Identifying non-coding drivers of cancer

Division Of Basic Sciences - Nci

Investigators

Linked publications, trials & patents

Abstract

After compiling and annotating publicly available, single nucleotide variants in Lung cancer, we quantify the effect of non-coding mutations using CAPE, a machine learning approach that estimates the probability of a mutation deactivating transcription factor (TF) binding. When we compare enhancer mutations with coding region mutations, we do not see significant differences between the scores. Further, when we examine if there is a correlation between high magnitude CAPE scores and changes in gene expression between normal and cancer tissue contexts, we again do not see any significant correlation. The results suggest that lung enhancers do not exhibit significantly higher deleterious effects when we use CAPE as a method for quantifying mutational consequences. We are considering alternative approaches and additional datasets to test our hypotheses. Related to the above effort, in parallel, we have developed a sequence-based deep-learning model to quantify enhancer activity of a DNA sequence, which enables us to assess the effect of single nucleotide mutations. Using this tool, we have now shown emergence of novel enhancers in the human prefrontal cortex by single nucleotide mutations in the human lineage. The manuscript reporting this is now under review. Going forward we will use this approach to assess cancer non-coding mutations. With regards to our sub-aim of prioritizing TF binding, We have compiled a set of 10 TF-tissue pairs and are in the process of developing a model for scoring using deep learning based on sequence features. Previously, we tried this with a decision tree model trained using motifs as features, and found weak correlation with different tests of functionality (e.g. conservation, overlap with an enhancer, found in replicate samples). For example, for HNF4A in liver tissue and GATA2 in K562 cells, we found significantly higher evolutionary conservation in the highly scored peaks vs. the low scoring peaks. We think that utilizing a deep learning model will achieve more success in distinguishing functional peaks.