PULSE framework generates metabolomic and proteomic profiles from routine blood tests with a new architecture for longitudinal clinical data

Precision medicine applications face a structural hurdle in the high cost and operational complexity of advanced biological measurements. While deep phenotyping through high-throughput proteomics and metabolomics offers granular insights into an individual's internal physiological dynamics, these assays remain largely confined to research settings and well-funded institutions. In daily practice, healthcare systems rely on sparse routine laboratory tests that capture only narrow facets of a patient's biological state. To bypass this constraint, researchers have introduced a self-supervised learning framework called the Patient-Unified Longitudinal Signal Engine (PULSE) in a study published in Nature Computational Science. The model is designed to process multimodal, longitudinal clinical data by drawing on a patient's physiological history to reconstruct missing complex measurements.

The framework builds on the premise that diverse clinical modalities represent biologically coupled, time-varying readouts of an evolving latent physiological state, rather than independent targets for direct cross-assay translation. Clinical data differs fundamentally from single-cell data, which typically captures isolated snapshots. Patients undergo repeated testing across multiple visits, and diagnostic value lies in evaluating current test results against the patient's own baseline and history of prior visits.Sound clinical reasoning does not evaluate a test result in isolation, but interprets it within the patient's individual trajectory and changes over time.

The model encodes the patient's prior state from the preceding visit through a dedicated branch, then merges it with measurements available at the current visit within a shared latent space using flexible distributional alignment. This alignment ensures physiological consistency across modalities and visits without imposing rigid constraints on the representation space, consolidating signals into a unified visit state used to decode and synthesize unmeasured modalities. During training, a random modality-masking strategy forced the model to infer masked data from remaining measurements and historical context, using training objectives that minimize distortions from missing-data patterns before passing the visit state through a temporal module representing cumulative physiological transitions.

The framework proved effective when evaluated on data from approximately 10,000 participants in the UK Biobank, where patient-level samples were split into a training set of 7,000 individuals and an independent test set of 3,000. Relying solely on 61 routine laboratory tests and historical visit context, the model reconstructed complete metabolic profiles of 251 biomarkers, statistically outperforming benchmark models adapted from single-cell research. Furthermore, the framework successfully integrated fundus images and electronic health records. Downstream models trained on the synthetically generated proteomic profiles achieved area under the receiver operating characteristic curve values between 0.72 and 0.83 in predicting six common diseases, performance comparable to models trained on true laboratory-measured proteomic data, alongside predicting 30-day intensive care readmissions across multiple clinical datasets.

This computational advance has direct implications for hospitals, clinical laboratories, and healthcare AI developers across the Gulf, Egypt, and the Levant. Regional health systems hold millions of cumulative electronic health records and routine lab panels from patients visiting clinics over years, whereas the cost of comprehensive proteomic and metabolomic testing remains a financial barrier to widespread adoption. This approach allows clinical and technical teams to extract deep phenotyping and early predictions for complex diseases from existing routine testing infrastructure, without expanding budgets for advanced molecular assays.Historical laboratory data becomes a diagnostic asset that algorithms can unlock.

This shift calls on health data and innovation leaders in the region to reassess longitudinal clinical record strategies, as tracking a patient's trajectory over time and linking historical panels to current tests becomes more valuable than holding an isolated snapshot. Investing in data infrastructure that supports longitudinal modeling is the practical step toward converting standard laboratory tests into advanced biomarkers that inform proactive clinical decisions.