CASTLE: A database of synthetic lethal sets predicted from genome-scale metabolic networks

SummaryGenome-scale metabolic networks have been reconstructed for hundreds of organisms over the last two decades, with wide-ranging applications, including the identification of drug targets. Constraint-based approaches such as flux balance analysis have been effectively used to predict single and combinatorial drug targets in a variety of metabolic networks. We have previously developed Fast-SL, an efficient algorithm to rapidly enumerate all possible synthetic lethals from metabolic networks. Here, we introduce CASTLE, an online standalone database, which contains synthetic lethals predicted from the metabolic networks of over 130 organisms. These targets include single, double or triple lethal set of genes and reactions, and have been predicted using the Fast-SL algorithm. The workflow used for building CASTLE can be easily applied to other pathogenic models and used to identify novel therapeutic targets.AvailabilityCASTLE is available at https://ramanlab.github.io/CASTLE/[email protected]

Application of artificial intelligence to predict a new class of novel synthetic lethal targets.

2598 Background: Synthetic lethal targets are proteins that are contextually vulnerable. Inhibitors of PARP1, for example, selectively produce a lethal phenotype in the context of cancer cells which have lost BRCA1 or BRCA2 function. As a high mutation rate is a hallmark of many cancers, targeting synthetic lethal interactions to selectively inhibit cancer cells with altered genetic backgrounds may increase the specificity and efficacy of therapeutics. Recently, clinical trials have targeted synthetic lethal pairs such as EGFR and BRAF, TP53 and BCL2, and PTEN and CHD1. Previous attempts to identify synthetic lethal targets have relied on empirical results from published studies of biological pathways perturbed in cancer cells. Developing strategies to rapidly identify synthetic lethals by combining multiple experimental and computational approaches would result in a new class of potential cancer drug targets beyond the existing efforts that rely on single experimental or computational methods alone. Methods: Here we present Expansive AI, an artificial intelligence augmented knowledge network that enables rapid hypothesis generation for accelerated discovery research. Using a purpose-built, hypergraph database of massive, integrated genomic and biomedical data, we can query all synthetic lethals and their component genes, as well as a wealth of data related to these genes. The database of biological data includes 11,000+ cancer genomes from TCGA, prior knowledge resources such as gene ontology and pathway resources, and experimental data including chemical and protein interaction and patent data. The hypergraph’s architecture allows for linking and nesting data, enabling efficient extraction of biologically-relevant features. Results: Using these features, a neural network classified 540 new candidate pairs that have previously not been reported. The candidate pairs were filtered to include only known oncogenes and least-studied genes. This produced a list of gene pairs which may represent the most novel class of synthetic lethal target candidates identified to date. Conclusions: We highlight the results of this AI-based approach and discuss validation efforts of the predicted interactions in specific cancer contexts.