Enter Note Done

Session


M. Stricker1, W-Y. Cheng2, S. Nahkuri3, P. Palamara1; 1Univ. of Oxford, Oxford, United Kingdom, 2Roche Pharma Res. & Early Dev. Informatics, Roche Innovation Ctr., New York, NY, 3Roche Pharma Res. & Early Dev. Informatics, Roche Innovation Ctr., Zurich, Switzerland

Epigenetics is known to play a key role in the regulation of adaptive and innate immune-system (AIIS) relevant genes that dictate the therapeutic course for many diseases such as cancer and COVID-19.
We built a model that leverages epigenetic data to perform supervised gene classification, with the goal of detecting AIIS genes. We leveraged known similarities between epigenetic states to encode high-dimensional epigenetic data (Epigenome Roadmap, Kundaje et al 2015; ChromHMM, Ernst & Kellis 2012) for different human tissues into images. We then trained a convolutional neural network from a handcrafted list of 477 known AIIS genes to perform image recognition, achieving a testing accuracy of 0.93 (SE 0.03). We used the trained model to scan the human genome for new AIIS regions and detected 1964 putative loci, of which 1129 do not have an associated Gene Ontology (GO) term. Some of these predicted novel AIIS regions harbour immune-cell specific alternative transcription variants, suggesting underlying biological mechanisms.
To evaluate the role played by these predicted AIIS regions in the genetic architecture of complex immune-related traits, we built a genome-wide annotation that reflects our model’s confidence of AIIS relevance at each locus. Regions with high predicted AIIS relevance (score > 0.5) were found to be enriched for GO terms related to immune-related function (p = 7e-21). We used linkage disequilibrium score regression (LDSC) coupled with genome-wide association summary statistics for 176 traits (average N = 262k) to test whether our AIIS annotation is predictive of regional heritability for these traits. We detected a significant heritability enrichment (LDSC |τ*| p < 0.05/176) for 20 out of 24 immune-related traits (including e.g. Crohn's disease). Analysis of additional 152 traits that are not directly immune-related revealed heritability enrichment for 5 traits, including respiratory and dermatological diseases for which immune function is likely to play a role.
To confirm the specificity of the detected enrichments for immune-related phenotypes, we performed a meta-analysis of 64 independent traits, including 4 autoimmune diseases. For all autoimmune diseases, the heritability enrichment remained strong (|τ*| > 0.5; p < 0.05/64) after conditioning on 97 other functional and evolutionary annotations. Immune-related traits and diseases were 3.45x (SE 0.09) more enriched for heritability (p = 9e-92) than non-immune system traits.
These results underscore the promise of leveraging machine learning algorithms and large epigenetic datasets to detect genomic regions implicated in specific classes of heritable traits and diseases.
Session Type
Poster Presentations
Track
Reviewers’ Choice
Topic
Statistical Genetics and Genetic Epidemiology