Reconstructing disease dynamics for mechanistic insights and clinical benefit

Diseases are dynamic processes involving multifaceted changes at various levels, from intracellular molecular states to systemic physiological changes. High-resolution understanding of disease dynamics is critical for developing preventive and therapeutic strategies. Time-series experimental designs offer an opportunity to study these dynamics. However, capturing disease dynamics is challenging due to complexity and high heterogeneity between individuals. Existing methods, such as PCA, FA, and diffusion maps, often suffer from low interpretability. Recent advances in differentially expressed feature analysis compare entire states (e.g., disease vs. control) but lack the ability to describe global disease dynamics. This research addresses this limitation by presenting TimeX, a novel method to capture disease dynamics using time-series data from multiple individuals. Inspired by multiple sequence alignment, TimeX performs multiple trajectory alignment, offering significant advantages for molecular interpretation and clinical diagnosis of acute and chronic diseases. The study demonstrates TimeX's applicability to diverse time-series data types, including transcriptomes and medical imaging features. The overall aim is to establish a high-resolution understanding of disease progression, discover underlying molecular mechanisms, and enhance clinical decision-making, including designing new interventions.

The introduction mentions the challenges in capturing disease dynamics due to complexity and heterogeneity. Existing methods like PCA, FA, and diffusion maps are noted for their limitations in interpretability. The authors point out that previous approaches focusing on differentially expressed features often compare entire states rather than the global dynamics of disease progression. The paper does not contain a dedicated literature review section beyond these introductory remarks.

TimeX is a three-step algorithm. First, it selects a subset of features with locally similar dynamics across patients, termed the 'conserved-dynamics-seed'. This seed isn't directly linked to the disease progression axis but serves as a backbone for comparing patient trajectories. The selection can be user-defined or computationally determined. Second, TimeX constructs a model approximating the shared dynamics across all patient trajectories. Third, this model is used to infer 'disease pseudotime' for each sample, representing its position along the shared disease progression axis. Disease pseudotime is independent of chronological time and captures the shared disease dynamics better. The algorithm explicitly models sample identity and sequential sampling. The researchers used TimeX to analyze influenza infection, urothelial bladder cancer (UBC), and age-related macular degeneration (AMD), employing various time series as input, including transcriptomics and features derived from OCT scans. For each disease, they contrasted TimeX's performance with chronological time and existing stratification methods. They evaluated the performance of the TimeX algorithm on various datasets (Influenza, UBC, and AMD). Each dataset required different pre-processing steps depending on the nature of the data (RNAseq, OCT scan features). Cell type deconvolution was also employed for UBC data. Statistical analysis including polynomial regression, linear regression, t-tests, log-rank tests, Fisher's exact tests, and Kolmogorov-Smirnov tests were used to assess the significance of the results. Pathway enrichment analysis was performed using GO, MsigDB Hallmark gene sets, Reactome, and KEGG pathways.

TimeX effectively captures shared disease dynamics across multiple patients, even with high heterogeneity. In influenza, disease pseudotime clearly distinguished symptomatic from asymptomatic patients, identifying novel genes associated with disease progression. In UBC, TimeX revealed a 'stromal pro-invasion point' (SPIP) characterized by immune cell infiltration and increased mortality. TimeX outperformed chronological time in identifying molecular associations. SPIP analysis revealed that transition from pre- to post-SPIP positions was associated with worse patient outcomes, independent of molecular subtype. TimeX provided higher resolution stratification of UBC patients compared to existing molecular subtyping frameworks, differentiating between early and late tumors within the same subtype. Analysis of early and late UroA tumors identified distinct molecular modules, suggesting mechanisms driving malignant transformation. The downregulated module included pseudogenes and microRNAs potentially suppressing transformation, while the upregulated module included genes involved in processes like the ubiquitin proteasome system, autophagy, oxidative phosphorylation, and mitotic kinetochore-microtubule interactions, pointing towards chromosomal instability. Analysis of AMD using OCT scans showed a strong association between disease pseudotime and disease severity (visual acuity), further demonstrating TimeX's versatility in handling diverse data types.

The findings demonstrate TimeX's superior ability to reconstruct disease dynamics compared to chronological time and existing stratification methods. The identification of SPIP in UBC highlights a clinically relevant transition point with prognostic value. The detailed molecular insights gained from TimeX analysis provide potential new targets for therapeutic intervention. The algorithm’s adaptability to different data types (transcriptomics, imaging) expands its applicability across various diseases. The study's success in uncovering novel molecular mechanisms underlying disease progression underscores the value of modeling continuous disease dynamics. Future research could explore TimeX's application to other diseases and data modalities.

TimeX offers a powerful approach to reconstructing disease dynamics, providing enhanced molecular interpretability and clinical utility. Its application across diverse diseases and data types demonstrates its versatility and potential to improve patient stratification, outcome prediction, and mechanistic understanding of disease progression. Future work should focus on expanding TimeX's capabilities, such as handling sparse data and incorporating additional data modalities, to further enhance its value in clinical research and practice.

While TimeX demonstrates significant advantages, limitations include the requirement for longitudinal data with at least three time points per patient. The performance might be influenced by the quality and quantity of the input data. The interpretation of molecular mechanisms requires further experimental validation, as highlighted by the discussion of chromosomal instability in UBC. The conserved-dynamics-seed selection method could be further refined.