Search

Summary: A major challenge in cancer treatment is predicting clinical response to anti-cancer drugs on a personalized basis. Using a pharmacogenomics database of 1,001 cancer cell lines, we trained deep neural networks for prediction of drug response and assessed their performance on multiple clinical cohorts. We demonstrate that deep neural networks outperform the current state in machine learning frameworks. We provide a proof of concept for the use of deep neural network-based frameworks to aid precision oncology strategies. : Sakellaropoulos et al. designed a machine learning workflow to predict drug response and survival of cancer patients. All pipelines are trained on a large panel of cancer cell lines and tested in clinical cohorts. DNN outperforms other machine learning algorithms by capturing pathways that link gene expression with drug response. Keywords: drug response prediction, precision medicine, machine learning, deep neural networks, DNN

Although B-cell acute lymphoblastic leukemia (B-ALL) is the most common malignancy in children and while highly curable, it remains a leading cause of cancer-related mortality. The outgrowth of tumor subclones carrying mutations in genes responsible for resistance to therapy has led to a Darwinian model of clonal selection. Previous work has indicated that alterations in the epigenome might contribute to clonal selection, yet the extent to which the chromatin state is altered under the selective pressures of therapy is unknown. To address this, we performed chromatin immunoprecipitation, gene expression analysis, and enhanced reduced representation bisulfite sequencing on a cohort of paired diagnosis and relapse samples from individual patients who all but one relapsed within 36 months of initial diagnosis. The chromatin state at diagnosis varied widely among patients, while the majority of peaks remained stable between diagnosis and relapse. Yet a significant fraction was either lost or newly gained, with some patients showing few differences and others showing massive changes of the epigenetic state. Evolution of the epigenome was associated with pathways previously linked to therapy resistance as well as novel candidate pathways through alterations in pyrimidine biosynthesis and downregulation of polycomb repressive complex 2 targets. Three novel, relapse-specific superenhancers were shared by a majority of patients including one associated with S100A8, the top upregulated gene seen at relapse in childhood B-ALL. Overall, our results support a role of the epigenome in clonal evolution and uncover new candidate pathways associated with relapse. Significance: This study suggests a major role for epigenetic mechanisms in driving clonal evolution in B-ALL and identifies novel pathways associated with drug resistance.

Planar cell polarity (PCP) and neural tube defects (NTDs) are linked, with a subset of NTD patients found to harbor mutations in PCP genes, but there is limited data on whether these mutations disrupt PCP signaling in vivo. The core PCP geneVan Gogh(Vang),Vangl1/2 in mammals, is the most specific for PCP. We thus addressed potential causality of NTD-associatedVangl1/2 mutations, from either mouse or human patients, inDrosophilaallowing intricate analysis of the PCP pathway. Introducing the respective mammalian mutations intoDrosophila Vangrevealed defective phenotypic and functional behaviors, with changes to Vang localization, post-translational modification, and mechanistic function, such as its ability to interact with PCP effectors. Our findings provide mechanistic insight into how different mammalian mutations contribute to developmental disorders and strengthen the link between PCP and NTD. Importantly, analyses of the human mutations revealed that each is a causative factor for the associated NTD.

The NSD2 p.E1099K (EK) mutation is observed in 10% of acute lymphoblastic leukemia (ALL) samples with enrichment at relapse indicating a role in clonal evolution and drug resistance. To discover mechanisms that mediate clonal expansion, we engineered B-precursor ALL (B-ALL) cell lines (Reh, 697) to overexpress wildtype (WT) and EK NSD2, but observed no differences in proliferation, clonal growth, or chemosensitivity. To address whether NSD2 EK acts collaboratively with other pathways, we used short hairpin RNAs to knockdown expression of NSD2 in B-ALL cell lines heterozygous for NSD2 EK (RS4;11, RCH-ACV, SEM). Knockdown resulted in decreased proliferation in all lines, decreased clonal growth in RCH-ACV, and increased sensitivity to cytotoxic chemotherapeutic agents, although the pattern of drug sensitivity varied among cell lines implying that the oncogenic properties of NSD2 mutations are likely cell context specific and rely on cooperative pathways. Knockdown of both Type II and REIIBP EK isoforms had a greater impact than knockdown of Type II alone, suggesting that both SET containing EK isoforms contribute to phenotypic changes driving relapse. Furthermore, in vivo models using both cell lines and patient samples revealed dramatically enhanced proliferation of NSD2 EK compared with WT and reduced sensitivity to 6-mercaptopurine in the relapse sample relative to diagnosis. Finally, EK-mediated changes in chromatin state and transcriptional output differed dramatically among cell lines further supporting a cell context–specific role of NSD2 EK. These results demonstrate a unique role of NSD2 EK in mediating clonal fitness through pleiotropic mechanisms dependent on the genetic and epigenetic landscape. Implications: NSD2 EK mutation leads to drug resistance and a clonal advantage in childhood B-ALL.

A major challenge in cancer treatment is predicting the clinical response to anti-cancer drugs on a personalized basis. The success of such a task largely depends on the ability to develop computational resources that integrate big "omic" data into effective drug-response models. Machine learning is both an expanding and an evolving computational field that holds promise to cover such needs. Here we provide a focused overview of: 1) the various supervised and unsupervised algorithms used specifically in drug response prediction applications, 2) the strategies employed to develop these algorithms into applicable models, 3) data resources that are fed into these frameworks and 4) pitfalls and challenges to maximize model performance. In this context we also describe a novel in silico screening process, based on Association Rule Mining, for identifying genes as candidate drivers of drug response and compare it with relevant data mining frameworks, for which we generated a web application freely available at: https://compbio.nyumc.org/drugs/. This pipeline explores with high efficiency large sample-spaces, while is able to detect low frequency events and evaluate statistical significance even in the multidimensional space, presenting the results in the form of easily interpretable rules. We conclude with future prospects and challenges of applying machine learning based drug response prediction in precision medicine.

The BCL2 family plays important roles in acute myeloid leukemia (AML). Venetoclax, a selective BCL2 inhibitor, has received FDA approval for the treatment of AML. However, drug resistance ensues after prolonged treatment, highlighting the need for a greater understanding of the underlying mechanisms. Using a genome-wide CRISPR/Cas9 screen in human AML, we identified genes whose inactivation sensitizes AML blasts to venetoclax. Genes involved in mitochondrial organization and function were significantly depleted throughout our screen, including the mitochondrial chaperonin CLPB. We demonstrated that CLPB is upregulated in human AML, it is further induced upon acquisition of venetoclax resistance, and its ablation sensitizes AML to venetoclax. Mechanistically, CLPB maintains the mitochondrial cristae structure via its interaction with the cristae-shaping protein OPA1, whereas its loss promotes apoptosis by inducing cristae remodeling and mitochondrial stress responses. Overall, our data suggest that targeting mitochondrial architecture may provide a promising approach to circumvent venetoclax resistance. Significance: A genome-wide CRISPR/Cas9 screen reveals genes involved in mitochondrial biological processes participate in the acquisition of venetoclax resistance. Loss of the mitochondrial protein CLPB leads to structural and functional defects of mitochondria, hence sensitizing AML cells to apoptosis. Targeting CLPB synergizes with venetoclax and the venetoclax/azacitidine combination in AML in a p53-independent manner. See related commentary by Savona and Rathmell, p. 831. This article is highlighted in the In This Issue feature, p. 813