Dynamic metrics-based biomarkers to predict responders to anti-PD-1 immunotherapy

Abstract

Background: Anti-PD-1 immunotherapies have shown clinical benefit in multiple cancers, but response was only observed in a subset of patients. Predicting which patients will respond is an urgent clinical need, but current companion diagnosis based on PD-L1 IHC staining shows limited predictability.

Methods: A dynamic, metrics-based biomarker was developed to discriminate responders from non-responders for anti-PD-1 immunotherapy in B16F10 melanoma-bearing mice.

Results: Similar to patients, there was considerable heterogeneity in response to anti-PD-1 immunotherapy in mice. Compared with the control group, 45% of anti-PD-1 antibody-treated mice displayed improved survival (defined as responders) and the remainder only gained little, if any, survival benefit from PD-1 blockade (non-responders). Interestingly, the dynamics of IFN-γ secretion by peripheral lymphocytes was associated with faster secretion onset (shorter lag time), stronger exponential phase, shorter time to half magnitude, and higher magnitude of secretion in responders at day 10 after tumour inoculation. To sufficiently predict responders from non-responders, IFN-γ secretion descriptors as well as phenotypic markers were subjected to multivariate analysis using orthogonal partial least-squares discriminant analysis (OPLS-DA).

Conclusions: By integrating phenotypic markers, IFN-γ secretion descriptors sufficiently predict response to anti-PD-1 immunotherapy. Such a dynamic, metrics-based biomarker holds high diagnostic potential for anti-PD-1 checkpoint immunotherapy.

Fig. 1

a Schematic diagram of immunotherapy regimen and sampling strategy on mice challenged with 5 × 10⁴ B16F10 cells. b Scanning cut-off for stratification of responders and non-responders to anti-PD-1 treatment. Histogram indicates individual survival time following anti-PD-1 treatment. Survival differences between non-responder and untreated groups were analysed by Log-rank test. P-value at each step was showed as line chart. Red dot indicates the best cut-off. c Kaplan–Meier survival curves of mice according to different treatment (cIg vs treatment). Responder and non-responder are sub-groups of PD-1 treatment. Log-rank test was used to compare survival curves. d Tumour growth following B16F10 inoculation. Tumour sizes are individually represented over time

Fig. 2

Cell phenotyping on peripheral lymphocytes at day 3 and day 10. a, b Per cent of CD4+ (a) and CD8+ (b) T effector cells of CD45.2+ cells. c, d Frequency of PD-1+ subset in CD4+ (c) and CD8+ (d) T cells. e Per cent of CD25+Foxp3+ regulatory T cells (Treg) of CD45.2+ cells. f CD8+ T cells to Treg ratio. g Frequency of MDSCs. Values shown are for individually analysed mice. Data are analysed by Wilcoxon–Mann–Whitney test (*p < 0.05)

Fig. 3

Graphical representation of IFN-γ secretion descriptors derived from simulated curves. C_max, Tc50, Hill exponent (h), and τ (tau) indicate magnitude, time to reach 50% C_max, slope, and lag time, respectively

Fig. 4

Dynamic IFN-γ secretion from peripheral CD4+ and CD8+ lymphocytes. a–d Simulated IFN-γ secretion based on Sigmoid E_max model from CD4+ (a, b) and CD8+ (c, d) T cells at day 3 (a, c) and day 10 (b, d). Each line represents individual mouse. Dynamic IFN-γ secretion was determined at day 3 and day 10. Supernatant IFN-γ was quantified by ELISA at 1, 2, 3, 4, 6, 8, 12, 24, and 40 h following T cell activation. e–l Model-based secretion descriptors from CD4+ (e–h) and CD8+ (i–l) T cells addressing magnitude (C_max) (e, i), time to reach 50% C_max (Tc50) (f, j), slope (h) (g, k), and lag time (tau) (h, l) of dynamic IFN-γ secretion

Fig. 5

OPLS-DA model outcomes for distinguishing between immune responses. OPLS-DA models were established based on only phenotypic markers (a, d), only secretion descriptors (b, e), and all variables (c, f) at day 3 (a–c) and day 10 (d–f). The variable changes from day 3 to day 10 were also used as predicting variables (g). All variables at day 3 and day 10 were finally integrated (h). The model was generated using one predictive component (t1) and the first orthogonal component (to1). Each dot stands for OPLS-DA score of individual mouse. Black hoteling ellipses defined global confidence limits. The values of R²X, R²Y, Q²Y, and RMSEE were shown under each plot. R²X and R²Y: percentage of predictor and response variance explained by the model. Q²Y: predictive performance of the model estimated by cross-validation. RMSEE: root mean square error of estimation. Per cent in colour (only shown when Q²Y is positive) indicates the accurate prediction of responder and non-responder by the model

Fig. 6

Prediction performance of OPLS-DA models at day 10. Internal validation was conducted on training (a) and test (b) subset by re-trained all-variable based model at day 10 using training subset. External validation (c) was conducted on an independent dataset. Prediction performance of all-variable (d), secretion descriptors (e) and phenotypic markers (f) based multivariate analysis at day 10 were compared. Prediction performance derived from representative univariate analysis (g–n) were also shown to compare with multivariate analysis. g–n described univariate analysis of CD4Cmax, CD4h, CD4tau (secretion descriptors of Cmax, h and tau derived from CD4+ lymphocytes), CD8Tc50, CD8tau (secretion descriptors of Tc50 and tau derived from CD8+ lymphocytes), CD8+, PD1+CD4+, MDSC (cell density of CD8+, PD1+% of CD4+ cells, cell density of MDSC in peripheral blood) respectively. Obs. R and Obs. N represent observed responders and non-responders respectively. Pred. R and Pred. N represent predicted responders and non-responders respectively. The diagonal in each panel represents conditions under which the prediction was matched with the observation. Colors indicate the percent (%) of samples that met a given condition. Left-top panel showing an example of prediction with 100% accuracy