Kaitlyn Gayvert

A Computational Approach for Identifying Synergistic Drug Combinations

A promising alternative to address the problem of acquired drug resistance is to rely on combination therapies. Identification of the right combinations is often accomplished through trial and error, a labor and resource intensive process whose scale quickly escalates as more drugs can be combined. To address this problem, we present a broad computational approach for predicting synergistic combinations using easily obtainable single drug efficacy, no detailed mechanistic understanding of drug function, and limited drug combination testing. When applied to mutant BRAF melanoma, we found that our approach exhibited significant predictive power. Additionally, we validated previously untested synergy predictions involving anticancer molecules. As additional large combinatorial screens become available, this methodology could prove to be impactful for identification of drug synergy in context of other types of cancers.

A Computational Drug Repositioning Approach for Targeting Oncogenic Transcription Factors

Mutations in transcription factor (TF) genes are frequently observed in tumors, often leading to aberrant transcriptional activity. Unfortunately, TFs are often considered undruggable due to the absence of targetable enzymatic activity. To address this problem, we developed CRAFTT, a computational drug-repositioning approach for targeting TF activity. CRAFTT combines ChIP-seq with drug-induced expression profiling to identify small molecules that can specifically perturb TF activity. Application to ENCODE ChIP-seq datasets revealed known drug-TF interactions, and a global drug-protein network analysis supported these predictions. Application of CRAFTT to ERG, a pro-invasive, frequently overexpressed oncogenic TF, predicted that dexamethasone would inhibit ERG activity. Dexamethasone significantly decreased cell invasion and migration in an ERG-dependent manner. Furthermore, analysis of electronic medical record data indicates a protective role for dexamethasone against prostate cancer. Altogether, our method provides a broadly applicable strategy for identifying drugs that specifically modulate TF activity.

A New Big-Data Paradigm For Target Identification And Drug Discovery

Drug target identification is one of the most important aspects of pre-clinical development yet it is also among the most complex, labor-intensive, and costly. This represents a major issue, as lack of proper target identification can be detrimental in determining the clinical application of a bioactive small molecule. To improve target identification, we developed BANDIT, a novel paradigm that integrates multiple data types within a Bayesian machine-learning framework to predict the targets and mechanisms for small molecules with unprecedented accuracy and versatility. Using only public data BANDIT achieved an accuracy of approximately 90% over 2000 different small molecules – substantially better than any other published target identification platform. We applied BANDIT to a library of small molecules with no known targets and generated ∼4,000 novel molecule-target predictions. From this set we identified and experimentally validated a set of novel microtubule inhibitors, including three with activity on cancer cells resistant to clinically used anti-microtubule therapies. We next applied BANDIT to ONC201 – an active anti- cancer small molecule in clinical development – whose target has remained elusive since its discovery in 2009. BANDIT identified dopamine receptor 2 as the unexpected target of ONC201, a prediction that we experimentally validated. Not only does this open the door for clinical trials focused on target-based selection of patient populations, but it also represents a novel way to target GPCRs in cancer. Additionally, BANDIT identified previously undocumented connections between approved drugs with disparate indications, shedding light onto previously unexplained clinical observations and suggesting new uses of marketed drugs. Overall, BANDIT represents an efficient and highly accurate platform that can be used as a resource to accelerate drug discovery and direct the clinical application of small molecule therapeutics with improved precision.

Abstract 362: Computational drug repositioning identifies dexamethasone as potential ERG inhibitor

Proceedings: AACR Annual Meeting 2014; April 5-9, 2014; San Diego, CA Mutations in genes coding for transcription factors (TFs) are frequently observed in tumors, many of which lead to aberrant transcriptional activity. Unfortunately, transcription factors are often considered “undruggable” due to the absence of targetable enzymatic activity and the large surface contacting DNA. To address the transcription factor “druggability” problem, we developed a computational repositioning approach to identify small molecules that can perturb the activity of transcription factors. Our approach involves identifying drugs that mobilize many of the target genes of a transcription factor. This approach uses Gene Set Enrichment Analysis to integrate genomewide binding data (ChIP-seq) with drug perturbation differential gene expression profiles. When applied to ENCODE ChIP-seq and the Connectivity Map expression profiles, our approach predicted 38,000 disruptive drug-TF interactions (FWER<0.1). These predictions included the known inhibitory effect of a BRD4 inhibitor (JQ1, FWER=0.000) on MYC and of HSP90 inhibitors (e.g.17-AAG, FWER=0.031) on HSF1. We used an integrated biological network with 22k genes and 7k drugs to identify predicted disruptive drug-TF interactions. Based on this approach network path length of predicted drug-TF was significantly shorter than non-predictions (p=2.2e-16). Many predicted drug-TF interactions involved only one protein intermediary between the drug and the TF, indicating that our predictions are not random and suggesting that many drugs might disrupt TFs by targeting their regulatory or interacting co-factors. We then decided to apply our approach to identifying candidate molecules that can inhibit ERG, a pro-invasive, oncogenic TF over-expressed in as many as 50% of prostate cancer patients. Using ERG ChIP-seq peaks in prostate cells, dexamethasone (FWER=0.086) was predicted to inhibit ERG transcriptional activity. Using cell invasion and migration assays, we found that dexamethasone significantly decreased cell invasion and migration in DU145 prostate cancer cells over-expressing ERG, but not in isogenic control cells. Dexamethasone also abrogated expression of PLAU, a known ERG target, and substantially decreased binding of ERG at the PLAU promoter. Analysis of ERG ChIP-seq peaks revealed a highly enriched AP-1 DNA motif and preferential mobilization by dexamethasone of genes near peaks containing the AP-1 motif (p=0.06). ChIP-seq experiments showed that dexamethasone reduced AP-1 binding at ERG-JUN promoter sites. These results suggest that dexamethasone inhibits ERG by disruption of AP1, a key co-factor. Altogether, this method provides a novel, broadly applicable strategy to computationally identify drugs that indirectly target transcription factors. This may be of further interest for other factors with oncogenic activity, such as FOXA1 or for reactivating deactivated tumor suppressive transcription factors such as p53. Citation Format: Kaitlyn Gayvert, Cynthia Cheung, David Rickman, Olivier Elemento. Computational drug repositioning identifies dexamethasone as potential ERG inhibitor. [abstract]. In: Proceedings of the 105th Annual Meeting of the American Association for Cancer Research; 2014 Apr 5-9; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2014;74(19 Suppl):Abstract nr 362. doi:10.1158/1538-7445.AM2014-362

A Machine Learning Approach Predicts Tissue-Specific Drug Adverse Events

One of the main causes for failure in the drug development pipeline or withdrawal post approval is the unexpected occurrence of severe drug adverse events. Even though such events should be detected by in vitro, in vivo, and human trials, they continue to unexpectedly arise at different stages of drug development causing costly clinical trial failures and market withdrawal. Inspired by the “moneyball” approach used in baseball to integrate diverse features to predict player success, we hypothesized that a similar approach could leverage existing adverse event and tissue-specific toxicity data to learn how to predict adverse events. We introduce MAESTER, a data-driven machine learning approach that integrates information on a compound’s structure, targets, and phenotypic effects with tissue-wide genomic profiling and our toxic target database to predict the probability of a compound presenting with different types of tissue-specific adverse events. When tested on 6 different types of adverse events MAESTER maintains a high accuracy, sensitivity, and specificity across both the training data and new test sets. Additionally, MAESTER scores could flag a number of drugs that were approved, but later withdrawn due to unknown adverse events – highlighting its potential to identify events missed by traditional methods. MAESTER can also be used to identify toxic targets for each tissue type. Overall MAESTER provides a broadly applicable framework to identify toxic targets and predict specific adverse events and can accelerate the drug development pipeline and drive the design of new safer compounds.

Abstract 1563: A machine learning approach to predict platform specific gene essentiality in cancer

Loss-of-function (LOF) screenings across a set of diverse cancer cell lines has the potential to reveal novel synthetic lethal interactions, cancer-specific vulnerabilities, and guide treatment options. These were traditionally done using shRNAs, but with the recent emergence of CRISPR technology there has been a shift in methodology. The Achilles project is to date the largest cancer LOF screening effort undertaken, however we found a large amount of inconsistency between their shRNA and CRISPR-Cas9 essentiality results for the same set of cell lines. Here we characterize the differences between genes found to be essential in either CRISPR or shRNA screens. We found that certain features such as gene expression, network connectivity and conservation could accurately separate out essential genes that were found exclusively in either one of these screens. This information could be tremendously useful in understanding the differences in the CRISPR and shRNA screening results. Furthermore, one limitation with Project Achilles was that they conducted shRNA screens on 216 cell lines, but only 33 cell lines in CRISPR. Therefore we developed a model that integrates these genetic, network, and population features to predict CRISPR results from shRNA screenings, and found that our model can accurately identify CRISPR essential genes better than approaches just based on the shRNA results (p-value Citation Format: Coryandar M. Gilvary, Neel S. Madhukar, Kaitlyn M. Gayvert, David S. Rickman, Olivier Elemento. A machine learning approach to predict platform specific gene essentiality in cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 1563. doi:10.1158/1538-7445.AM2017-1563

Abstract 5039: A data driven approach to predicting tissue-specific adverse events

Adverse events are currently one of the main causes of failure in drug development and withdrawal after approval. As a result, predicting drug side effects is an incredibly important part of drug discovery and development. With the emergence of precision medicine there has been a surge in interest on creating drugs for specific protein targets, however we lack accurate ways to connect drug targets and mechanisms to specific side effects. Here we take a target-centric approach to in-silico drug side effect prediction. We have mined drug side effect databases and grouped sets of side effects to the originating human tissue. For each of 30+ tissues, we defined a set of “toxic targets”- proteins that are only targeted by drugs with toxicity in that tissue - and “safe targets” - proteins only targeted by drugs with no related tissue toxicities. We found that toxic targets are consistently more highly expressed than safe targets, indicating that their mechanisms may be more crucial in their respective tissue. Furthermore we found that toxic targets have higher network connectivity. Using published gene knockdown screens, we also found that toxic targets for each tissue are significantly more likely to be essential than safe targets and are more likely to be enriched for GO terms related to cell death. These pieces of information all reinforce the proposed relationship between the identified toxic targets and drug induced tissue toxicities. We next leveraged this information to draw insights into unexpected drug toxicity events. We applied the BANDIT drug target prediction tool to drugs misclassified by the PrOCTOR toxicity prediction method and drugs with a specific type of tissue toxicity that were not known to hit any of our identified toxic targets. We found that new drug-target predictions explained a large number of these toxicities, correctly classifying approximately five times as many side effects as would have been expected by random chance. These results all supported our target-centric hypothesis of drug side effect prediction. Therefore we built a set of machine-learning models that would integrate drug targets with tissue-wide expression patterns and gene-specific features to predict specific side effects for a given drug. We found that these methods could significantly outperform other prediction techniques and random chance. For instance, our method for predicting drug induced liver injury (DILI) had ~70% accuracy at pinpointing specific drugs known to cause DILI and its likelihood score correlated with the FDA’s reported DILI severity score. Overall these findings show how a target-centric approach to drug development can not only help us understand the relation between targets and specific phenotypic effects, but can help drug developers predict side-effects before costly and time-consuming clinical studies. Our hope is that adoption of these methods will lead to overall increase in drug development efficiency and bring safer drugs to the market quicker. Note: This abstract was not presented at the meeting. Citation Format: Kaitlyn M. Gayvert, Neel Madhukar, Coryandar Gilvary, Olivier Elemento. A data driven approach to predicting tissue-specific adverse events [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2017; 2017 Apr 1-5; Washington, DC. Philadelphia (PA): AACR; Cancer Res 2017;77(13 Suppl):Abstract nr 5039. doi:10.1158/1538-7445.AM2017-5039

Abstract 3916: A “moneyball” approach to predicting clinical trial toxicity events

Over the past decade, significant strides have been made in the management and treatment of various diseases. Despite this progress, clinical attrition rates have continued to substantially rise. Clinical trials can fail for a variety of reasons, ranging from design issues to drug efficacy and safety problems. Drug-likeness approaches, as first proposed by Lipinski almost two decades ago, have become a key tool for the pre-selection of compounds that are likely to have manageable toxicity in clinical studies. However all these methods consider molecular properties of the drug itself alone. In general, these approaches struggle to simultaneously well-characterize the properties of both FDA approved drugs (which we term the sensitivity) and drugs that fail clinical trials (specificity). We introduce an approach that integrates chemical properties of a compound, along with that of its targets, to provide a new quantitative measure that helps predict whether drugs in clinical trials will fail for toxicity reasons. When trained on failed clinical trials and FDA approved drugs, this method performs at a high accuracy, specificity and sensitivity (∼0.75), as well as high area under the ROC curve (>0.80). In comparison, none of the drug-likeness approaches were able to successfully maintain both high sensitivity and specificity. A feature analysis of the model indicates that it is critical to consider both structural properties and properties of the drug target, with the target9s network connectivity and liver toxicity as two important features. The approach was further evaluated by testing the predictions of the trained model on an established independent dataset. We found that our method was able to significantly distinguish a representative set of bioavailable drugs from a representative set of toxic drugs (D = 0.2133, p Citation Format: Kaitlyn Gayvert, Neel Madukhar, Olivier Elemento. A “moneyball” approach to predicting clinical trial toxicity events. [abstract]. In: Proceedings of the 107th Annual Meeting of the American Association for Cancer Research; 2016 Apr 16-20; New Orleans, LA. Philadelphia (PA): AACR; Cancer Res 2016;76(14 Suppl):Abstract nr 3916.

Abstract LB-106: Using a data-driven Bayesian approach to predict the targets of orphan small molecules and ways to overcome drug resistance

For natural products and phenotypic screen derived small molecules, one of the greatest bottlenecks is target identification. Since most experimental approaches are failure-prone and require a large investment of time and resources, computational methods have the potential to substantially improve target identification efforts. Recently there has been an explosion of genomic, chemical, clinical, and pharmacological datasets that characterize a small molecule9s mechanism of action, but no method has been invented to integrate the multiple, independent pieces of evidence provided by each data type into a cohesive prediction framework To address this problem we developed BANDIT: a Bayesian ANalysis to determine Drug Interaction Targets. BANDIT integrates over 20,000,000 pieces of unique information from seven distinct data types to predict drug targets. By integrating these data types within a Bayesian network we outperformed other target prediction methods and achieved a target prediction accuracy of over 90%. Furthermore, we observed that BANDIT was able to predict the precise targeting mechanism as well as the protein target. We further validated BANDIT9s accuracy by using it to reproduce the results of a previously published experimental kinase inhibitor screen without any additional data. We observed that our target predictions for the tested small molecules were significantly more likely to be the targets with the highest levels of inhibition in the experimental screen (P = 3e-06). We next used BANDIT to predict targets and mechanisms for over 50,000 small molecules with no known targets. We identified 24 molecules with varying structures and efficacies that we predicted to inhibit microtubule polymerization. Using immunofluorescence and tubulin binding assays we were experimentally validated microtubules as targets for 18 of these compounds - a success rate much higher than expected by random chance. Moreover, one of the greatest challenges with current microtubule chemotherapy is the development of drug resistance, which is lethal for the patients. Remarkably, we identified a subset of our compounds to be active against tumor cells resistant to Eribulin, a microtubule depolymerizing agent FDA approved for breast cancer, and cross-resistant to Vincristine and Colchicine, known depolymerizing drugs. This demonstrated that BANDIT could not only rapidly identify drug mechanisms and targets, but also determine novel molecules with the potential to act on refractory tumors. Finally we used BANDIT to observe how different drug types could interact with one another. We observed many unexpected and exciting results. For instance we observed a tight clustering of the seemingly unrelated classes of beta-blockers with anti-Parkinson9s medications. This interaction potentially indicates a shared protein target interaction and provides support to studies proposing the clinical use of beta-blockers in the treatment of tremors in Parkinson9s patients. Examples such as this indicate the validity of BANDIT9s clustering approach and reveal how previously unknown shared target interactions could cause the unexplained phenotypic effects. This approach could further be expanded to determine undiscovered relationships between classes of drugs, discover which drugs could be repositioned for new uses, and predict clinical drug synergies. Citation Format: Neel S. Madhukar, Prashant Khade, Linda Huang, Kaitlyn Gayvert, Giuseppe Galletti, Paraskevi Giannakakou, Olivier Elemento. Using a data-driven Bayesian approach to predict the targets of orphan small molecules and ways to overcome drug resistance. [abstract]. In: Proceedings of the 107th Annual Meeting of the American Association for Cancer Research; 2016 Apr 16-20; New Orleans, LA. Philadelphia (PA): AACR; Cancer Res 2016;76(14 Suppl):Abstract nr LB-106.

Abstract 887: N-Myc drives neuroendocrine prostate cancer

Emerging observations from clinical trials suggest that a subset of castration resistant prostate adenocarcinomas (CRPC) eventually evolve or progress to a predominantly neuroendocrine phenotype (NEPC). This transition is emerging as an important mechanism of treatment resistance. This cell plasticity is characterized by loss of androgen receptor (AR) and prostate specific antigen (PSA), and significant over-expression and gene amplification of MYCN (encoding N-Myc). While N-Myc is a bona fide driver oncogene in several rare tumor types, the molecular mechanisms that underlie N-Myc driven NEPC have yet to be characterized. Integrating a novel genetically engineered mouse (GEM) model of prostate specific N-Myc overexpression, human prostate cancer cell line modeling, and human prostate cancer transcriptome data, we found that N-Myc over-expression leads to the development of poorly differentiated, invasive prostate cancer that is molecularly similar to human NEPC tumors. To determine if N-Myc plays a causal role in driving the NEPC phenotype, we generated GEM lines that carry a CAG-driven lox-stop-lox human MYCN gene integrated into the ROSA26 (LSL-MYCN) locus and either a Tmprss2 driven tamoxifen-activated Cre recombinase (T2-Cre) or probasin (Pb)-Cre. Since PTEN deletion is a frequent alteration in CRPC and PI3K/AKT signaling can enhance N-Myc protein stability we also engineered the mice with a floxed Pten locus. N-Myc over-expression in the context of Ptenf/+ at 3 months post-induction leads to focal mouse high-grade prostatic intraepithelial neoplasia (mHGPIN). T2-Cre; Ptenf/f; LSL-MYCN+/+ mice develop highly proliferative, diffuse mHGPIN which consists of proliferations of cells with nuclear atypia that expand the glands, imparting irregular borders and inducing a mild stromal response, mitotic figures, and incipient necrosis. RNAseq data from N-Myc these mHGPIN lesions show they are molecularly similar to NEPC based on RNAseq data from 203 human CRPC and NEPC samples. At 6 months, Pb-Cre; Ptenf/f; LSL-MYCN+/+ mice develop poorly differentiated, highly proliferative, invasive prostate cancer. Based on the RNAseq data from the N-Myc GEM line, GEM-derived mouse prostate cancer organoid cultures and isogenic cell lines, we found that N-Myc regulates a specific NEPC-associated molecular program that includes a repression of AR signaling, enhanced AKT signaling and repression of Polycomb Repressive Complex 2 target genes. We further showed that N-Myc interacts with AR and this interaction depends on Enhancer of Zeste Homolog 2 (EZH2). Finally, N-Myc expressing cell lines and organoids displayed an enhanced sensitivity to inhibitors targeting the AKT pathway, EZH2 and Aurora-A. Altogether, our data shows that N-Myc drives the neuroendocrine phenotype in prostate cancer and provides rationale for the development of new therapeutic strategies for treating this aggressive subtype of prostate cancer. Citation Format: Etienne Dardenne, Himisha Beltran, Kaitlyn Gayvert, Matteo Benelli, Adeline Berger, Loredana Puca, Joanna Cyrta, Andrea Sboner, Zohal Noorzad, Theresa MacDonald, Cynthia Cheung, Dong Gao, Yu Chen, Martin Eilers, Juan Miguel Mosquera, Brian D. Robinson, Mark A. Rubin, Olivier Elemento, Francesca Demichelis, David S. Rickman. N-Myc drives neuroendocrine prostate cancer. [abstract]. In: Proceedings of the 107th Annual Meeting of the American Association for Cancer Research; 2016 Apr 16-20; New Orleans, LA. Philadelphia (PA): AACR; Cancer Res 2016;76(14 Suppl):Abstract nr 887.