Knowledge graph-based recommendation framework identifies drivers of resistance in EGFR mutant non-small cell lung cancer

Abstract

Resistance to EGFR inhibitors (EGFRi) presents a major obstacle in treating non-small cell lung cancer (NSCLC). One of the most exciting new ways to find potential resistance markers involves running functional genetic screens, such as CRISPR, followed by manual triage of significantly enriched genes. This triage process to identify ‘high value’ hits resulting from the CRISPR screen involves manual curation that requires specialized knowledge and can take even experts several months to comprehensively complete. To find key drivers of resistance faster we build a recommendation system on top of a heterogeneous biomedical knowledge graph integrating pre-clinical, clinical, and literature evidence. The recommender system ranks genes based on trade-offs between diverse types of evidence linking them to potential mechanisms of EGFRi resistance. This unbiased approach identifies 57 resistance markers from >3,000 genes, reducing hit identification time from months to minutes. In addition to reproducing known resistance markers, our method identifies previously unexplored resistance mechanisms that we prospectively validate.

EMBL’s European Bioinformatics Institute (EMBL-EBI) in 2022

“…In May 2022, the resources were made available in the cloud via Google BigQuery and AWS Open Data. This integration and accessibility enables the data to be used in AI applications, such as machine learning to identify novel target-disease associations ( 50 ), building knowledge graphs for different biological insights ( 51–54 ) and for benchmarking new computational methods for drug target prioritisation ( 55 , 56 ).…”

Building a knowledge graph to enable precision medicine

“…Nevertheless, many barriers to linking and efficiently exploiting medical information across healthcare organizations and biological scales slow down the research and development of individualized care 2 . While many have acknowledged the difficulties in linking existing biomedical knowledge to patient-level health records [2][3][4][5] , few realize that biomedical knowledge is itself fragmented. Biomedical knowledge about complex diseases comes from different organizational scales, including genomics, transcriptomics, proteomics, molecular functions, intra-and inter-cellular communications, phenotypes, therapeutics, and environmental exposures.…”

“…A resource that comprehensively describes the relationships of diseases to biomedical entities would enable the large-scale, data-driven study of human disease. Understanding the connections between diseases, drugs, phenotypes, and other entities could open the doors for many types of research to leverage recent computational advances, including but not limited to the study of disease phenotyping [6][7][8] , disease etiology 9 , disease similarity 10 , disease diagnosis [11][12][13] , disease treatments 14 , drug-disease relationships [15][16][17] , mechanisms of drug action 18 and resistance 3 , drug repurposing [19][20][21] , drug discovery 22,23 , adverse drug events 24,25 , combination drug therapies 26 , and so forth. Many researchers have developed knowledge graphs for individual diseases that have helped advance computational precision medicine within their respective disease area [27][28][29][30][31][32][33][34][35][36][37][38][39][40][41][42] .…”

TigerLily: Finding drug interactions in silico with the Graph

“…Publicly available graphs [7] are differentiated by each other based on the types of nodes and edges present in the graph, various node type hierarchies [7], the size of the biological graph [9] and the use case that was driving the creation of the graph [11]. These heterogeneous biological graphs can be used by drug discovery researchers in the pharmaceutical industry [7] as an input for off-and on target drug repurposing [10], gene target identification [12], and compound interaction prediction [33] systems.…”

“…We split the drug pair features and the labels with the train_test_split function into training and testing parts (line 5). We create a LightGBMClassifier instance, learn the model from the training set drug pairs, score on the test set, compute the AUROC score and print the score by taking the first few digits (lines [7][8][9][10][11][12][13][14][15][16][17]. This snippet demonstrated that TigerLily interfaces with existing machine learning libraries smoothly.…”