scarce in population cohorts, limiting their translational impact. Metabolomic data, while widely available, carry substantial pre-analytical noise from fasting status, circadian timing, and recent dietary intake, which obscures stable disease-relevant signals. Here, we present AugMent, a contrastive alignment framework that uses the plasma proteome as a biological supervisory signal to distill the most stable and disease-informative components of the metabolome. While the framework supports various architectures, a CLIP-based contrastive implementation yielded the highest fidelity in the UK Biobank. When these distilled representations were used in standard Cox proportional hazards models, they improved prediction across 82 diseases. Mechanistic analysis reveals that the framework preferentially captures latent relationships involving lipoprotein remodeling and lipid transport, particularly pathways where proteins and metabolites are tightly coupled, adding predictive signals to cardiometabolic and related diseases. Together, our results demonstrate that proteome-guided distillation can unlock disease-relevant signals already present in the metabolome but otherwise masked by noise.

Foundation models like Geneformer and scGPT have shown impressive performance on biological prediction tasks, but a prediction without a mechanism is just a correlation. Ihor will present work applying mechanistic interpretability, a discipline from technical AI safety research, to open the hood of these models and ask what they have actually learned about biology. We find that bio AI models build remarkably organized internal representations: protein-protein interaction networks, pathway membership, functional modules, and subcellular localization, all encoded in structured geometric arrangements that mirror known cellular organization. But we also find an important limitation: often these models encode co-expression and pathway structure, not causal regulatory logic. They know which genes go together, but not which gene controls which.
Ihor will also present the discovery and extraction of a compact developmental algorithm from inside scGPT: a representation of blood cell differentiation that we surgically removed from the model as a standalone tool roughly a thousand times smaller, faster, and competitive with established bioinformatics methods. This points toward a new way of creating bioinformatics algorithms, not by designing them, but by extracting them from foundation model internals. Ihor will close by discussing where biological foundation models can be trusted, where they should not be, and how interpretability bridges the gap between black-box AI and the mechanistic understanding that biology demands.

Graph Neural Networks (GNNs) are powerful black-box models for biological data, as they provide a robust way to model complex structural relationships, yet their clinical adoption is hindered by a lack of trustworthy explanations. Current GNN explainability methods provide instance-specific attributions but fail to reveal the global, recurring concepts a model has learned. We propose Graph Explainable Attribution (GEA): a new framework that reframes GNN explanation as a dictionary learning problem. By training a sparse autoencoder on a GNN's graph-, node-, and edge- level embeddings, we decompose dense representations into a sparse combination of interpretable features. Each feature in this learned dictionary corresponds to a fundamental biological motif the GNN uses for prediction. Unlike existing XAI methods that offer fragmented, local views, GEA provides a holistic, dataset-wide perspective that captures systemic model logic. We demonstrate the utility of GEA through use cases on Inflammatory Bowel Disease patient-specific gene co-expression networks, where GEA successfully recovers known biological motifs and uncovers the underlying importance of specific subgraphs, node configurations, and edge patterns that drive the model’s decisions. This approach provides a structured, decomposable view of the GNN's internal logic and a new paradigm for creating more reliable and scientifically valuable explanations.

Despite growing interest in applying machine learning to tandem mass spectrometry, progress has been hindered by the lack of standardized datasets and evaluation protocols. We present MassSpecGym, the first comprehensive benchmark for molecular discovery from MS/MS data, providing 231,000 high-quality labeled spectra across 29,000 molecules and defining three challenges: de novo structure generation, molecule retrieval, and spectrum simulation. We discuss key design decisions, including a novel molecular edit distance-based data split that resolves data leakage issues pervasive in prior work. We also present the first systematic evaluation audit of the MassSpecGym ecosystem, identifying recurring failure modes in recent works—including data contamination, reward hacking, and metric implementation divergence—that materially affect benchmarking conclusions, and release MassSpecGym v1.5 with more robust evaluation conditions.

Recently we have discovered that the human body is full of bacteria and viruses that protect us from disease and such microbes vastly outnumber the ones that cause disease. Analysis of metagenomic data can help us to detect new bacteria and viruses that were not known before, and if they protect from disease, they can be administered to prevent at treat diseases instead of drugs. Discovery of novel microbes in metagenomic data requires approaches that are able to detect microbial dark matter. We previously used protein sequence alignments to define families of viruses that were not found in public databases. These viral families were later found to protect children from developing asthma. However, protein alignments are not quite sensitive enough to bridge the huge diversity viruses on earth, so many viruses are still unclustered singletons. The protein language models that have been developed in recent years are extremely sensitive, but we still do not have the tools to integrate them into existing microbial classification tasks. Here, we applied protein language models for classification of distant viral taxa, noting that they vastly outperformed traditional bioinformatics workflows and even manual curation. However, interpretability remains a challenge, and there is a huge need to assign biological meaning to embedding dimensions.

TBA