Application Note | August 31, 2026
An AI Foundation Model of Human Metabolic State
Get the Resource
Download NowPatients who share the same diagnosis, driver mutation, or baseline lab value can respond to the same therapy in markedly different ways and progress along distinct trajectories. That heterogeneity is a primary driver of the industry’s most expensive failures: pivotal trials diluted by patients the drug was never going to help, molecules advanced into late-stage development without ever engaging their intended biology, and safety signals identified too late to address. The differences lie in each patient’s physiological state: the integrated, real-time output of genotype, microbiome, diet, exposure, and organ function.
Sapient’s DynamiQ™ metabolic foundation model reads out this dynamic human physiology from a single blood sample, revealing the metabolic state that determines disease trajectory, drug response, and toxicity. It gives drug development teams a powerful tool to enrich trials, discontinue failing programs earlier, and detect safety signals that are hidden to genomic, structural, and clinical record models.
Download the white sheet for an in-depth look at Sapient’s DynamiQ™ metabolic foundation model and its applications for de-risking drug development.
What is a metabolic foundation model, and what insights can it reveal?
A metabolic foundation model leverages AI in large-scale, nontargeted metabolomics datasets to learn a reusable representation of the circulating metabolome: the functional layer that reflects a patient’s dynamic physiology and metabolic state which cannot be comprehensively captured by genomics, routine clinical measures, or singular biomarkers alone.
As the functional integrator in which genotype, microbiome, diet, exposure, and physiology converge, the metabolome provides a measurable readout of a patient’s dynamic, whole-body physiological state rather than one that is static or tissue-bound. The model is trained on metabolomics data from both healthy individuals and patients with varied demographics, diagnoses, and treatments – providing the scale, diversity, and depth of biological data necessary to map human metabolism across broad populations, including the phenotypic commonalities and distinctions that govern this biology over time.
The representation is learned once and can be reused across therapeutic areas, programs, and development phases to answer critical questions that drug development teams are asking each day: which patients to enroll, whether a molecule is engaging its intended biology, and whether a safety liability is emerging.
Sapient's DynamiQ™ metabolic foundation model: trained on our one-of-a-kind dataset
At Sapient, we have developed our own metabolic foundation model trained by self-supervised learning on Sapient’s DynamiQ database, the largest standardized resource of human metabolism measures assembled for this purpose and comprised of:
67K+ human plasma samples
collected longitudinally from more than 13,000 individuals.
15K+ metabolite & lipid features per sample
including thousands of features encoding biochemical activity invisible to other omics.
Matched EHR data & clinical outcomes
spanning more than 60 diseases including cardiometabolic, hepatic, and renal conditions.
The model is able to reconstruct masked metabolites and preserve the covariation structure across pathways, mapping the intrinsic organization of human metabolic biology rather than optimizing for a single labeled outcome. Because the representation is learned from population-scale metabolic structure, it is robust to cohort shift, missing data, and sparsely observed phenotypes, and it captures the latent states that drive differential drug response and disease progression.
The defensibility of the model rests on the data beneath it. Model architectures are commoditizing quickly, but the input is not. Broad, standardized metabolomics at this scale cannot be assembled without a purpose-built analytical platform and years of harmonized data collection, which is exactly what DynamiQ provides.
Use cases for the DynamiQ metabolic foundation nodel in drug development
Historically, patient heterogeneity in a trial has been approached through the lens of diagnosis, genotype, and standard laboratory values. The DynamiQ metabolic foundation model reframes the approach with predictions and measurements that enable:
Patient Stratification and Trial Enrichment
Response to metabolic and immune therapies is itself largely a metabolic phenotype, so a baseline blood signature can separate likely responders from non-responders before enrollment and focus a trial on those patients most likely to benefit from the drug.
The same capability supports prognostic enrichment for event-driven trials. Enriching for a higher-risk population decreases the required trial size, which on a study costing hundreds of millions of dollars can represent tens of million of dollars in savings.
Disease State and Pharmacodynamics
Because mass spectrometry can measure what is present in the blood at a given moment, the model reads out biological state directly and non-invasively. This is particularly beneficial for disease states such as metabolic-associated liver disease, where tissue sampling is difficult.
On treatment, it detects whether a molecule is shifting the metabolome along its intended mechanistic axis, providing a pharmacodynamic and target engagement readout weeks before a clinical endpoint moves. This converts a slow and expensive failure into an early, low-cost go or no-go decision.
On-treatment metabolome shifts can read out target engagement weeks before the clinical endpoint.
Safety and Toxicity Profiling
Human toxicity is a leading cause of program failure but still difficult to predict with standard omics measures. Muscle and lean-mass catabolism, drug-induced liver injury, and nephrotoxic or metabolic signatures appear in circulating biochemistry – often before standard clinical measures register them – and the DynamiQ foundation model captures this chemistry to read out emerging organ injury and metabolic stress.
Continuous metabolic safety monitoring surfaces these signals while they are still a protocol decision in early development, rather than a label restriction or cause for program withdrawal.
Injury and metabolic-stress signatures flag toxicity while it is still a protocol decision.
The DynamiQ differentiator compared to other foundation model classes
Biological foundation models have emerged as a new and powerful asset for pharma and biotech R&D, but the data that trains them defines both their strengths and their limits.
Genomic, RNA-based, and clinical record models capture a static or localized property, such as a sequence, an inherited variant, or a snapshot of a single tissue. None measure the dynamic, whole-body physiological state that separates one patient from another once a program is underway.
That state is carried by the circulating metabolome, which the DynamiQ metabolic foundation model is uniquely built to capture – giving drug development teams a systemic view of each patient rather than one that is static or tissue-bound.
Equipped with this model, teams can probe questions such as: what patients are most likely to respond to a given therapy? Can an outcome trial be enriched to reduce trial size and costs? Is a metabolic safety signal emerging before it becomes a label or franchise risk? The answers align patient selection and trial interpretation with the physiology that actually drives outcomes.
Evaluate our model with your own samples.
To discuss how Sapient can scope a retrospective analysis on a subset of your banked baseline samples from a completed or ongoing program to demonstrate the model’s results, schedule a time to meet with our team.