The determination of molecular features that mediate clinically aggressive phenotypes in prostate cancer remains a major biological and clinical challenge^1,2. Recent advances in interpretability of machine learning models as applied to biomedical problems may enable discovery and prediction in clinical cancer genomics^3–5. Here we developed P-NET—a biologically informed deep learning model—to stratify patients with prostate cancer by treatment-resistance state and evaluate molecular drivers of treatment resistance for therapeutic targeting through complete model interpretability. We demonstrate that P-NET can predict cancer state using molecular data with a performance that is superior to other modelling approaches. Moreover, the biological interpretability within P-NET revealed established and novel molecularly altered candidates, such as MDM4 and FGFR1, which were implicated in predicting advanced disease and validated in vitro. Broadly, biologically informed fully interpretable neural networks enable preclinical discovery and clinical prediction in prostate cancer and may have general applicability across cancer types.

Allogeneic natural killer (NK) cell adoptive transfer has shown the potential to induce remissions in relapsed or refractory leukemias and lymphomas, but strategies to enhance NK cell survival and function are needed to improve clinical efficacy. Here, we demonstrated that NK cells cultured ex vivo with interleukin-15 (IL-15) and nicotinamide (NAM) exhibited stable induction of l -selectin (CD62L), a lymphocyte adhesion molecule important for lymph node homing. High frequencies of CD62L were associated with elevated transcription factor forkhead box O1 (FOXO1), and NAM promoted the stability of FOXO1 by preventing proteasomal degradation. NK cells cultured with NAM exhibited metabolic changes associated with elevated glucose flux and protection against oxidative stress. NK cells incubated with NAM also displayed enhanced cytotoxicity and inflammatory cytokine production and preferentially persisted in xenogeneic adoptive transfer experiments. We also conducted a first-in-human phase 1 clinical trial testing adoptive transfer of NK cells expanded ex vivo with IL-15 and NAM (GDA-201) combined with monoclonal antibodies in patients with relapsed or refractory non-Hodgkin lymphoma (NHL) and multiple myeloma (MM) (NCT03019666). Cellular therapy with GDA-201 and rituximab was well tolerated and yielded an overall response rate of 74% in 19 patients with advanced NHL. Thirteen patients had a complete response, and 1 patient had a partial response. GDA-201 cells were detected for up to 14 days in blood, bone marrow, and tumor tissues and maintained a favorable metabolic profile. The safety and efficacy of GDA-201 in this study support further development as a cancer therapy.

Bladder cancer is a genetically heterogeneous disease, and novel therapeutic strategies are needed to expand treatment options and improve clinical outcomes. Here, we identified a unique subset of urothelial tumors with focal amplification of the RAF1 (CRAF) kinase gene. RAF1-amplified tumors had activation of the RAF/MEK/ERK signaling pathway and exhibited a luminal gene expression pattern. Genetic studies demonstrated that RAF1-amplified tumors were dependent upon RAF1 activity for survival, and RAF1-activated cell lines and patient-derived models were sensitive to available and emerging RAF inhibitors as well as combined RAF plus MEK inhibition. Furthermore, we found that bladder tumors with HRAS- or NRAS-activating mutations were dependent on RAF1-mediated signaling and were sensitive to RAF1-targeted therapy. Together, these data identified RAF1 activation as a dependency in a subset making up nearly 20% of urothelial tumors and suggested that targeting RAF1-mediated signaling represents a rational therapeutic strategy.

B7-H3 (CD276) is a transmembrane glycoprotein of the B7 immune checkpoint superfamily that has emerged as a promising therapeutic target. To better understand the applicability of B7-H3–directed therapies, we analyzed 156,791 samples comprising 50 cancer types to interrogate the clinical, genomic, transcriptomic, and immunologic correlates of B7-H3 mRNA expression. DNA (592-gene/whole-exome) and RNA (whole-transcriptome) sequencing was performed from samples submitted to Caris Life Sciences. B7-H3 high versus low expression was based on top and bottom quartiles for each cancer type. Patients’ overall survival was determined from insurance claims data. Pathway analysis was performed using gene set enrichment analyses. Immune cell fractions were inferred using quanTIseq. B7-H3 is expressed across several human malignancies including prostate, pancreatic, ovarian, and lung cancers. High B7-H3 expression is associated with differences in overall survival, possibly indicating a prognostic role of B7-H3 for some cancers. When examining molecular features across all cancer types, we did not identify recurrent associations between B7-H3 expression and genetic alterations in TP53, RB1, and KRAS. However, we find consistent enrichment of epithelial-to-mesenchymal transition, Wnt, TGFβ, and Notch signaling pathways. In addition, tumors with high B7-H3 expression are associated with greater proportions of M1 macrophages, but lower fractions of CD8+ T cells. We have begun to define the genomic, transcriptomic, clinical, and immunologic features associated with B7-H3 expression in 50 cancer types. We report novel clinical and molecular features of B7-H3–high tumors which may inform how current B7-H3 therapeutics should be deployed and prioritized. Significance: B7-H3–targeting therapeutics have shown promising results in initial clinical trials. In this pan-cancer analysis of B7-H3 mRNA expression, we found that B7-H3 exhibits robust expression in many common cancer types. These results may inform further development of B7-H3–targeting therapeutics and may guide clinical decisions for patients with limited treatment options.

Collateral lethality occurs when loss of a gene/protein renders cancer cells dependent on its remaining paralog. Combining genome-scale CRISPR/Cas9 loss-of-function screens with RNA sequencing in over 900 cancer cell lines, we found that cancers of nervous system lineage, including adult and pediatric gliomas and neuroblastomas, required the nuclear kinase vaccinia-related kinase 1 (VRK1) for their survival in vivo. VRK1 dependency was inversely correlated with expression of its paralog VRK2 . VRK2 knockout sensitized cells to VRK1 loss, and conversely, VRK2 overexpression increased cell fitness in the setting of VRK1 loss. DNA methylation of the VRK2 promoter was associated with low VRK2 expression in human neuroblastomas and adult and pediatric gliomas. Mechanistically, depletion of VRK1 reduced barrier-to-autointegration factor phosphorylation during mitosis, resulting in DNA damage and apoptosis. Together, these studies identify VRK1 as a synthetic lethal target in VRK2 promoter–methylated adult and pediatric gliomas and neuroblastomas.

Sarcomere and cytoskeleton genes, or actomyosin genes, regulate cell biology including mechanical stress, cell motility, and cell division. While actomyosin genes are recurrently dysregulated in cancers, their oncogenic roles have not been examined in a lineage-specific fashion. In this report, we investigated dysregulation of nine sarcomeric and cytoskeletal genes across 20 cancer lineages. We found that uterine cancers harbored the highest frequencies of amplification and overexpression of the gamma actin gene, ACTG1. Each of the four subtypes of uterine cancers, mixed endometrial carcinomas, serous carcinomas, endometroid carcinomas, and carcinosarcomas harbored between 5~20% of ACTG1 gene amplification or overexpression. Clinically, patients with ACTG1 gains had a poor prognosis. ACTG1 gains showed transcriptional patterns that reflect activation of oncogenic signals, repressed response to innate immunity, or immunotherapy. Functionally, the CRISPR-CAS9 gene deletion of ACTG1 had the most robust and consistent effects in uterine cancer cells relative to 20 other lineages. Overall, we propose that ACTG1 regulates the fitness of uterine cancer cells by modulating cell-intrinsic properties and the tumor microenvironment. In summary, the ACTG1 functions relative to other actomyosin genes support the notion that it is a potential biomarker and a target gene in uterine cancer precision therapies.