Precision oncology matches tumors to targeted therapies based on the presence of actionable molecular alterations. However, most tumors lack actionable alterations, restricting treatment options to cytotoxic chemotherapies for which few data-driven prioritization strategies currently exist. Here, we report an integrated computational/experimental treatment selection approach applicable for both chemotherapies and targeted agents irrespective of actionable alterations. We generated functional drug response data on a large collection of patient-derived tumor models and used it to train ScreenDL, a novel deep learning-based cancer drug response prediction model. ScreenDL leverages the combination of tumor omic and functional drug screening data to predict the most efficacious treatments. We show that ScreenDL accurately predicts response to drugs with diverse mechanisms, outperforming existing methods and approved biomarkers. In our preclinical study, this approach achieved superior clinical benefit and objective response rates in breast cancer patient-derived xenografts, suggesting that testing ScreenDL in clinical trials may be warranted.

Selection of the treatments most likely to combat a patient’s tumor is a central aim of precision oncology. We are currently developing a functional precision oncology program at the University of Utah and Huntsman Cancer Institute that combines the multi-omic characterization of a patient's tumor with the functional screening of relevant drug candidates in patient-derived organoid tumor models to identify which drugs are uniquely capable of combating the patient’s cancer. Here we present our model-based approach for utilizing the multi-omic tumor data to computationally predict the patient’s response to each drug. In this approach, we first identify the genomic and transcriptomic vulnerabilities of the tumor as genes harboring somatic DNA mutations or copy number changes (from paired tumor/normal WGS/WES DNA sequencing data), as well as genes whose expression levels have been significantly altered in the tumor (based on bulk or single-cell RNA sequencing data). Second, using targeting information from drug-gene interaction databases (DGIdb) we compile a list of genes targeted by each drug relevant in the patient’s treatment. Third, we utilize gene interaction database knowledge (KEGG) to construct a gene interaction graph to link each drug’s gene targets with all of the tumor’s vulnerability genes. We then identify all gene interaction paths connecting specific drug target genes with specific vulnerability genes; and score each path and combine all path-specific scores for the target-vulnerability gene pair. Subsequently, we further combine all pairwise scores across all drug target genes and all tumor vulnerability genes and determine the statistical significance of the resulting score by sampling the background distribution of such scores across random drugs, target genes, and vulnerability genes. We have applied our algorithm to predict drug response in advanced breast cancer patients as well as using publicly available tumor cell line data (GDSC2). We have found a high level of concordance between our computational prediction and organoid/cell line screening results, clearly separating sensitive and non-sensitive models. Using Bayesian probability, we compare the drug-specific score distributions of sensitive and non-sensitive models and are able to identify sensitive cell lines/patients with high accuracy. By encapsulating available information on direct gene-gene interactions, the drug’s direct gene targets and the collected omic vulnerabilities of the tumor, our model is not only capable of predicting sensitivity to both targeted and chemotherapy agents, but can also provide a mechanistic understanding for the targeting of the tumor. Our approach will be validated and fine-tuned on a large cohort of breast and brain cancer patients as well as patient PDX models, interrogating gene-target interactions to identify novel relevant target genes/pathways. Citation Format: Szabolcs Tarapcsak, Yi Qiao, Xiaomeng Huang, Tony Di Sera, Matthew H. Bailey, Bryan E. Welm, Alana L. Welm, Gabor T. Marth. Model-based cancer therapy selection by linking tumor vulnerabilities to drug mechanism [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 2723.

Diagnosis of rare, or complex Mendelian conditions is a difficult and time-consuming process that typically requires integrating knowledge and expertize from a team of experts. For example, treating physicians have a comprehensive knowledge of a patient's family history, and have performed a differential diagnosis based on the clinical presentation. Medical geneticists have a deep understanding of disease etiology as well as genetic factors that contribute to different diseases. It is often the case, however, that time constraints and/or a lack of computational expertise, or familiarity with genomic data formats and bioinformatics tools, limits the ability of these critical experts to easily contribute to analyses.

OBJECTIVES/SPECIFIC AIMS: The objective of the study was 2-fold; to identify potentially deleterious alleles in a child with Treacher Collins syndrome, and; to demonstrate the value of the iobio analysis platform for intuitively and rapidly analyzing genomic data. METHODS/STUDY POPULATION: We used the iobio suite of web-based applications to analyze quality metrics for the sequencing data and called variants for the proband and his parents. We then visually interrogated variants in genes potentially associated with the syndrome in real-time, using the intuitive gene.iobio application. We sought high impact variants that demonstrated a predicted impact on the protein function, and were simultaneously at low allele frequency in the general human population. Variants were also compared against the ClinVar database of known mutations to identify variants that have already been associated with this, or related syndromes in the literature or clinical studies. Finally, the gene.iobio tool allows users to interrogate the primary sequencing data to ensure that no variants had been missed by the primary variant calling pipeline. This analysis pipeline was performed using intuitive web-based apps in real time, and consequently represents a system that is available to users that traditionally are excluded from these analyses. RESULTS/ANTICIPATED RESULTS: The iobio suite was used to rapidly assess data quality and interrogate genetic variants for a child with Treacher Collins syndrome. A compound heterozygote consisting of 2 missense alleles in the TCOF1 gene was identified as a compelling pathogenic allele, necessitating further functional investigation. The study helped validate the use of the intuitive iobio tools in such analyses, strengthening the case for greater involvement of medical professionals in data analysis. DISCUSSION/SIGNIFICANCE OF IMPACT: The performed analyses demonstrated that the whole genome sequencing data for the family being studied was of a very high quality, although 1 gene demonstrated a local region of almost zero coverage. This ensured that study conclusions can be presented with confidence. A variant associated with Treacher Collins syndrome 1 in ClinVar was uncovered in the TCOF1 gene, however, given it’s benign rating, this variant was not considered further. The most interesting candidate was a compound heterozygote, consisting of 2 missense mutations, also in the TCOF1 gene. These mutations occurred with allele frequencies of 22% and 8% in the general population, and additional molecular and functional studies are currently being pursued.