Despite impressive durable responses elicited by cancer immunotherapy, the majority of patients do not see long-term benefit with treatment [1, 2]. However, the molecular mechanisms that determine therapeutic resistance remain poorly understood, particularly genetic interactions. To systematically interrogate such genetic interactions that mediate immune resistance, we designed a Combinatorial Antineoplastic Drug Resistance Experiment (CADRE) screening strategy with an asymmetric library design (Fig. 1A, B). The CADRE library was synthesized via oligonucleotide array and cloned into lentiviral vectors (Additional file 1: Fig. S1A), and the representations of double knockouts (DKOs), single knockouts (SKOs), and double non-targeting controls (DNTCs) in the library were verified by next-generation sequencing (NGS) (Additional file 1: Fig. S1C, D). We transduced B16F10;OVA;Cas9 cells at a low multiplicity of infection (MOI) (MOI < 0.2) at a coverage of approximately 500X. We NGS verified that the transduced pre-selection cell pool retained the vast majority of the CADRE library (Additional file 1: Fig. S1E).
We then performed co-culture assays on library and non-library infected B16F10;OVA;Cas9 clone #3 cells (BC3 cells) with OT-I CD8 + T cells. Both library and non-library transduced cells showed comparable survival across E:T ratios, with the exception of at the high E:T ratio conditions (E:T ratios > 1) where the mutant pool demonstrated a significant increase in resistance (Fig. 1C). We then performed the co-culture screen with BC3-CADRE cells and OT-I CD8 + T-cells, followed by library NGS readout (Additional file 1: Fig. S2A). Clustering analysis showed distinct clusters between plasmid, cell populations before co-culture, and cell populations post co-culture (Additional file 1: Fig. S2B), suggesting a high-quality screen and NGS readout between the cell pool conditions and E:T ratios 2–5 (Fig. 1D). There are strong shifts between pre-selection and post-selection co-cultures (Additional file 1: Fig. S2C), indicative of strong selection seen at sgRNA library levels.
At a false-discovery rate (FDR) of 1.19% we identified 222 enriched sgRNA pairs of which 194 (87.4%) are associated with Janus kinase 1 (Jak1) or Janus kinase 2 (Jak2), including DKO and SKO constructs. Bulk analysis revealed that Jak-associated sgRNAs dominated the enrichment in the screen post-selection (Fig. 1E; Additional file 1: Fig. S3A–D). We found that Jak1/2-associated gene pairs were the most statistically significantly different from their constitutive SKOs (Fig. 1E, F; Additional file 1: Fig. S3E, F), suggestive of potential gene interactions. We observe that gene pairs Jak1_Trp53, Jak1_Nf1, and Jak1_Rb1 have higher observed enrichment for double knockout than expected (adjusted p-value < 0.001) suggesting potential additive gene interaction (Fig. 1F), while gene pairs Jak1_Apc, Jak1_Vhl, Jak1_Kmt2c, Jak1_Kmt2d, Jak1_Arid1a, Jak1_Fbxw7, Jak1_Ctnnb1 have lower observed enrichment for double knockout than expected, suggesting potential subtractive gene interaction (Fig. 1F). Boxplots of normalized read counts for Jak1_Kmt2d, Jak2_Kmt2d, Jak1_Trp53 and Jak2_Trp53 (Fig. 2A–D) also suggest potential subtractive and additive phenotypic interactions to Jak1/2 perturbation for Kmt2d and Trp53, respectively. However, it should be noted that although significant, the putative gene interaction signals appear to be modest in part due to the strong resistance phenotype of single knockout of JAK1/2.
We looked at the global gene expression profiles of KMT2D, JAK1, TP53, and IFNGR1 across all tumor samples and paired normal tissues (Additional file 1: Fig. S4A–D) and more specifically for KMT2D and JAK1 in the SKCM cohort (Fig. 2E–G) and identified tumor-type specific expression patterns. We found that the effector and exhaustion T cell signatures were upregulated in the tumor samples in melanoma patients (Fig. 2G). Cell proportion deconvolution analyses revealed increased estimated proportions of CD8 T cells, memory-activated CD4 T cells, and Tregs in the tumor samples, with a decrease of naïve and memory resting CD4 T cells (Fig. 2H).
Genes negatively correlated with KMT2D were further analyzed using DAVID gene ontology functional annotation (Fig. 2F). We found positive and significant correlations for both JAK1 and IFN-gamma signaling gene signatures across both the SKCM cohort and across 33 different cancer types from TCGA (Additional file 1: Fig. S5A). KMT2D and JAK1 are both frequently mutated in melanoma patients (Additional file 1: Fig. S5B). Mutual exclusivity and co-occurrence analyses for all pairwise combinations of KMT2D, JAK1, JAK2, IFNGR1, and TP53 suggest that all mutation combinations except JAK2-IFNGR1 co-occur at a significant rate (Additional file 1: Fig. S5C).
We also performed survival analyses on patient cohorts with the public database TCGA (The Cancer Genome Atlas) (Additional file 1: Figs. S4E, F, S5D, E). Survival maps (Additional file 1: Fig. S4G, H) revealed cancer-type specific effects of KMT2D, JAK1, JAK2, IFNGR1, or TP53 expression levels on patient survival. The KMT2D-low patient group demonstrated increased CTL-associated overall survival benefit, whereas high levels of KMT2D abolished the overall survival benefit of CTL-high patients (Fig. 2I).
Altogether, we demonstrate how dual loss-of-function CRISPR screens with asymmetric library designs can resolve complex phenotypes such as resistance to T cell killing.