2021/22 Recipients

Nilah Ioannidis, UC Berkeley
Department of Electrical Engineering and Computer Sciences

Title: Computational methods for causal genetic variant discovery in polygenic disease, with application to kidney disease

Nilah IoannidisGenome-wide genetic variation affects our risks for complex polygenic diseases, but identifying the variants causal for disease risk and determining their mechanisms of action is an unsolved challenge in human genetics. Genome-wide association studies (GWAS) identify variants associated with disease, but most of these associations are non-causal. Since these associations predominantly fall in non-protein-coding regions of the genome, annotation of noncoding regulatory elements greatly improves fine-mapping of causal variants at associated loci. Deep learning methods trained on experimentally defined genome-wide epigenetic features can predict the effects of genetic variants on such regulatory elements, but optimizing these methods for use in GWAS fine-mapping and causal variant discovery remains a challenge.

In this project, the Ioannidis lab will collaborate with Profs. Jeremy Reiter and Gabriel Loeb at UCSF, combining computational and experimental expertise to address this challenge and advance methods for causal variant discovery, with application to kidney disease. We are using single-cell experimental measurements of chromatin accessibility, histone modifications, and gene expression in primary kidney cells to train deep learning models to predict the cell-type specific effects of genetic variants on these molecular phenotypes. We will advance the design and training of these models to improve their sensitivity to changes caused by single nucleotide variants, as well as their performance in disease-relevant regions of the genome containing kidney-specific regulatory elements. The methods developed here will not only advance our understanding of the genetic mechanisms underlying kidney disease, but will be widely applicable across many complex polygenic diseases.

 

Timothy McCalmont, UC San Francisco
Department of Pathology and Laboratory Medicine

Title: Quantifying Invasive Tumor Volume by Deep Learning to Improve Malignant Melanoma Prognosis

Timothy McCalmontMelanoma is the deadliest skin cancer with the fastest rising incidence of cancer in the United States. The most important predictor of melanoma patient survival is the volume of invasive tumor at the first biopsy. However, the current standard of care for outcome prediction is to manually measure a single-dimension tumor thickness, which acts as a proxy for volume. The predictive accuracy of this single measurement depends on a host of subjectively defined measurements which can lead to incorrect risk assessment and management planning. Recent advances in deep learning as applied to digital pathology will allow us to objectively assess the entirety of the invasive tumor and its interrelated variables. Here, we propose to use deep learning methods to automatically and accurately quantify the invasive tumor’s cross-sectional area, at the sensitivity of single tumor cells. We will calculate the relationship between the automatically calculated invasive melanoma areas to patient survival, and develop a survival prediction algorithm. We hypothesize that this efficient and reproducible method will better predict patient survival than current methods.