Ekin Deniz Aksu
@ekindea
Scientist and medical doctor. Biology AI/ML methods, gene regulation, DNA sequence models, single cells. Doing a PhD in computational biology at @molgen.mpg.de.
Corgi can also predict eQTL variant effects in held-out tissue types. 8/9
Corgi+ can identify important cell type-specific regulators zero shot, using attribution/contribution scores of regulator expression. 7/n
We also describe Corgi+, an improved version specifically designed for the RNA-to-epigenome data imputation task. We find that it has state-of-the-art performance in this benchmark. 5/n
Corgi reaches top accuracy in joint cross-cell type and cross-sequence functional genomics predictions (see below and Figure 2&3). Also, it does not underperform relative to Borzoi when tested cross-sequence. 4/n
We aimed to design a neural network architecture that imitates cellular gene regulation. i) input gene expression vector is limited to trans-regulators (TFs, cofactors, histone modulators, RBPs) ii) trans-features modulate cis-features with affine transformations (FiLM) 3/n
I'm very happy to say that my first first-author paper has been published in Bioinf. Advances! 1/2 academic.oup.com/bioinformati...
Today I was happy to present Corgi at the @broadinstitute.org Broad Institute at the ML in Drug Discovery Symposium. If you want to use it for predicting genomic tracks, Corgi is now published on GitHub: github.com/ekinda/corgi
It's conference week in Cambridge. Yesterday I presented Corgi in the MIT/MGB AI Cures conference, and today I'm at the Broad Institute for the Variant to Function conference.
Looking forward to spending the next couple of months at MIT, collaborating with @bergerlab.bsky.social on deep learning methods in computational biology!
I will present my recent work Corgi, a context-aware sequence-to-activity model, at the IMSB/ECCB 2025 next week at the RegSys track. Send me a message if you want to meet! #ISMBECCB2025
New preprint alert! Corgi imitates cellular gene regulation and integrates DNA sequence and trans-regulator information. This allows Corgi to make accurate predictions in unseen cell types. Also, it can simulate trans-regulator perturbations in silico.