Genome-level selection in tumors as a universal marker of resistance to therapy

Abstract

Tumor evolution is shaped by selective pressures imposed by physiological factors as the tumor naturally progresses to colonize local and distant tissues, as well as by therapy. However, the distinction between these two types of pressures and their impact on tumor evolution remain elusive, mainly, due to extensive intra-tumor heterogeneity. To disentangle the effects of these selective pressures, we analyze data from diverse cohorts of patients, of both treated and untreated cancers. We find that, despite the wide variation across patients, the selection strength on tumor genomes in individual patients is stable and largely unaffected by tumor progression in the primary settings, with some cancer-specific signatures detectable in the progression to metastases. However, we identify a nearly universal shift toward neutral evolution in tumors that resist treatment and demonstrate that this regime is associated with worse prognosis. We validate these findings on both published and original datasets. We suggest that monitoring the selection regime during cancer treatment can assist clinical decision-making in many cases.

Fig. 1. Evolution regimes in Untreated Cancers.

a The relationship between dN/dS of primary (P) samples and metastatic (M) samples, prior to treatment, in 21 colorectal cancer patients (top), and the relationship between dN/dS of early primary (P1) samples and late primary (P2) samples in 9 patients (bottom). b The relationship between dN/dS of P samples and M samples, in 10 esophageal cancer patients (top), and the relationship between dN/dS of P1 samples and P2 samples in 9 patients (bottom). c The relationship between dN/dS of 3 primary samples in 36 quantifiable lung cancer patients are shown for sample 1 vs. sample 2 (top) and sample 2 vs. sample 3 (bottom). R-square (R²) of a least mean square (LMS) fit, and the linear regression fit (y=ax+b) with respective P-values of the slope (a) and intersection (b) are shown. Confidence Intervals (95%) are denoted by dashed curves. See main text for the one-way ANOVA tests comparing between two linear regressions (“Methods”).

Fig. 2. Evolution regimes in resistant treated cancers.

a The relationship between dN/dS at diagnosis and after chemotherapy at relapse of 16 pediatric ALL patients. Cases that developed a hypermutator genotype after treatment are marked (red). b The relationship between dN/dS before and after BCL-2 inhibition treatment of 7 CLL patients. The cases C811, C812, C586, and C577 (see main text) are marked. c The relationship between dN/dS of treatment-naïve primary samples and post-treatment distant metastases, of 7 ER-positive breast cancer patients following ER-targeted therapy. d The relationship between dN/dS of a treatment-naïve primary sample and the average dN/dS of primary (green) and metastases (red) samples post treatment, of 16 bladder cancer patients. 3 cases that previously received BCG are marked (arrow). R² of a Bi-square (BIS) fit is also shown due to the outliers. e The relationship between dN/dS before and after radiotherapy of glioblastoma patients (top). Cases that developed hypermutator genotypes following treatment are marked (red). Linear fit curve is shown with (thick) and without (thin) the hypermutator cases. Heatmap of P values of the Kaplan-Meier survival analysis of glioblastoma patients, with respect to dN/dS of paired samples (before and after treatment), comparing patients in the neutral regime or approaching it after treatment with patients far-from or escaping the neutral regime (bottom, left). The point marked in red (Distance = 0.38, Delta = 0.24) serves an example to demonstrate patient’s classification and the respective survival curves (bottom, right). R-square (R²) of a least mean square (LMS) fit, and the linear regression fit (y = ax+b) with respective P values of the slope (a) and intersection (b) are shown. Confidence Intervals (95%) are denoted by dashed curves. See main text for the one-way ANOVA tests comparing between the linear regressions of each cancer relative to a reference linear model (“Methods”).

Fig. 3. Analysis of published datasets of untreated and treated breast cancers.

a The relationship between dN/dS values of early and late primary untreated samples, combined from two studies^,. b The relationship between dN/dS values of primary and metastatic untreated samples. Metastatic samples are comprised of only local LN (without distant metastases) in one study, or distant metastases in another study. Divergent cases in the former and advanced cases in the latter and their fit are displayed (red). Fit without these cases is also shown (black). See also Fig. S8. c The relationship between dN/dS values of primary tumors before and after treatment, combined from two studies^,. The cases treated with HER-positive directed therapy from and the resistant (R) tumors from are shown. d The case of resistant breast tumors to ER-directed agent comparing primary and metastases (cf. Fig. 2c), is re-plotted for convenience. R-square (R²) of a least mean square (LMS) fit, and the linear regression fit (y = ax+b) with respective P values of the slope (a) and intersection (b) are shown. Confidence Intervals (95%) are denoted by dashed curves. See main text for the one-way ANOVA tests comparing between the linear regressions of each cancer relative to a reference linear model (“Methods”).

Fig. 4. Analysis of the multiple myeloma patients cohort.

a The distribution of dN/dS values across 780 samples (upper). The median of dN/dS values, percentage of genes affected by CNA, MSI score and the no. of N mutations across different statuses of the disease progression, error bars denote standard deviation of the mean (SEM) (bottom). b The relationship between dNdS values of early (1) and late (2) samples of the same patient are shown for 23 patients with untreated samples, including the healthy benign states MGUS and Smoldering and the disease state at diagnosis (left), 17 patients with disease state before (pre) and after (post) treatment (middle), and 73 patients who experienced disease relapse under continuous treatment (right). Colors code denotes comparisons of the same status, and when the status changes (arrow). The results of linear regression fits are denoted. One-way ANOVA F-test comparing untreated and treated linear regressions provides P value < 0.073 (see main text). c Heatmap of the P values of Kaplan-Meier (KM) survival analysis with respect to dN/dS, comparing values in the range µ±σ (as shown in panel A) with values outside this range, of samples in the healthy statuses (i.e., MGUS and smoldering) prior to treatment (left), in the disease status at diagnosis prior to treatment (middle) and in the relapse status (i.e., early and late) under treatment (right). Circles correspond to P values < 0.05, where cyan circles indicate better prognosis (valleys in tumor fitness) and black circles indicate worse prognosis (hills in tumor fitness) of samples in the chose range. An example of the actual KM curves for the choice µ = 1, σ = 0.24 is shown on the right. d Cox regression analysis of different drugs with respect to the distance from neutrality post treatment, |1-dN/dS | , indicating that most of the drugs lead to regime close to neutrality (HR > 1). e Heatmap of P-values of KM analysis for paired samples before and after lines of therapy, comparing patients in the neutral regime or approaching it after treatment with patients far-from or escaping the neutral regime (left). The point marked in red (Distance = 0.32, Delta = 0.3) serves an example to demonstrate patient’s classification and the respective survival curves (right).

Fig. 5. Analysis of single patients with multiple samples in the multiple myeloma cohort.

a Representative examples of phylogenetic trees of patients with multiple samples; trajectories are denoted as arrows from pre-treatment (P) to early (E) and late (L) recurrences of sequential biopsies (upper). Dominant COSMIC mutational signatures along the tree branches (color coded), mirroring the dominant signatures in the entire cohort (see Fig S12). dN/dS values of trunk and branches (blue) and of each leaf (black) are denoted. The respective clonal compositions, of sequential biopsies, are shown for each patient, with the dN/dS of each sample (bottom). b The distributions of N mutations and dN/dS values across patients with at least 3 samples (n = 27), as deduced from the phylogenetic trees of each patient, for all mutations, clonal (trunk, shared by all) mutations and subclonal (branch, private) mutations. Box plots denote the median (red) and the edges of the box are the 25th and 75th percentiles, with whiskers extending to the most extreme data points. See Fig. S13 for similar analysis in other cohorts, indicating that positive selection in the trunk and negative selection at the branches is a universal feature. c Histogram of the difference in the no. of clones between consecutive biopsies under treatment in a patient, indicating that the no. of clones is highly stable (but with dynamic proportions as in panel A), with a weak tendency for reduced clonality.

Fig. 6. Correlation between dN/dS and N as function of allele frequency (AF) in single patients.

a Heatmap of Pearson correlation coefficients of the correlation between dN/dS and N as function of AF, across studies, for patients with at least 5 samples. Black vertical line divides the plane from early (AF>values) to late (AF<values) mutations. The horizonal line separates between studies, with the upper panel corresponding to colorectal patients, and below it two single patients: blader patient (P117) with 16 samples from and the single melanoma patient Mel-37 with 37 samples. b Example of a colorectal cancer patient (V750, with late metastatic seeding), which includes untreated (U) primary (n = 4) and LN (n = 3) samples and treated brain metastases (n = 5). dN/dS and N distribution are shown, along with the positive correlation for clonal-enriched early mutations (AF > 0.25) and the negative correlation for subclonal-enriched late mutations (AF < 0.33). See Fig. S15 for patient V930 (with early metastatic seeding). c Example of the bladder patient, exhibiting sharp decline in N following therapy with increased load in the metastases, compared to the untreated primary (green lines), as well as dominant positive correlation between dN/dS and N across most ranges of AF, but specifically in late phases. Only samples that pass the criteria of sufficient no. of mutations in each test are analyzed and shown (“Methods”). Box plots (in b, c) denote the median (red) and the edges of the box are the 25th and 75th percentiles, with whiskers extending to the most extreme data points. Data points are superimposed (x).

Fig. 7. Summary of the findings on tumor evolution regimes and their potential translational value.

a The validated non-monotonic, be-phasic evolutionary regime of tumor evolution, based on theory (black) and empirical evidence across primary cancers (colored lines; each line represent a cancer cohort), showing that tumors operate predominantly near the neutral regime, with signatures of positive selection in early stages (and low N) representing the accumulation of drivers, and signatures of negative selection in late stages (and high N) indicating the removal of deleterious passengers. b The relationship between dN/dS values of early and late samples prior to treatment (green) indicating the roughly linear relationship in the natural progression of primary tumors (solid ellipse), and cancer specific signature in the natural progression to metastases (dashed ellipses). The relationship between the dN/dS values before and after treatment (red), depicting the near universal observed tendency towards neutral evolution (i.e., dN/dS values approach 1) when tumors resist treatment. c The translational value of monitoring dN/dS values of biopsies to guide treatment. If a patient starts with dN/dS values around neutrality (dN/dS ~ 1), and following treatment, the dN/dS value changes toward a selective regime, the treatment is likely working, and the recommendation would be to continue treatment. In contrast, if dN/dS values before treatment are far from neutrality but steadily approach neutrality following treatment, the treatment is likely to fail, and the recommendation would be to stop or change the treatment.