Reconstructing disease dynamics for mechanistic insights and clinical benefit

Diseases are dynamic processes spanning intracellular molecular states and systemic physiological changes. Heterogeneity in individual disease courses obscures shared progression dynamics, hindering diagnostics, mechanistic understanding, and treatment design. Traditional stratification clusters patients into discrete subtypes or stages, losing the continuity of progression and often failing to capture temporal dynamics. Dimensionality reduction and trajectory inference methods (e.g., PCA, FA, diffusion maps) can be driven by inherent variation and offer limited interpretability for disease trajectories. The study aims to model a shared, continuous disease progression axis (disease pseudotime) across individuals using longitudinal, high-dimensional measurements, enabling higher-resolution patient stratification, mechanistic insight, and improved clinical prediction.

The paper situates its contribution among approaches for analyzing temporal data. It notes that standard dimensionality reduction (PCA, FA) and trajectory inference methods (diffusion maps) may be dominated by non-temporal variation, limiting interpretability for progression. Prior advances have focused on differential expression over time when comparing broad states (e.g., case vs control), not on reconstructing global disease dynamics. The authors draw analogy to multiple sequence alignment, proposing multiple trajectory alignment to uncover shared dynamics. References include work on diffusion maps, time-course differential expression models, and molecular subtyping frameworks for bladder cancer (e.g., Lund and consensus classifications), highlighting their limitations for capturing continuous progression.

TimeAx (also referred to as TimeX in the abstract) models shared disease dynamics from longitudinal data collected across multiple individuals (minimum 3 time points per individual; time points can differ in number and timing). Core steps: (1) Feature selection to identify a conserved-dynamics-seed—features whose dynamics are locally similar across patients—either user-specified or selected computationally; (2) Multiple trajectory alignment across patients using the conserved-dynamics-seed to anchor shared dynamics; (3) Model construction that approximates the shared trajectory and assigns each sample a disease pseudotime reflecting its position along the progression axis. The model explicitly accounts for sample identity and sequential sampling and iteratively builds the trajectory. Data modalities and cohorts:

• Influenza: Training on an H3N2 human challenge cohort (17 adults, 13–15 timepoints over ~108 hours post infection) with 1,258 total samples analyzed, including a longitudinal cohort and two validation cohorts (an adult H1N1 challenge dataset and a pediatric H1N1 dataset). Pseudotime was inferred and associated with symptomatology and gene expression dynamics; polynomial regression with FDR control assessed associations; pathway enrichment via KS tests over correlation distributions (GO, Hallmark, Reactome, KEGG).
• Urothelial bladder cancer (UBC): Training on longitudinal microdissected tumors from 15 patients with recurring non–muscle-invasive bladder cancer, 4–6 samples per patient collected over up to 15 years (UBC longitudinal cohort). A conserved-dynamics-seed of 100 genes was used. Validation in additional 28 patients (longitudinal test cohort), a microarray cohort of 276 cystectomy patients (GSE83536), and TCGA RNA-seq cohort (n=430). Associations with pseudotime vs chronological time assessed via linear regression; cell type deconvolution (e.g., LM22 and tumor-derived single-cell signatures) to estimate tumor microenvironment composition; survival analyses (log-rank, Fisher’s exact) in TCGA; comparisons with molecular subtyping frameworks (LundTax, consensus). Differential expression and co-expression analyses contrasted ‘early’ vs ‘late’ tumors within UroA subtype; functional enrichments assessed.
• Age-related macular degeneration (AMD): OCT-based modeling trained on segmented features from 157 patients (15–79 scans each; 4,953 scans total) using a deep U-Net semantic segmentation ensemble to derive features (e.g., retinal atrophy, fibrosis, thickness, subretinal hyperreflective material, RPE, fibrovascular PED, drusen, choroid). The trained model predicted pseudotime for 34,836 scans from 1,641 patients (test cohort). Associations between pseudotime and clinical measures (visual acuity, anti-VEGF usage) were evaluated and contrasted with chronological time. Statistical analyses: Gene–pseudotime associations via polynomial regression with FDR correction (q cutoffs as specified for figures). Pathway enrichment using KS tests on correlation distributions (−log10 transformed scores). Survival via log-rank tests; group comparisons via t-tests and Fisher’s exact tests. Tumor purity and cell-type proportions assessed across pseudotime; visualization via boxplots and distributions.

• Across diseases and modalities, disease pseudotime better captures progression than chronological time, enabling clearer molecular and clinical associations.
• Influenza: Pseudotime increased in symptomatic but not asymptomatic patients; significant separation reported (e.g., p = 7.89 × 10^-3). Validation cohorts (adults and children) showed strong significance (reported p-values including 10^-7 and 0.003). 3,432 genes were significantly associated only with disease pseudotime (~29% of genes; 79% of genes with any association), revealing pathways enriched among positively associated genes, including inflammation and heme metabolism.
• UBC longitudinal cohort: Pseudotime varied markedly across patients relative to time since diagnosis, revealing stronger molecular signal than chronological time. 7,484 genes (≈32% of genes; 95% of genes with detected signal) were significantly associated solely with disease pseudotime. Notable associations included known biomarkers (CCL2, ITF2) and newly highlighted markers (SPGL1). Enriched pathways included EMT, interleukin signaling, and cell cycle checkpoints (e.g., q < 10^-6, q < 10^-2, q < 10^-3). A stromal pro-invasion point (SPIP) at high pseudotime marked a sharp decrease in tumor purity and profound tumor microenvironment remodeling with increased immune infiltration (macrophages, T cells) and fibroblasts. TCGA validation showed worse survival for tumors post-SPIP (log-rank), and basal/squamous tumors post-SPIP had higher death rates (Fisher’s exact).
• UBC subtype dynamics: Disease pseudotime formed a common axis shared across luminal and basal subtypes, capturing transitions not reflected by discrete subtyping. Urothelial-like (Uro) tumors showed wide pseudotime variation; within UroA, later pseudotime associated with lower survival (e.g., p < 0.01), and splitting UroA into ‘early’ vs ‘late’ pseudotime suggested stage-related survival differences.
• UBC mechanistic insights (UroA early vs late): Identified 2,642 differentially expressed genes forming two co-regulated modules: a downregulated module (1,587 genes) enriched for pseudogenes and microRNAs (q < 10^-4, 0.03) and pathways including ligand-gated ion channels and GPCRs; and an upregulated module (1,055 genes) enriched for ubiquitin–proteasome system components, APC and ubiquitin ligases, autophagy, and oxidative phosphorylation, indicating proliferative and metabolic reprogramming. Genes at the kinetochore–microtubule interface were upregulated, suggesting increased chromosomal instability/aneuploidy; PPP2R3A (PP2A B56) upregulation could stabilize kinetochore–MT attachments, potentially counteracting Aurora B–mediated error correction.
• AMD: Pseudotime generally increased over chronological time but varied widely across patients, capturing patient-specific progression rates. Pseudotime, but not chronological time, was strongly associated with disease severity (visual acuity). Higher pseudotime correlated with increased anti-VEGF usage (p < 10^-2), consistent with fluid accumulation in later disease stages.
• Overall: TimeAx delivered improved molecular interpretability and clinical utility for patient stratification and outcome prediction compared with chronological time or static subtyping.

Modeling shared, continuous disease dynamics across individuals addresses the limitations of discrete subtyping and chronological time, revealing latent temporal structure that aligns with symptomatology, molecular programs, and outcomes. In influenza, pseudotime recapitulated host response trajectories and uncovered gene/pathway dynamics otherwise obscured. In AMD, pseudotime linked imaging-derived structural changes to clinical severity more robustly than chronological time, enabling scalable progression monitoring. In UBC, the pseudotime axis unified luminal/basal subtypes and exposed a stromal pro-invasion transition characterized by immune infiltration, decreased purity, and poorer survival—refining risk stratification even within the same molecular subtype. Within UroA, early-to-late transitions corresponded to coordinated transcriptomic remodeling involving reduced GPCR/ion channel signaling and increased UPS, autophagy, OXPHOS, and kinetochore–microtubule interface gene expression, pointing to mechanisms of malignant transformation and chromosomal instability. These results demonstrate that TimeAx-derived pseudotime provides a mechanistically meaningful, clinically actionable measure of disease state that improves prediction and stratification across diverse diseases and data modalities.

The study introduces TimeAx, a multiple trajectory alignment framework that reconstructs shared disease dynamics from longitudinal, high-dimensional data to infer a continuous disease pseudotime. Applied to influenza, AMD, and UBC, TimeAx improved detection of molecular programs and clinical associations compared with chronological time, enabled higher-resolution patient stratification beyond discrete subtyping, identified a stromal pro-invasion point with adverse outcomes in UBC, and uncovered potential mechanistic drivers of progression, including alterations in ubiquitin–proteasome, metabolic pathways, and kinetochore–microtubule interactions. Future directions include extending seed detection to complex (non-monotonic) feature dynamics, accommodating fewer than three time points per individual, meta-analyses across datasets, modeling recovery dynamics and non-disease processes (e.g., immune age), and integrating additional modalities (proteomics, microbiome, epigenetics, routine clinical markers, and multi-omics). Experimental validation of proposed mechanistic hypotheses (e.g., chromosomal instability drivers) will further translate insights into therapeutic strategies.

• Current seed detection focuses on monotonically increasing/decreasing features; more complex temporal patterns are not yet modeled.
• Requires at least three time points per individual; many clinical datasets have sparser sampling.
• Heterogeneity and environmental covariates may still influence trajectories; model captures shared dynamics but not all individual-specific effects.
• Some mechanistic inferences (e.g., kinetochore–microtubule stabilization and CIN in UBC) are correlative and require experimental validation.
• Discrete subtyping comparisons may be impacted by classification uncertainties and cohort composition; generalizability across institutions and assays needs further evaluation.
• Imaging-to-molecular linkage in AMD is indirect; treatment effects (e.g., anti-VEGF) and technical variation may confound associations and require careful adjustment.