Reconstructing disease dynamics for mechanistic insights and clinical benefit

TimeAx reveals shared disease dynamics across multiple patients

Patient cohorts tend to display high heterogeneity in patients’ disease courses, masking shared disease progression dynamics and the underlying biological mechanisms that are shared across all patients. Often, the naive solution is clustering patients into disease-specific, clinical stages or subtypes, which many times fails to capture the continuous dynamics of the disease and its progression over time (Fig. 1A). To overcome this problem, TimeAx quantitatively models a representation of the shared disease dynamics over time. TimeAx relies solely on measured features (i.e, genes, clinical markers, etc.) collected longitudinally from multiple patients (3 or more time points per patient). Importantly, patient time points can differ in number and in collection time. Broadly, the TimeAx process consists of three steps (Fig. S1A, see Methods): First, a feature selection step, in which TimeAx uses either a user-predefined or an unsupervised, computationally-selected, set of features whose dynamics are loosely similar across patients (“conserved-dynamics-seed”). While not directly associated with the hidden disease progression axis, their shared dynamics point toward their ability to serve as the backbone for the comparison between patients’ disease trajectories (Fig. 1B). Second, TimeAx builds a model, which approximates the shared dynamics across all patients’ disease trajectories (Fig. 1C). Finally, TimeAx can leverage the model to identify the disease state of a particular individual at a particular time point, referred to as ‘disease pseudotime’. By capturing a shared representation of disease progression dynamics across patients, TimeAx-inferred disease pseudotime can simulate patient-specific disease states, discovering disease progression-related mechanisms and providing predictive clinical utility. By explicit modeling of sample identity and sequential sampling as well as iterative building of the trajectory, TimeAx outperforms standard models using chronological time or trajectory inference methods (Fig. 1D; additional information in Supplementary Note 3).

Disease pseudotime captures disease progression dynamics better than chronological time

To showcase the molecular and clinical benefits of explicitly modeling disease progression from longitudinal data, we contrasted disease pseudotime with chronological time as a measure of disease progression, in both simulations (see Supplementary Note 3and Fig. S1C–G) and real-life human disease datasets, using various data types as input. We first focused on influenza infection, an acute disease which affects the immune-system over short periods of time. In this study, 17 healthy adults were challenged with influenza and then profiled longitudinally for whole blood gene expression by RNA-seq at 13–15 fixed time points within the first 108 h following infection (Fig. 2A, see Methods; total of 268 samples⁷, denoted as ‘Longitudinal influenza cohort’). As samples were collected at fixed times, similarly across all patients, chronological time points could not be used to differentiate between symptomatic and asymptomatic patients (Fig. S2A). On the other hand, based on TimeAx modeling, we projected disease pseudotime positions for all samples, observing a clear gradual increase in disease pseudotime in symptomatic but not in asymptomatic patients (Fig. 2B and Fig. S2B; p < 10⁻²⁹), an observation which we validated using two additional cohorts, including both children and adult patients (Figure S2C–D; p < 10⁻⁴⁷ and 0.0003 for longitudinal adult and children cohorts respectively, see Methods). Moreover, we identified changes in molecular processes over the course of disease pseudotime that would otherwise be missed using chronological time as a basis for disease progression (Fig. 2C and Fig. S2E, see Methods). In addition to genes which were associated with both disease pseudotime and chronological time, we identified 3432 genes which were only significantly associated with disease pseudotime, accounting for ~29% of the genes and 79% of genes with any association. Of note, genes with positive associations to disease pseudotime were highly enriched for multiple pathways, including the interferon pathway and heme metabolism (Fig. S2F), which is in concordance with previous findings emphasizing these pathways as related to augmented immune responses^8,9,10,11 with implications to disease severity and patients’ clinical outcomes during influenza infections.

A An illustration of the influenza infection dynamics TimeAx modeling and disease pseudotime inference, based on the longitudinal influenza cohort. B TimaAx disease pseudotime (y-axis) is different from chronological time (time from infection, x-axis). Shown are the differences between symptomatic (green lines) and asymptomatic (red lines) patients from the longitudinal influenza cohort. P value was calculated by comparing the prediction of disease pseudotime (by ANOVA), using only time or the interaction between time and symptoms as predictors. Trend lines represent the average levels in each of the groups ± standard error. C Gene associations (−log10 transformed, FDR-corrected, Q values based on linear regression), across all symptomatic and asymptomatic patients, using either sampling time (x-axis) or disease pseudotime (y-axis), and applying a Q value threshold of 10⁻⁵ (Dashed lines) (See Methods). Genes are colored based on their association with the two time axes. D An illustration of the UBC dynamics TimeAx modeling and disease pseudotime inference, based on the UBC longitudinal cohort. E TimeAx disease pseudotime (y-axis) is different from chronological time (time from primary tumor, x-axis), exemplified in six patients in the UBC longitudinal cohort. F Gene associations (−log10 transformed, P values based on linear regression) with sampling time (x-axis) and disease pseudotime (y-axis), using a P value threshold of 10⁻² (Dashed lines) (See Methods). Genes are colored based on their association with the two time axes, displaying significant associations almost entirely only with the disease pseudotime. G An illustration of the AMD dynamics TimeAx modeling and disease pseudotime inference, based on segmented features extracted from OCT scans of the patients’ retina. Optical coherence tomography (OCT). H TimeAx disease pseudotime (y-axis) is different from sampling time (time from first encounter, x-axis), exemplified in five patients in the AMD train cohort. I The distribution of AMD test cohort patients’ disease pseudotime positions (top; n = 29205 biologically independent samples) and times from diagnosis (bottom; n = 11075 biologically independent samples) (y-axis), across different visual severity states, determined according to the patients’ visual acuity levels (logMAR; x-axis). Boxes represent the 25th, 50th, and 75th percentiles; whiskers show maxima and minima. P values were calculated based on linear regression. We thank Yuval Abraham for his contribution in the design and creation of (A, D and G).

Diseases often advance slowly and progress at different rates across patients, making it difficult to track their dynamics and understand their molecular drivers. A common approach is to cluster patient samples into disease subtypes, clinical or data-driven, which risks losing the continuous aspect of disease dynamics due to large differences in rates of disease progression between patients^12,13. To highlight the ability of TimeAx to capture long-term processes of disease progression through explicit quantitative modeling of disease dynamics, we studied urothelial bladder cancer (UBC). UBC is a tumor with high recurrence rates after cancer removal or treatment that predominantly presents as a non-muscle invasive tumor with a small proportion of patients progressing to its muscle-invasive form, increasing the risk of developing metastases¹⁴. We trained a TimeAx model using time series microarray data from 18 patients with recurring non-muscle invasive bladder cancer who ultimately progressed to advanced disease and who were sampled longitudinally during each incidence of tumor recurrence (Fig. 2D, see Methods; ‘UBC longitudinal cohort’). In this cohort, each patient had 4–6 samples, collected up to 15 years apart from first to last recurrence¹².

Using TimeAx, we inferred disease pseudotime positions for this cohort, which exhibited high patient-specific variability when considering the chronological time that had passed since their primary diagnosis (Fig. 2E). UBC disease pseudotime also uncovered strong molecular associations, which could not be observed when modeling the data using chronological time (Fig. 2F; using linear regression). Specifically, we identified 7484 genes (~32% of the genes and 95% of genes with detected signal) as significantly associated solely with disease pseudotime and not with the chronological time (Fig. 2F, upper left quarter). These included known clinical biomarkers of UBC progression such as CCL2 and IFITM2, as well as negatively associated SGPL1, a marker linked to positive outcomes in cancer^15,16,17 (Fig. S3A). This signal enhancement was also observed at the pathway level, where a TimeAx based analysis identified stronger associations for known cancer-related processes such as the epithelial-mesenchymal transition (EMT)^18,19, TNFα signaling²⁰, interferon gamma²¹ and G2/M cell cycle checkpoint²² (Fig. S3B, see Methods; q < 10⁻⁴⁸, q < 10⁻²⁵, q < 10⁻¹⁹ and q < 10⁻⁹, respectively).

Dynamic modeling by TimeAx is applicable to multiple data types, including imaging technologies commonly used for patient diagnosis and monitoring in the clinic. To further demonstrate the scope of TimeAx’s utility, we sought to apply TimeAx to age-related macular degeneration (AMD), a chronic disease monitored periodically using low cost and non-invasive optical coherence tomography (OCT)²³. AMD is an irreversible progressive chronic disease of the retina, resulting in decreased visual acuity and is one of the leading causes of blindness in developed countries^24,25. We generated a TimeAx model of AMD progression, using segmented features, generated from OCT scans of 157 patients, each with 15 to 79 consecutive scans over several years (4953 scans overall; AMD train cohort). We then used the generated model to predict disease pseudotime positions for an additional 34,836 OCT scans, collected from 1641 different patients (Fig. 2G, see Methods; 2–90 consecutive scans per patient; denoted as ‘AMD test cohort’²⁶). The original analysis, based on OCT scans from a subset of both cohorts, used chronological time but did not identify any changes in retinal morphology associated with disease progression²⁶. Consistent with this finding, we observed that individuals’ disease pseudotime generally increased over time, and was highly variable between patients (Fig. 2H), suggesting that disease pseudotime accounts for patient variability—which is only partly captured by chronological time. Indeed, we observed a strong association between the increase in disease pseudotime and the severity of the patients’ disease burden as assessed by visual acuity (Fig. 2I, top), while no significant association was found for chronological time (Fig. 2I, bottom). Moreover, we observed an increase, at higher disease pseudotime positions, in the usage of anti-VEGF injections —a clinical procedure for reducing the accumulation of retinal fluids that are associated with worst visual acuity (Fig. S4A; p < 10⁻¹⁹²). This supports the notion that retinal fluids appear, mostly, in late stages of the disease (see Supplementary Note 4 for full the AMD progression analysis, including the identification of disease progression-related segmented features and clinical applications). Taken together, these results suggest that TimeAx dynamic modeling enables molecular interpretability and may provide increased clinical utility compared to naive longitudinal monitoring based solely on chronological time.

TimeAx uncovers an advanced tumor state with unfavorable clinical outcomes

To further highlight the utility of TimeAx, we focused on the UBC progression model where the long time scales of disease progression result in large differences between patients’ disease progression rates on the one hand, and on the other, the high resolution molecular data enables the study of cellular and molecular mechanisms of tumorigenesis. To confirm that our TimeAx model captures tumor progression, we compared the disease pseudotime positions in the UBC longitudinal cohort with the tumor stage of the patients (see Methods; based on TMN staging), and saw a clear association between increased disease pseudotime and more advanced stages (Fig. 3A, p < 0.0005 by linear regression), compared to no association using chronological time (Fig. S5A). We validated these results using held-out longitudinal UBC samples from the same dataset, and two additional cross-sectional UBC test cohorts ‒ the former consisting of microarray data of 276 UBC patients²⁷ and the latter of 430 UBC RNA-seq samples from the Cancer Genome Atlas (TCGA) Program (Fig. S5B, see Methods; respectively denoted as ‘longitudinal’, ‘microarray’ and ‘TCGA’ test cohorts). Tumors with higher disease pseudotime positions were also more transitional/invasive (Fig. S5C; p < 10⁻⁷), compared to more papillary tumors found at lower pseudotime positions. In addition, patients with prior malignancy were positioned further along disease pseudotime (defined by TCGA; Fig. S5D; p < 10⁻⁴).

A Disease pseudotime clinical applicability. Shown are disease pseudotime (y-axis) relation with tumor stage (x-axis) for samples within the UBC longitudinal cohort. P value was calculated based on linear regression. B Tumor purity scores (y-axis) along the disease pseudotime (left) and the time from primary tumor (right) (x-axis), displaying a sharp decrease in high disease pseudotime positions (SPIP; dashed line). Disease pseudotime ranges pre and post the SPIP are marked by colored bars. C Cell type deconvolved cell contributions (y-axis), displaying major differences between pre- and post- SPIP samples. * p < 0.05, ** p < 0.01 based on a two-sided t-test (p < 10⁻⁴, 0.003, 0.05, 0.05, 10⁻⁴; Urothelial cells, Macrophages, naive T cells, memory T cells and Fibroblasts, respectively). D Survival plot for UBC patients within the TCGA test cohort, comparing patients’ tumors with disease pseudotime positions lower and higher the SPIP (pre versus post, respectively; color coded). P value was calculated based on a log-rank test. E Survived (black) versus deceased (white) percentages of basal/squamous patients within the TCGA test cohort pre and post SPIP (x-axis). p value was calculated using Fisher’s exact test. In (A and C), boxes represent the 25th, 50th, and 75th percentiles; whiskers show maxima and minima and n = 84 biologically independent samples. Stromal pro-invasion point (SPIP).

We hypothesized that the TimeAx model, based on whole bulk-tissue transcriptomes, represents a continuous shift in complex multicellular programs in which both cell compositions and cell-states may change along the disease pseudotime and result in tumor progression. To explore our hypothesis, we first looked at tumor purity, a measure for the fraction of cancer cells in a tumor sample (calculated as in ref. ¹²), along the TimeAx UBC disease progression model. We identified a position along the disease pseudotime (disease pseudotime of 0.7), where a sharp decrease in tumor purity occurred, a trend undetectable when ordering samples by the time from primary tumor occurrence (Fig. 3B, see Methods; r = −0.44 and r = −0.1, respectively). A decrease in tumor purity has been previously associated with elevated immune infiltration and overall poor prognosis^28,29. This sharp decrease in tumor purity was not associated with patients’ demographics, including sex (Fig. S5E) and age (Fig. S5F). Taken together, the TimeAx model highlights the existence of a stromal pro-invasion point (SPIP), characterized by a change in the immune-stroma composition within the tumor microenvironment^29,30.

To understand how cellular composition and regulatory programs change along the disease pseudotime and how these relate to the decrease in tumor purity at the SPIP (Fig. 3B), we deconvolved the cell composition of all samples in all four cohorts and, predicted the compositions of seven major cell types: urothelial cells, muscle cells, basal tumor cells, endothelial cells and fibroblasts, as well as T cells, macrophages and additional immune cell subtypes (see Methods). Samples localized to disease pseudotime positions ‘beyond the SPIP’ showed a significant decrease in the abundance of urothelial cells, accompanied by an increase in the abundance of activated macrophages and fibroblasts (Fig. 3C and Fig. S5G; p < 0.003), a transition from naive to memory CD4 + T-cells (Fig. 3C and Fig. S5G; p < 0.05) and no significant change in basal tumor cells (Fig. S5G–H; p > 0.1). We reasoned that differences in cell compositions along the TimeAx model may be due to either differences in disease severity or technical variation during biopsy sampling. We therefore decided to leverage the clinical outcome data available in the TCGA test cohort to test whether mapping pre- versus post- the SPIP carried clinically meaningful signal (see Methods). Indeed, we observed a significantly lower survival probability for patients mapping to the TimeAx disease progression axis past the SPIP (Fig. 3D; p < 0.003 by Kaplan-Meyer). The classification of patients into either pre- or post- SPIP improved survival rate prediction even after accounting for other covariates with known association with survival, including age, sex and the clinical stage of the disease (Supplementary Data 1; p < 0.02, Cox Proportional-Hazards Model). Specifically, even within a specific molecular cancer subtype, such as urothelial-like and basal/squamous tumors, patients with disease pseudotime positions past the SPIP displayed a higher mortality percentages (Fig. 3E; 72% compared to 39%, p < 0.03) and more rapid rates of mortality (Figure S5I, J). These observations suggest that the SPIP we observed reflects a biological milestone in UBC progression, in line with previous findings associating the infiltration of activated immune cells and cancer-associated fibroblasts into the tumor microenvironment with tumor progression and poor clinical outcome^31,32. Taken together, this suggests that the TimeAx model accurately reflects tumor developmental processes and improves clinical prediction over the previously suggested tumor classification frameworks.

Disease pseudotime captures variation undetectable by current stratification frameworks

Current clinical assessment of UBC progression, as well as the selection of interventions and therapies, relies mostly on histopathological staging. As patients’ disease courses in UBC are highly heterogeneous, these traditional methodologies, focusing on a relatively small set of markers, are not sufficient for optimal clinical decision making. Recently, molecular profiling analyses divided urothelial carcinomas into two major molecular types, luminal and basal, with the latter showing down-regulation of urothelial differentiation markers, a higher incidence in muscle invasive tumors and association with worse clinical prognosis³³. The luminal type, which represents most of the non-muscle invasive tumors, can be further divided into urothelial-like and genomic-unstable subtypes, with the former harboring alterations in the FGFR3 pathway^34,35,36. While these subtyping frameworks aim to stratify patients, they are not widely applied in the clinic due to their high complexity and uncertainty surrounding their utility for clinical prognosis over traditional frameworks³³. Importantly, the grouping of patients into small sets of disease subtypes results in the loss of continuity in assessing disease progression. For example, though the original analysis¹² of the ‘UBC longitudinal cohort’ (Fig. 2D and Methods) divided patients’ UBC recurrences into distinct molecular subtypes, it also showed that clinical progression occurred in all patients regardless whereas molecular subtyping remained predominantly stable in most patients. Moreover, in some cases patients were simultaneously associated with two different subtypes¹². This suggests that discretized molecular subtyping does not reflect disease state, nor necessarily severity, and in some cases, dynamical changes in disease progression may be interpreted as different disease subtypes.

We hypothesized that the disease pseudotime we identified represents a generalized UBC disease dynamics axis, shared across luminal and basal molecular subtypes. Indeed, relying on previously published molecular typing (‘LundTax’ tumor molecular classification³⁷), modeling disease dynamics, while excluding patients with basal tumors, allocated all patients to nearly the same disease pseudotime positions (Fig. S6A; r = 0.91), and yielded similar molecular associations as in the original model (Fig. S6B). This suggests that these molecular subtypes reside in the same disease dynamics axis, which accommodates transitions between molecular subtypes along the disease pseudotime. Supporting this, we observed higher disease pseudotime levels in more progressive molecular tumor subtypes, such as basal and mesenchymal, compared to lower values in urothelial-like subtypes (Fig. 4A; p < 10⁻⁵ by one-way anova). This classification by molecular subtypes was strongly associated with patients’ allocation to either the pre- and post- SPIP groups (Fig. 4B) but the resolution achieved by this classification was insufficient for the correct assignment to patients’ histological staging (Fig. S6C, D).

A Disease pseudotime distribution across tumor molecular classifications in patients within the UBC longitudinal cohort based on the ‘LundTax’ molecular subtyping framework. B Distribution of ‘LundTax’ tumor molecular classifications for pre (red) and post (blue) stromal pro-invasion samples within the UBC longitudinal cohort. C Comparison of disease pseudotime (y-axis) between primary and recurrent tumors in pre-SPIP samples within the UBC longitudinal cohort. D Disease pseudotime distribution across tumor molecular classifications in patients pre-SPIP within the UBC longitudinal cohort based on the ‘LundTax’ molecular subtyping framework. E Percent of surviving patients in the TCGA test cohort with UroA tumors (y-axis) across disease pseudotime bins (bin size = 0.1; x-axis) pre-SPIP. p value was calculated based on linear regression. In (A, C and D), boxes represent the 25th, 50th, and 75th percentiles; whiskers show maxima and minima and n = 84 biologically independent samples.

We reasoned that in the absence of long-term molecular follow up of patients, understanding the relationship between disease progression and disease subtypes can only be garnered through further mechanistic understanding and the observation of clinically actionable predictions. As early events in UBC tumorigenesis are largely unresolved, we focused on disease pseudotime positions prior to the SPIP. TimeAx allowed us to discern high resolution dynamics and observe differences between early (primary) and recurring tumors with a large variation in disease pseudotime between recurrent tumors (Fig. 4C; p < 10⁻⁹). Considering the present UBC molecular subtypings, we observed that while most samples pre-SPIP were classified as low grade tumor stages (Lum-P, Lum-U by the consensus molecular subtyping¹³ and Urothelial-like and genomically-unstable by Lund taxonomy³⁷), those patients showed large variation in disease pseudotime (Fig. 4D and Fig. S6E). These observations suggest that the current molecular subtyping systems of UBC tumors only allow for a coarse stratification of patients, precluding the high resolution granularity provided by modeling the continuous dynamics of disease progression.

Of all molecular subtypes, UroA exhibited the largest variation along the disease progression axis whereby UroA tumors at higher disease pseudotime positions were associated with lower survival percentages (Fig. 4E; p < 0.01). Consistent with this, by dividing the UroA tumors pseudotime continuum to ‘early’ and ‘late’, respectively, using disease pseudotime cutoff of 0.25, we observed lower survival percentages in late tumors (Fig. S6F; p < 0.1)—suggesting that these tumors represent different UBC progression stages.

Disease pseudotime uncovers molecular mechanisms promoting UBC progression

We next explored whether we can detect molecular patterns associated with differences between early and late progression within the UroA tumors, and whether those differences manifest different oncogenic transformation stages. We identified 2642 differentially expressed genes between early and late UroA progression tumors, which were highly co-regulated into two main modules (q < 0.05, Fig. 5A and Supplementary Data 2), one downregulated and one upregulated along the disease progression axis. In contrast, we observed no significant changes, between the two groups of UroA tumors, in the expression of established urothelial markers, including CCND1, FGFR3, FOXA1, RB1, CDKN2A, GATA3, ERBB2, PPARG and XBP1 (Fig. S6G).

A Co-expression matrix for genes (rows, columns) differentially expressed (q < 0.05) between early and late UroA tumors. Two gene sets were discovered and are highlighted in green and purple. B High enrichment of different pseudogene families within the downregulated module, compared to the upregulated module set and a random gene set (color-coded). C Pathway enrichment for the downregulated module, including pathway enrichment scores (left) and a heatmap of the expression levels of these pathways genes (columns), in early and late UroA tumors (rows) within the UBC longitudinal cohort. D Pathway enrichment scores for the upregulated module. E An illustration of the molecular model of UBC tumorigenesis, discovered by the TimeAx-based analysis, based on the increasing gene module. The illustration includes the association of known hallmarks of cancer with disease pseudotime (top) and the suggested mitotic kinetochore- spindle microtubules (MT) interaction (middle), which is also presented as boxplots of differential expression of its subunits (bottom; n = 40 biologically independent samples, boxes represent the 25th, 50th, and 75th percentiles; whiskers show maxima and minima). G protein-coupled receptors (GPCRs). We thank Yuval Abraham for his contribution in the design and creation of (E).

The downregulated module was composed of 1587 downregulated genes within late UroA tumors (Fig. S6H). This module was highly enriched in pseudogenes and microRNAs, (q < 10⁻⁴, 0.03 respectively; Fig. 5B, see Supplementary Data 3 for functional enrichment). Interestingly, we noted that for many of these pseudogenes, their coding paralogs were transcribed by RNA polymerase III³⁸ and involved in biosynthetic processes promoting cancer cell proliferation³⁹, suggesting pseudogenes actively suppress the transformation process by downregulation of their coding equivalents. In addition, we noted an enrichment for protein coding genes in this module transcribing membrane channel proteins that were downregulated along disease pseudotime. Specifically, these included ligand gated ion channels (q < 0.05, Fig. 5C, see Supplementary Data 4), primarily calcium, potassium and sodium voltage- gated channels, controlling cellular ion homeostasis and downstream cell cycle and cell death processes. Interestingly, we also detected down regulation of G protein-coupled receptors (GPCRs) including neurotransmitter, hormone and free fatty acids receptors (q < 0.05, Fig. 5C, see Supplementary Data 4). GPCR associations with tumor progression have remained unclear, with evidence showing both tumor progressor⁴⁰ and suppressor⁴¹ functions, likely explained both by differences in functionality between GPCRs and tissue and malignancy dependency. Our analysis showing downregulation of GPCRs as the disease shifts toward advanced UroA, pinpoints these GPCRs as likely tumor suppressors in bladder cancer.

The upregulated disease progression module contained 1055 genes (Fig. S6I), and was highly enriched for pathways associated with malignant transformation as well as known hallmarks of cancer including sustained proliferative signaling, activating invasion and metastasis, genome instability and mutation and deregulation of cellular energetics (q < 0.05; Fig. 5D, see Supplementary Data 5). Specifically, we detected a striking upregulation of genes in the ubiquitin proteasome (UPS) system including structural components of the 26 S proteasome, components of the anaphase promoting complex and ubiquitin ligases. Post-translational polyubiquitylation of key regulatory proteins, results in their proteasomal degradation and alters regulation of cell cycle and epithelial to mesenchymal transition⁴². In addition, late UroA tumors showed upregulation of genes functioning in cellular metabolism, including autophagy and oxidative phosphorylation, allowing the tumor to meet the increasing energetic demands and support cell proliferation during oncogenic transformation⁴³.

Interestingly, we detected that the shift from early to late UroA tumor progression was associated with an abundance of genes involved in mitotic kinetochore- spindle microtubules (MT) interaction which suggested it as a mechanism for malignant transformation in UBC, undetectable without TimeAx modeling. These genes spanned six integral kinetochore constituents⁴⁴ which together suggest that the shift from early to late UroA involves increasing chromosomal instability and aneuploidy. Specifically, these include MIS12⁴⁵, Ska^46,47, RZZ⁴⁸ components of the outer kinetochore, components of the nuclear pore complex NUP107–160^49,50 localized to the kinetochore during mitosis, the MT binding protein CLASP-2 and the chromatin remodeler RSF1⁵¹, all required for stabilizing dynamic MT to kinetochores⁵². Similarly, late UroA tumors showed upregulation in genes coding for proteins localized to the spindle, including a subunit of the augmin complex (Haus) functioning in MT end capture⁵³ and KIF2A functioning in MT depolymerization, bi-polar spindle formation and chromosome pulling⁵⁴ (Fig. 5E and Supplementary Data 2). Taken together the signal detected via TimeAx disease progression modeling versus molecular subtyping suggests that modeling of disease progression not only provides high utility for understanding mechanistic changes in the biology of disease but is likely a necessary step prior to unsupervised sub-typing of diseases.