How single-cell metabolomics and AI could transform toxicity prediction and drug discovery

Metabolomics offers a uniquely direct view of cellular function. While genomics can indicate what a cell is capable of doing and transcriptomics can reveal changes in gene expression, metabolomics captures the downstream molecular activity that more closely reflects what a cell is doing at a given moment.

For Theodore Alexandrov, CEO and co-founder of DeepCyte, this makes metabolomics particularly valuable for understanding how cells respond to drugs, toxicants, and changes in their environment. By combining single-cell metabolomics with artificial intelligence, DeepCyte is working to turn these functional molecular measurements into predictive models for toxicity, drug safety, and mechanisms of action.

Alexandrov’s work in this field spans more than a decade. Before founding DeepCyte, his academic team developed SpaceM, a family of methods for single-cell metabolomics, and METASPACE, a global cloud platform for spatial metabolomics. He also co-founded SCiLS, a scientific software company later acquired by Bruker.

Today, that background is informing a broader ambition: to move metabolomics from a technology primarily used to describe biological states toward one capable of helping predict cellular responses before they occur.

Why metabolomics provides a functional view of cellular biology

Among the different omics technologies available to researchers, Alexandrov sees metabolomics as particularly close to phenotype and cellular function.

“While genomics tells you what a cell could do and transcriptomics what it plans to do, metabolomics shows you what a cell is actually doing,” he explains.

That distinction is important because metabolic changes can happen rapidly. When an immune cell is activated, for example, its metabolism can be rewired within minutes. Measuring metabolites can therefore provide a direct readout of these functional changes.

Metabolomics can also reveal biology extending beyond classical metabolic pathways. Metabolites are increasingly understood to have signaling and regulatory roles, meaning that metabolic measurements can provide insight into broader molecular programs operating within a cell. This functional dimension makes metabolomics an important complement to genomics, transcriptomics, and proteomics, rather than simply another parallel layer of molecular information.

How single-cell metabolomics reveals cellular heterogeneity

The shift from bulk to single-cell analysis adds another level of biological resolution. Bulk metabolomics produces an average measurement across many cells. While useful, this can obscure distinct cellular states and subpopulations that may respond differently to the same biological stimulus or chemical exposure. Single-cell metabolomics instead makes it possible to observe molecular variation from cell to cell.

“It captures the diversity of molecular programs manifested across cell subpopulations — details that are invisible to bulk methods, which average everything out,” says Alexandrov.

This can be especially important in complex biological systems where a relatively small population of cells may behave very differently from the majority. From a technology perspective, Alexandrov also highlights practical advantages of single-cell metabolomics, including relatively simple sample preparation, high throughput, and affordability compared with some other single-cell omics approaches. Together, these characteristics could help make metabolomics increasingly valuable for applications in drug discovery and toxicology, where understanding heterogeneous cellular responses is often critical.

Why spatial metabolomics matters for understanding cells in tissues

Single-cell resolution alone does not capture the full biological context. For cells within intact tissues, location can profoundly influence molecular state and function. Alexandrov argues that preserving this spatial context is particularly important in metabolomics because cellular metabolism can change extremely quickly in response to stress or sample handling. Methods that require tissue dissociation or homogenization can therefore create two problems: they remove information about where cells were located, and they can alter the molecular state researchers are trying to measure.

“Metabolic states of cells can shift within seconds under stress, and dissociation can release intracellular metabolites, collapsing gradients and signals that define cell states in situ,” Alexandrov explains.

The result is that a measurement obtained after extensive sample preparation may not accurately represent what a cell was doing inside the original tissue. Minimizing these perturbations is therefore essential. Approaches such as flash-freezing tissues or preparing cells in ways that preserve metabolic state can help retain biologically meaningful information.

Spatial metabolomics also allows researchers to connect molecular profiles with cell-cell interactions, tissue architecture, and the local microenvironment. These relationships can be particularly important when interpreting how metabolism changes in disease or following drug exposure.

Combining single-cell metabolomics and AI for predictive toxicology

The information density of metabolomics creates both an opportunity and a challenge. Metabolic profiles reflect the combined output of many upstream molecular pathways, producing high-dimensional datasets rich in functional information. At the same time, metabolism can be difficult to interpret. Individual metabolites may participate in several biological pathways, and many metabolic processes remain incompletely understood. This is where Alexandrov sees artificial intelligence becoming particularly powerful.

“When applied to a tractable problem and supplied with sufficient data and well-designed evaluation, AI can capitalize on the benefits of metabolomics even without full mechanistic understanding of metabolism,” he says.

Rather than requiring every metabolic interaction to be understood before useful predictions can be made, AI models can learn patterns within large datasets that are associated with specific biological outcomes. Alexandrov describes this as the possibility of AI functioning as an “oracle”: using metabolomics to address biological or medical questions even before researchers have completely decoded the mechanisms responsible for the observed signal. At DeepCyte, this concept is already being applied to toxicity prediction.

How DeepCyte uses metabolomics to predict drug toxicity and mechanism

DeepCyte was founded around the idea that established approaches to toxicity testing and prediction are not sufficient on their own. The company is using large, curated, and experimentally designed single-cell metabolomics datasets as a functional input for machine-learning models. The objective is not only to predict whether a compound produces toxicity, but also to identify the mechanisms underlying that response.

According to Alexandrov, DeepCyte has already generated proof points demonstrating that its machine-learning framework applied to single-cell metabolomics can outperform structure-based toxicity prediction. The company has also found that the approach can identify mechanisms of toxicity that structure-activity methods fail to detect.

“This is the advantage of measuring what cells actually do in response to a compound rather than inferring toxicity from its molecular structure alone,” he says.

The distinction is central to DeepCyte’s strategy. Structure-based approaches infer biological effects from chemical characteristics, whereas metabolomics provides a direct measurement of the cellular response after exposure. For Alexandrov, combining this functional information with AI creates an opportunity to move toxicology towards a more predictive model.

Why metabolomics adds value to multiomics studies

Metabolomics also has a broader role within multiomics research. By adding a functional molecular layer alongside genomic, transcriptomic, and proteomic information, metabolomics can reveal aspects of biology that are difficult to access using other technologies alone.

Research has already highlighted signaling and regulatory functions of metabolites, as well as interactions between metabolites and macromolecular complexes. While its relevance to metabolic diseases is intuitive, Alexandrov argues that metabolomics is proving useful far beyond disorders traditionally classified as metabolic.

He points to large-scale research from the Nightingale group showing that metabolomic profiles can provide predictive information across a wide range of diseases, including conditions such as depressive disorders and chronic obstructive pulmonary disease. For Alexandrov, these findings raise a broader question for biomedical research and drug development: how much clinically relevant information may be missed when metabolomics is absent from multiomics studies?

The implication is that metabolomics may not simply provide additional detail about metabolism itself. It may capture functional signatures associated with complex biology that are difficult to predict from genetic information alone.

Moving drug safety from descriptive biology to prediction

Much of traditional molecular profiling is fundamentally descriptive. Researchers expose cells or tissues to a compound, measure the resulting changes, and then attempt to understand what has happened. DeepCyte is aiming to move beyond that paradigm. By training machine-learning models on carefully designed single-cell metabolomics datasets, the company is working towards systems capable of predicting toxicity and mechanisms of action from cellular response patterns.

Although DeepCyte has only recently emerged from stealth, Alexandrov believes the underlying approach demonstrates how AI can help overcome some of metabolomics’ historical limitations. The aim is to capitalize on the depth of information contained within metabolic measurements without requiring researchers to fully interpret every molecular interaction before those data become useful. In this model, metabolomics becomes not only a tool for characterizing biology after an event but also a foundation for anticipating what is likely to happen next.

Digital cell models could predict responses before experiments begin

The longer-term trajectory extends beyond toxicity classifiers. Advances in single-cell omics and artificial intelligence are increasingly enabling the development of predictive digital models of cells. For Alexandrov, this means models capable of forecasting how a cell’s metabolic state will change in response to a drug or toxicant before the corresponding experiment is performed.

“The discussion in the field now is not about feasibility but rather about how such models should be used and validated,” he says.

Pharmacology and toxicology may provide especially suitable applications for these technologies because both fields address clearly defined, experimentally testable questions. Predictive cell models could ultimately help researchers identify promising or problematic compounds earlier in the drug development process, improve safety assessment, and reduce dependence on animal testing.

“The goal is clear: to predict more efficacious and safer drugs faster, cheaper, and with less reliance on animal testing,” Alexandrov says.

For DeepCyte, the path towards that goal begins with generating high-quality single-cell metabolomics data and using those measurements to build increasingly capable predictive systems. The eventual destination is a model of drug discovery and toxicology in which researchers can move from observing cellular responses experimentally to predicting them computationally. As single-cell metabolomics and AI continue to converge, that shift could fundamentally change how researchers assess drug safety, investigate toxicity mechanisms, and understand cellular function.

Want the latest science news straight to your inbox? Become a SelectScience member for free today>>