Five latent factors underlie response to immunotherapy

Abstract

Only a subset of patients treated with immune checkpoint inhibitors (CPIs) respond to the treatment, and distinguishing responders from non-responders is a major challenge. Many proposed biomarkers of CPI response and survival probably represent alternative measurements of the same aspects of the tumor, its microenvironment or the host. Thus, we currently ignore how many truly independent biomarkers there are. With an unbiased analysis of genomics, transcriptomics and clinical data of a cohort of patients with metastatic tumors (n = 479), we discovered five orthogonal latent factors: tumor mutation burden, T cell effective infiltration, transforming growth factor-beta activity in the microenvironment, prior treatment and tumor proliferative potential. Their association with CPI response and survival was observed across all tumor types and validated across six independent cohorts (n = 1,491). These five latent factors constitute a frame of reference to organize current and future knowledge on biomarkers of CPI response and survival.

Fig. 1. Extracting features from the HMF-CPI cohort.

a–c, For 479 patients with metastatic cancer in the HMF-CPI database of different cancer types, we obtained 18 clinical features, 19 germline HLA allotype features, 18,382 somatic features (based on single base substitutions, indels, copy number variants and other structural variants affecting specific genomic elements or summaries thereof) and 8,817 transcriptomic features, corresponding to all expressed genes. d, Numeric feature values were rescaled and re-normalized (Methods), yielding a large table describing the cohort. LOH, loss of heterozygosity, RECIST, Response Evaluation Criteria, CNV, copy number variant; SV, structural variant; WGD, whole-genome doubling; CR, complete response; PR, partial response; SD, stable disease; PD, progressive disease; OS, overall survival; PFS, progression-free survival.

Fig. 2. Three latent factors associated with CPI response.

a, Logistic regression analysis (represented as a volcano plot) identified features significantly associated with CPI response (Methods and Supplementary Note 1). Dots with larger sizes represent significant features, and they are colored following the type of feature. P values shown in the plots were computed by logistic regressions. These are, by definition, two-sided. b, All significant features were selected and clustered based on their pairwise correlations. The colors denoting the clusters are inherited from the type of feature included in each of them according to the color legend in a. c, Mean expression values of cluster R3 (x-axis), and the 'T-cell effector' gene set (y-axis), across patients (dots). The Pearson's correlation coefficient is indicated. d, To discern the nature of cluster R3, the correlation of its mean to 255 gene sets collected from the literature was computed across patients (as illustrated in panel (c)). Dots represent gene sets. e, Relationship between the significance of the association with the response (y axis) and the correlation (x axis) to the mean of the cluster of the features in each cluster. P values shown in the plots were computed by logistic regressions. These are, by definition, two-sided. Dots in these three panels appear in darker color if they represent features significantly associated with CPI response and with a correlation coefficient above 0.5 with the mean of their respective cluster. In (a), (d) and (e), the horizontal dashed lines represent the significance threshold according to the Benjamini–Yekutieli correction.

Fig. 3. Two latent factors associated with survival.

a, Features significantly associated with survival residuals, that is, after correction for the three latent factors associated with response. Larger dots represent significant features. Features with high correlation (Pearson coefficient of >0.5) with any of the three previously identified latent factors are removed. P values shown in the plots were computed by Cox regressions. These are, by definition, two-sided. b, Clusters of features based on their pairwise correlations. c,d, Cluster S1 and cluster S2.1 are highly correlated with gene sets representing the tumor proliferative potential and the activity of TGF-β in the tumor microenvironment, respectively. Dots represent the mean expression of two gene sets (y axis) and the mean expression of the genes in clusters S1 and S2.1 (x axis) across patients. Pearson's correlation coefficients are indicated. e, Features significantly associated with overall survival. Larger dots represent significant features. f, Significance of the association with the response (y axis) and the correlation (x axis) to the mean of the cluster of the features in each cluster. P values shown in the plots were computed by Cox regressions. These are, by definition, two-sided. g, Depiction of the five latent factors associated with CPI response and survival. Upwards arrows, positive association with response and/or survival; downwards arrows, negative associations. In (a), (e) and (f), the horizontal dashed lines represent the significance threshold according to the Benjamini–Yekutieli correction.

Fig. 4. Validation of the latent factors across independent cohorts.

Forest plots illustrating the association of latent factors across groups of tumors with different origins in the HMF-CPI cohort (left) and across six independent cohorts (right) with CPI response and overall survival. The value of each latent factor was computed as the mean of the cluster of features obtained in the HMF-CPI cohort across each validation cohort, except in the VHIO cohort, where the transcriptomics latent factors were estimated from alternative sets of genes (Methods). In the forest plots, the dots represent the strength (coefficients estimated through multivariate logistic or Cox regression) of the association between the latent factor and response or survival across cohorts. The horizontal bars denote the 95% confidence intervals. Gray dots represent latent factors whose estimates are within one standard error at either side of 0, dots with a light color (green or red) represent non-significant associations with coefficient estimates above (or below) one standard error of 0 and dark-colored dots represent significant associations. Green dots represent positive associations with improved outcomes (higher response odds or lower hazard ratio), while red dots represent negative associations (lower response or higher hazard ratio). Mixed denotes cohorts integrated by patients with multiple tumor types.

Fig. 5. Multivariate models to predict patients’ response and survival.

a, The values of the representative biomarkers of the five latent factors across patients in the HMF-CPI cohort were used to train hybrid (pan-cancer-informed tumor type-specific) gradient-boosting models to predict CPI response and survival. The performance of the models was assessed through cross-validation (Methods and Supplementary Note 1). b, Stratifying patients based on model predictions. We separated the patients in the HMF-CPI cohort into three groups based on their predicted probability of response (histograms) and the three-segment bar below. We then calculated the fraction of responders within each group (bar plots below each histogram). c, Differences in overall survival between the three groups of patients are represented by Kaplan–Meier curves. The P value for each cohort (annotated in the plot) was calculated with a one-sided log-rank test. The line colors correspond to the three groups of patients defined in a. d, The TMB for each patient in the HMF-CPI cohort (with complete data for all five latent factors) was computed with a measure commonly used in the clinic: the number of mutations per genomic megabase. Tumors were classified as low-TMB or high-TMB based on a simple cutoff (10 mutations per megabase). The bars are colored according to the fraction of patients with high or low TMB in each of them. Interestingly, a number of patients with high-TMB tumors are predicted to have a low probability of response, whereas some patients with low-TMB tumors appear in the high probability of response group. The bottom bar plots present the percentage of patients in the low-TMB and high-TMB groups that showed clinical response to CPIs. OS, overall survival; BOR, best overall response according to RECIST; MB, megabase.

Extended Data Fig. 1. Identification of latent factors associated with CPI response and survival across the HMF-CPI cohort.

The figure provides a broad comparison of the landscape of features identified as significantly associated with CPI response (BOR), Progression Free Survival (PFS) and Overall Survival (OS) through the systematic use of univariate regression models corrected with different sets of covariables (see main manuscript and Supplementary Note 1). The three panels illustrate the results of the systematic analysis using no covariables (a), only the tissue as covariable (b), or the tissue, age, biopsy site and tumor purity as covariables (c) for the regressions. All analyses described in the main manuscript were carried out taking into account all covariables described in c. Features of different nature are colored following the same legend as in the main Figures. All p-values shown in the plots were computed via logistic (response) or Cox (survival) regressions, as in Figs. 2 and 3 of the main manuscript. These are, by definition, two-sided, denoted by positive or negative odds ratios (logistic regressions) or the reverse of hazard estimates (Cox regressions). OS: overall survival; PFS: progression-free survival; BOR: best overall response according to RECIST.

Extended Data Fig. 2. The five latent factors are integrated by highly correlated and significant features, and are mutually orthogonal.

All graphs present the relationship between the significance of the association between individual features with CPI response or survival and their correlation to the mean of the clusters of features representing each latent factor. a) TMB cluster. Features integrating this latent factor are significantly associated with CPI response and survival. b) Pretreatment cluster. Only very few features, all capturing different treatments, appear correlated with the mean of this cluster. Their association is also apparent with CPI response and survival. c) Effective T-cell infiltration cluster. Features integrating this latent factor are significantly associated with CPI response and survival. d) TGF-β activity in the microenvironment cluster. Features included in this cluster are highly correlated with the mean of the cluster, while some features included in the effective T-cell infiltration cluster show a moderate correlation (~0.5). These features are only significantly associated with CPI survival (including survival residuals), but not with response. e) Proliferative potential cluster. These features are only significantly associated with CPI survival residuals. Features of different nature are colored following the same legend as in the main Figures. All p-values shown in the plots were computed via logistic (response) or Cox (survival) regressions, as in Figs. 2 and 3 of the main manuscript. These are, by definition, two-sided. OS: overall survival; PFS: progression-free survival; BOR: best overall response according to RECIST.

Extended Data Fig. 3. Interpretation of significant expression features using genesets.

a) Heatmap representing the pairwise correlation between genesets highlighted in Fig. 2 of the main paper. b) Significance of the association of 255 genesets with CPI survival residuals and their correlation with the mean of cluster S1 (left) and S2.1 (right). Significant genesets and correlation above 0 are highlighted. c) Heatmap representing the pairwise correlations between genesets that appear significantly associated with CPI survival residuals and correlated with cluster S2.1. d) All significant features from the volcano plot represented in Fig. 3e which do not belong to any of the response clusters previously identified (TMB, T-cell effective infiltration, prior treatment) were selected and clustered based on their pairwise correlations. One large cluster (along a few unclustered features) is apparent, called cluster Survival. e) We computed the correlation of the mean value of the Survival cluster with 255 genesets. It was highly correlated with genesets representing the activity of TGF-β in the tumor microenvironment (purple dots). Other significant genesets (uncorrelated with cluster Survival) represent T-cell effective infiltration (red dots). f) Pairwise correlations between all genesets that appear significantly associated with CPI overall survival not corrected by TMB, T-cell effective infiltration and prior treatment. Two clusters are apparent. One of them represents T-cell effective infiltration. The other represents TGF-β activity in the microenvironment. P-values shown in the plot were computed via logistic (response) or Cox (survival) regressions, as in Figs. 2 and 3 of the main manuscript. These are, by definition, two-sided. OS: overall survival; PFS: progression-free survival; BOR: best overall response according to RECIST.

Extended Data Fig. 4. Association of the five latent factors with anti-cancer systemic therapies other than CPI.

Association of the five latent factors with the response to treatment (a) and overall survival (b) of patients in the HMF cohort who received CPI (left) or other therapies (right). All patients with an annotation of having received a treatment (other than CPI) for the metastatic tumor and for which an annotation of the organ of origin of the primary tumor was available were included in this group (N = 2,497). In each of the graphs the horizontal dotted line represents the threshold of statistical significance, while the vertical dotted line separates the positive (increased response or survival) and negative (decreased response or survival) effects. The association of each of the latent factors with CPI response or survival has been assessed using a univariate regression (on the values of the representative of the latent factor computed across tumors). Hence, a circle in the top right quadrant denotes a latent factor significantly associated with a positive outcome (increased response or survival); a circle in the top left quadrant represents a latent factor associated with a negative outcome (decreased response or survival). A circle in either of the two bottom quadrants represents a latent factor not significantly associated with the outcome measured. P-values shown in the plots were computed via logistic (response) or Cox (survival) regressions, using as independent variable, in each case, the estimator of each latent factor. These are, by definition, two-sided, denoted by positive or negative odds ratios (logistic regressions) or the reverse of hazard estimates (Cox regressions).

Extended Data Fig. 5. The five latent factors capture all the signal of features associated with CPI response and survival.

a) Features of different types significantly associated with CPI response or survival. The three first graphs correspond to Extended Data Figure 1C. The fourth graph presents the regression of survival residuals (that is, controlling for the features identified as associated with response) on all features. b) Volcano plots resulting from the regression analyses presented in panel A, including only features with correlation coefficient above 0.8 with the mean of any latent factor. Significant features from all regression analyses show high correlation to the clusters’ mean (as the clusters are precisely constructed from them). Other non-significant features show equally high correlation with the clusters. c) Volcano plots as in panels A and B, but showing only features with correlation coefficient below 0.3 to the mean of the clusters defining the latent factors. Only scattered features uncorrelated to the five latent factors appear significantly associated with CPI response or survival, indicating the absence of any other mutually orthogonal latent factor in the HMF-CPI cohort at the level of statistical significance set by the stringent False Discovery Rate used. Features of different nature are colored following the same legend as in the main Figures. The p-values and effect sizes shown result from logistic or Cox regressions. P-values shown in the plots were computed via logistic (response) or Cox (survival) regressions, as in Figs. 2 and 3 of the main manuscript. These are, by definition, two-sided. OS: overall survival; PFS: progression-free survival; BOR: best overall response according to RECIST.

Extended Data Fig. 6. Univariate analyses reveal the association of latent factors with CPI response across different tissues in the HMF-CPI cohort and six validation cohorts.

a) Left panel: Forest plot illustrating the association (calculated through univariate regression models) of the five latent factors with CPI response and survival across groups of patients with different types of tumors in the HMF-CPI cohort. Right panel: Idem across six validation cohorts. Red or green dots denote clear association (regression coefficients estimate more than 1 (light) / 1.96 (dark) standard errors from 0) of a latent factor with response or survival, while gray dots denote lack of association. Dark color denotes significance of the association, while light color represents non-significant associations. In the forest plots, the dots represent the strength (coefficients estimated through multivariate logistic or Cox regression) of the association between the latent factor and response or survival across cohorts. The horizontal bars across dots denote the 95% confidence intervals. Gray dots represent latent factors whose estimates are within one standard error of 0, dots with light color (green or red) represent non-significant associations with coefficient estimates above (or below) one standard error of the 0, while dark colored dots represent significant associations. Green dots represent positive associations with improved outcomes (higher response odds or lower hazard ratio), while red dots represent negative associations (lower response or higher hazard ratio). b) Stability of transcriptomics latent factors across validation cohorts. We computed the relationship between the distance of each feature to all the members of its cluster (defined in the HMF-CPI cohort) and all members of other clusters (silhouette score; Methods). The silhouette scores thus computed for genes in the TGF-beta activity in the microenvironment across HMF-CPI and four validation cohorts are represented in the first five bar plots in the top panel. Two genes, one with relatively high silhouette score, and another showing more variability across all cohorts appear highlighted. The ranks of the genes (sorted according to their silhouette scores) are aggregated across all cohorts, and a significance score (reflecting genes that are ranked consistently better than expected) is computed (right-hand bar plot). The three graphs at the bottom of the panel represent the relationship between the silhouette score of the genes in each transcriptomics latent factor in the HMF-CPI cohort (x-axis) and their aggregated score (y-axis). Sample sizes for all datasets tested can be found in Supplementary Table 1.

Extended Data Fig. 7. Relative importance of the five latent factors in the prediction of response or overall survival across patients in the HMF-CPI cohort.

The line plots represent the contribution of the values of each latent factor (Scaled feature values) across patients to the predictions cast by the response (BOR) and overall survival (OS) multivariate models. The effects are illustrated through the Shapley Values (Methods and Supplementary Note 1). Thus, in each plot, the line corresponding to each latent factor follows the relative influence of the values of the feature used to measure the latent factor on the predictions obtained through the model across all patients. Lines with positive slope correspond to latent factors that increase either the probability of response or the hazards with the increase in their value. The bar plots below the line plots represent the overall importance of each latent factor (using the standard deviation of the Shapley values) across all predictions of each model in each cohort. a) Representation of the relative importance of the latent factors in the prediction of response to CPI across the pan-cancer cohort and each tumor type separately within the HMF-CPI cohort. b) Representation of the relative importance of the latent factors in the prediction of overall survival (hazards) to CPI across the pan-cancer cohort and each tumor type separately within the HMF-CPI cohort.

Extended Data Fig. 8. Comparison of response and survival models using Shapley values.

a) Showing a comparison of response and survival hazard estimates. The points are color coded red for low responders (<10% probability response), yellow for medium responders (10-50% probability) and green for high responders(>50%). The estimates we obtained from XGboost models trained on representative biomarkers of the five latent factors across patients in the HMF-CPI cohort to predict CPI response and survival. b) Exploring the determinants of the distribution of hazards across patients with low probability of response (scatterplot). The patients in this group have been subdivided into two smaller groups based on their predicted hazard, represented by dots of different shades of red separated by the horizontal line in the value of predicted hazard 1.5. The line plots represent the distribution (quantiles) of Shapley values (see Methods) calculated for these two subgroups of patients for the five latent factors. The two lines appear more separated in the distributions of Shapley values of tumor proliferative potential and TGF-beta activity in the microenvironment. This indicates that it is the values of these two latent factors that contribute the most to the separation between these two groups of patients. c) Example of the predicted CPI response and survival of one patient in the HMF-CPI cohort broken down by Shapley values.

Extended Data Fig. 9. Stratification of patients in validation cohorts using multivariate machine learning models.

a) The histograms represent the distribution of the probability of response to CPI of patients across three of the validation cohorts (those with complete data on all five latent factors), either combined or separate. The bars are colored red (probability of response below 0.1, low), yellow (probability between 0.1 and 0.5, medium) or green (probability above 0.5, high). The absolute number of patients across the three cohorts in each group (low, medium, high) are shown in the horizontal bar below the combined histogram. The barplots below present the percentage of patients in each of the groups who actually showed response to CPI according to the data of each cohort. b) Top panel: Kaplan-Meier curves resulting from the aforementioned stratification of patients across the three cohorts, either combined or separate. Bottom panel: Kaplan-Meier curves resulting from stratifying the patients across the three cohorts based on their predicted probability of survival according to the hybrid models trained on survival data. The p-value for each cohort (annotated in the plot) was calculated via a one-sided logrank test.

Extended Data Fig. 10. Application of multivariate machine learning models to identify patients with high probability to respond to CPI across the entire HMF cohort.

Bars represent the number of patients with metastatic tumors from different sites of origin in the HMF cohort who received (top) or did not receive (bottom) CPI as treatment. All patients with an annotation of having received a treatment (other than CPI) for the metastatic tumor and for which an annotation of the organ of origin of the primary tumor was available were included in this group (N = 2,497). The colored segments in the bars at the left represent the absolute number of patients with low (below 0.1), medium (between 0.1 and 0.5) or high (above 0.5) predicted probability of response. These bars have been separated based on the total number of patients of each tumor type, and x-axes representing the relative scales of each plot have been added. To the right side of the plot, the percentage of patients of each tumor type including more than 15 cases are represented as stacked bar plots, to facilitate comparability between tumor types. An important fraction of patients with tumors from the same origin as those in the HMF-CPI cohort (for example, in the lung) present high predicted probability of response to CPI. Interestingly, patients with tumors of other origins, who are not typically considered as candidates for CPI treatment also exhibit high predicted response probability.