Connie Huang

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2023-83

May 10, 2023

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-83.pdf

State-of-the-art deep learning models have been recently utilized to understand factors that affect gene expression. Current models are able to predict gene expression well across differ- ent genes in the reference genome directly from the DNA sequence alone, but their performance on individual level variation has not yet been studied deeply. To investigate model capabilities on individual variation, we evaluate four models - Enformer, Basenji2, Expecto, and Xpresso - on personal genome data and find limited performance when explaining variation in expression across individuals. Some genes have strong negative correlations between predicted and observed expression levels, suggesting that the models have identified causal regulatory variant(s) but incorrectly predicted their direction of effect. Our comparison of all four models reveals that the models often disagree with one another on the predicted direction of genetic effects on expression, and that models agree more often on the magnitude than on the direction of genetic effects.

Advisors: Nilah Ioannidis

---

BibTeX citation:

@mastersthesis{Huang:EECS-2023-83,
    Author= {Huang, Connie},
    Editor= {Ioannidis, Nilah},
    Title= {Performance and comparison of current sequence-based models on individual gene expression prediction},
    School= {EECS Department, University of California, Berkeley},
    Year= {2023},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-83.html},
    Number= {UCB/EECS-2023-83},
    Abstract= {State-of-the-art deep learning models have been recently utilized to understand factors that affect gene expression. Current models are able to predict gene expression well across differ- ent genes in the reference genome directly from the DNA sequence alone, but their performance on individual level variation has not yet been studied deeply. To investigate model capabilities on individual variation, we evaluate four models - Enformer, Basenji2, Expecto, and Xpresso - on personal genome data and find limited performance when explaining variation in expression across individuals. Some genes have strong negative correlations between predicted and observed expression levels, suggesting that the models have identified causal regulatory variant(s) but incorrectly predicted their direction of effect. Our comparison of all four models reveals that the models often disagree with one another on the predicted direction of genetic effects on expression, and that models agree more often on the magnitude than on the direction of genetic effects.},
}

EndNote citation:

%0 Thesis
%A Huang, Connie 
%E Ioannidis, Nilah 
%T Performance and comparison of current sequence-based models on individual gene expression prediction
%I EECS Department, University of California, Berkeley
%D 2023
%8 May 10
%@ UCB/EECS-2023-83
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-83.html
%F Huang:EECS-2023-83