Longitudinal transcriptomic studies frequently capture only a small number of molecular snapshots per patient. These sparse visit schedules make it difficult to reconstruct how individual patients progress over time, to compare patients when their progression rates differ, and to infer molecular changes that occur between observed visits. The work summarized here presents a framework designed to address these challenges by inferring continuous, patient-specific molecular trajectories from such sparse data.
The authors frame the problem as one of reconstructing continuous trajectories and organizing them at the cohort level so that shared progression patterns can be identified. This approach aims to go beyond discrete visit labels and provide a finer-grained progression scale suitable for biological interpretation and downstream prediction tasks.
The authors introduce CohortFM, a flow-matching–based framework that reconstructs continuous trajectories for each patient from sparse longitudinal molecular observations. CohortFM integrates a flow-matching learning mechanism with a mixture model that organizes individual trajectories into a compact set of cohort-level progression patterns.
Key aspects of the framework reported by the authors include:
The authors emphasize that flow matching is central to how the model learns continuous dynamics from discrete samples.
CohortFM explicitly separates molecular progression from calendar time. Rather than relying on chronological timestamps alone, the framework places patients onto a shared progression scale. This shared scale enables comparisons across patients even when their sampling schedules and progression rates differ.
By decoupling molecular progression from calendar time, CohortFM can align patients according to similar biological states rather than similar clock times, facilitating cohort-level pattern discovery and comparison.
After learning cohort-level progression patterns, CohortFM maps the inferred trajectories back to actual patient samples. This mapping supports biological interpretation: cohort patterns are not purely abstract trajectories but are grounded in observed molecular states from real patients. The authors report that this mapping aids in interpreting the biological programs and temporal dynamics captured by the learned trajectories.
The authors evaluated CohortFM across four infectious-disease contexts: COVID-19, sepsis, influenza, and tuberculosis. These datasets represent different sampling patterns and clinical contexts. According to the report, CohortFM generalized well across these varied datasets and consistently produced meaningful trajectories despite differences in how patients were sampled.
The framework reportedly captured biological differences between patients that were not explainable from the observed samples alone, indicating that the inferred continuous trajectories add information beyond discrete visit measurements.
CohortFM was compared with state-of-the-art trajectory baselines. The authors report that CohortFM produced more biologically distinct trajectories and offered a more complete representation of the patient population than the baselines tested. In particular, the finer progression scale provided by CohortFM revealed biological programs and temporal dynamics that were not captured by visit-level analyses.
To assess the model’s ability to recover unobserved states, the authors evaluated CohortFM on interpolation tasks and on true forward-prediction (forecasting) challenges. The reported results indicate that CohortFM remained competitive with strong interpolation baselines and was among the strongest learned predictors in the more difficult forecasting setting.
These evaluations suggest CohortFM can both impute intermediate states and produce useful predictions of future molecular states from sparse longitudinal data.
The authors performed ablation experiments to probe which components of the design drive performance. They report that flow matching is the core learning mechanism underpinning continuous trajectory reconstruction. Additional design choices implemented by the authors were reported to improve continuous trajectory learning, though specific ablation details are contained in the original preprint.
According to the authors’ conclusions, CohortFM offers a practical framework for converting sparse molecular snapshots into continuous, biologically interpretable trajectories. By providing a shared progression scale and cohort-level patterns, the method enables study of disease progression and the prediction of future molecular states even when patients are sampled only a few times.
The finer progression resolution that CohortFM provides can reveal biological programs and temporal dynamics missed by visit-level analyses, potentially informing mechanistic studies and cohort stratification based on molecular progression rather than calendar time.
All results and experimental details referenced here come from the cited preprint. The source reports evaluations across COVID-19, sepsis, influenza, and tuberculosis and states that ablations and comparative baselines were used. Specific numerical results, dataset sizes, training details, and other experimental parameters are reported in the original preprint but are not restated here beyond the high-level summaries provided in the source.
The authors declared no competing interests. The preprint is available via bioRxiv under a CC-BY-NC 4.0 International license and was posted October 10, 2026.