Artificial Intelligence in Cancer Research and Precision Medicine.

Artificial intelligence (AI) is rapidly reshaping cancer research and personalized clinical care. Availability of high-dimensionality datasets coupled with advances in high-performance computing, as well as innovative deep learning architectures, has led to an explosion of AI use in various aspects of oncology research. These applications range from detection and classification of cancer, to molecular characterization of tumors and their microenvironment, to drug discovery and repurposing, to predicting treatment outcomes for patients. As these advances start penetrating the clinic, we foresee a shifting paradigm in cancer care becoming strongly driven by AI. SIGNIFICANCE: AI has the potential to dramatically affect nearly all aspects of oncology-from enhancing diagnosis to personalizing treatment and discovering novel anticancer drugs. Here, we review the recent enormous progress in the application of AI to oncology, highlight limitations and pitfalls, and chart a path for adoption of AI in the cancer clinic.

A Bayesian machine learning approach for drug target identification using diverse data types.

Drug target identification is a crucial step in development, yet is also among the most complex. To address this, we develop BANDIT, a Bayesian machine-learning approach that integrates multiple data types to predict drug binding targets. Integrating public data, BANDIT benchmarked a ~90% accuracy on 2000+ small molecules. Applied to 14,000+ compounds without known targets, BANDIT generated ~4,000 previously unknown molecule-target predictions. From this set we validate 14 novel microtubule inhibitors, including 3 with activity on resistant cancer cells. We applied BANDIT to ONC201-an anti-cancer compound in clinical development whose target had remained elusive. We identified and validated DRD2 as ONC201's target, and this information is now being used for precise clinical trial design. Finally, BANDIT identifies connections between different drug classes, elucidating previously unexplained clinical observations and suggesting new drug repositioning opportunities. Overall, BANDIT represents an efficient and accurate platform to accelerate drug discovery and direct clinical application.

Dopamine Receptor D5 is a Modulator of Tumor Response to Dopamine Receptor D2 Antagonism.

PURPOSE: Dopamine receptor D2 (DRD2) is a G protein-coupled receptor antagonized by ONC201, an anticancer small molecule in clinical trials for high-grade gliomas and other malignancies. DRD5 is a dopamine receptor family member that opposes DRD2 signaling. We investigated the expression of these dopamine receptors in cancer and their influence on tumor cell sensitivity to ONC201. EXPERIMENTAL DESIGN: The Cancer Genome Atlas was used to determine DRD2/DRD5 expression broadly across human cancers. Cell viability assays were performed with ONC201 in >1,000 Genomic of Drug Sensitivity in Cancer and NCI60 cell lines. IHC staining of DRD2/DRD5 was performed on tissue microarrays and archival tumor tissues of glioblastoma patients treated with ONC201. Whole exome sequencing was performed in RKO cells with and without acquired ONC201 resistance. Wild-type and mutant DRD5 constructs were generated for overexpression studies. RESULTS: DRD2 overexpression broadly occurs across tumor types and is associated with a poor prognosis. Whole exome sequencing of cancer cells with acquired resistance to ONC201 revealed a de novo Q366R mutation in the DRD5 gene. Expression of Q366R DRD5 was sufficient to induce tumor cell apoptosis, consistent with a gain-of-function. DRD5 overexpression in glioblastoma cells enhanced DRD2/DRD5 heterodimers and DRD5 expression was inversely correlated with innate tumor cell sensitivity to ONC201. Investigation of archival tumor samples from patients with recurrent glioblastoma treated with ONC201 revealed that low DRD5 expression was associated with relatively superior clinical outcomes. CONCLUSIONS: These results implicate DRD5 as a negative regulator of DRD2 signaling and tumor sensitivity to ONC201 DRD2 antagonism.

Bioinformatics Approaches to Predict Drug Responses from Genomic Sequencing.

Fulfilling the promises of precision medicine will depend on our ability to create patient-specific treatment regimens. Therefore, being able to translate genomic sequencing into predicting how a patient will respond to a given drug is critical. In this chapter, we review common bioinformatics approaches that aim to use sequencing data to predict sample-specific drug susceptibility. First, we explain the importance of customized drug regimens to the future of medical care. Second, we discuss the different public databases and community efforts that can be leveraged to develop new methods for identifying new predictive biomarkers. Third, we cover the basic methods that are currently used to identify markers or signatures of drug response, without any prior knowledge of the drug's mechanism of action. We further discuss how one can integrate knowledge about drug targets, mechanisms, and predictive markers to better estimate drug response in a diverse set of samples. We begin this section with a primer on popular methods to identify targets and mechanism of action for new small molecules. This discussion also includes a set of computational methods that incorporate other drug features, which do not relate to drug-induced genetic changes or sequencing data such as drug structures, side-effects, and efficacy profiles. Those additional drug properties can aid in gaining higher accuracy for the identification of drug target and mechanism of action. We then progress to discuss using these targets in combination with disease-specific expression patterns, known pathways, and genetic interaction networks to aid drug choice. Finally, we conclude this chapter with a general overview of machine learning methods that can integrate multiple pieces of sequencing data along with prior drug or biological knowledge to drastically improve response prediction.

A Predictive Model for Selective Targeting of the Warburg Effect through GAPDH Inhibition with a Natural Product.

Targeted cancer therapies that use genetics are successful, but principles for selectively targeting tumor metabolism that is also dependent on the environment remain unknown. We now show that differences in rate-controlling enzymes during the Warburg effect (WE), the most prominent hallmark of cancer cell metabolism, can be used to predict a response to targeting glucose metabolism. We establish a natural product, koningic acid (KA), to be a selective inhibitor of GAPDH, an enzyme we characterize to have differential control properties over metabolism during the WE. With machine learning and integrated pharmacogenomics and metabolomics, we demonstrate that KA efficacy is not determined by the status of individual genes, but by the quantitative extent of the WE, leading to a therapeutic window in vivo. Thus, the basis of targeting the WE can be encoded by molecular principles that extend beyond the status of individual genes.

Cancer stem cell-related gene expression as a potential biomarker of response for first-in-class imipridone ONC201 in solid tumors.

Cancer stem cells (CSCs) correlate with recurrence, metastasis and poor survival in clinical studies. Encouraging results from clinical trials of CSC inhibitors have further validated CSCs as therapeutic targets. ONC201 is a first-in-class small molecule imipridone in Phase I/II clinical trials for advanced cancer. We have previously shown that ONC201 targets self-renewing, chemotherapy-resistant colorectal CSCs via Akt/ERK inhibition and DR5/TRAIL induction. In this study, we demonstrate that the anti-CSC effects of ONC201 involve early changes in stem cell-related gene expression prior to tumor cell death induction. A targeted network analysis of gene expression profiles in colorectal cancer cells revealed that ONC201 downregulates stem cell pathways such as Wnt signaling and modulates genes (ID1, ID2, ID3 and ALDH7A1) known to regulate self-renewal in colorectal, prostate cancer and glioblastoma. ONC201-mediated changes in CSC-related gene expression were validated at the RNA and protein level for each tumor type. Accordingly, we observed inhibition of self-renewal and CSC markers in prostate cancer cell lines and patient-derived glioblastoma cells upon ONC201 treatment. Interestingly, ONC201-mediated CSC depletion does not occur in colorectal cancer cells with acquired resistance to ONC201. Finally, we observed that basal expression of CSC-related genes (ID1, CD44, HES7 and TCF3) significantly correlate with ONC201 efficacy in >1000 cancer cell lines and combining the expression of multiple genes leads to a stronger overall prediction. These proof-of-concept studies provide a rationale for testing CSC expression at the RNA and protein level as a predictive and pharmacodynamic biomarker of ONC201 response in ongoing clinical studies.

A Data-Driven Approach to Predicting Successes and Failures of Clinical Trials.

Over the past decade, the rate of drug attrition due to clinical trial failures has risen substantially. Unfortunately it is difficult to identify compounds that have unfavorable toxicity properties before conducting clinical trials. Inspired by the effective use of sabermetrics in predicting successful baseball players, we sought to use a similar "moneyball" approach that analyzes overlooked features to predict clinical toxicity. We introduce a new data-driven approach (PrOCTOR) that directly predicts the likelihood of toxicity in clinical trials. PrOCTOR integrates the properties of a compound's targets and its structure to provide a new measure, the PrOCTOR score. Drug target network connectivity and expression levels, along with molecular weight, were identified as important indicators of adverse clinical events. Our method provides a data-driven, broadly applicable strategy to identify drugs likely to possess manageable toxicity in clinical trials and will help drive the design of therapeutic agents with less toxicity.

Prediction of Genetic Interactions Using Machine Learning and Network Properties.

A genetic interaction (GI) is a type of interaction where the effect of one gene is modified by the effect of one or several other genes. These interactions are important for delineating functional relationships among genes and their corresponding proteins, as well as elucidating complex biological processes and diseases. An important type of GI - synthetic sickness or synthetic lethality - involves two or more genes, where the loss of either gene alone has little impact on cell viability, but the combined loss of all genes leads to a severe decrease in fitness (sickness) or cell death (lethality). The identification of GIs is an important problem for it can help delineate pathways, protein complexes, and regulatory dependencies. Synthetic lethal interactions have important clinical and biological significance, such as providing therapeutically exploitable weaknesses in tumors. While near systematic high-content screening for GIs is possible in single cell organisms such as yeast, the systematic discovery of GIs is extremely difficult in mammalian cells. Therefore, there is a great need for computational approaches to reliably predict GIs, including synthetic lethal interactions, in these organisms. Here, we review the state-of-the-art approaches, strategies, and rigorous evaluation methods for learning and predicting GIs, both under general (healthy/standard laboratory) conditions and under specific contexts, such as diseases.

Organization of enzyme concentration across the metabolic network in cancer cells.

Rapid advances in mass spectrometry have allowed for estimates of absolute concentrations across entire proteomes, permitting the interrogation of many important biological questions. Here, we focus on a quantitative aspect of human cancer cell metabolism that has been limited by a paucity of available data on the abundance of metabolic enzymes. We integrate data from recent measurements of absolute protein concentration to analyze the statistics of protein abundance across the human metabolic network. At a global level, we find that the enzymes in glycolysis comprise approximately half of the total amount of metabolic proteins and can constitute up to 10% of the entire proteome. We then use this analysis to investigate several outstanding problems in cancer metabolism, including the diversion of glycolytic flux for biosynthesis, the relative contribution of nitrogen assimilating pathways, and the origin of cellular redox potential. We find many consistencies with current models, identify several inconsistencies, and find generalities that extend beyond current understanding. Together our results demonstrate that a relatively simple analysis of the abundance of metabolic enzymes was able to reveal many insights into the organization of the human cancer cell metabolic network.