Sasha Chernenkoff

PASTL: Fine-tuning genomic foundation models for protein abundance prediction University of Calgary MSc Thesis Project
Protein abundance is shaped by transcription, post-transcriptional control, translation efficiency and post-translational control. mRNA levels alone are a weak proxy for protein output, and protein labels are too sparse to train a sequence model from scratch. This project fine-tunes a genomic foundation model pretrained on transcriptional readouts to predict protein abundance directly from DNA sequence. It then uses in silico mutagenesis to study how retargeting the model changes the sequence features it responds to. The model's variant scores are also tested as a functional prior for proteome-wide association.

Code: github.com/sashachernenkoff/pastl

Diploid, patient-specific gene expression modelling University of Calgary Research Project
Most genomic sequence-to-function models operate on haploid reference sequences. This makes them poor predictors of real patient regulatory variation. They fail to capture cis-epistasis: the non-additive compounding effect of multiple regulatory variants interacting on the same haplotype. This is a limitation for models intended for clinical use. This project uses diploid, patient-specific genome sequences to predict personalized gene expression coverage. It keeps the trunk of a genomic foundation model frozen, trains LoRA adapters on its attention layers, and trains a new prediction head. The project investigates how much of the model's improvement on personalized data reflects genuine genotype-driven signal rather than simply learning a new data distribution.

HERL: VAE-based heritability estimation from genotype data University of Calgary Research Project
Standard heritability estimation methods, such as GCTA-GREML, build genetic relationship matrices (GRMs) as linear summaries of genotype data. These may not capture nonlinear structure such as dominance, epistasis or population structure. This project trains a categorical variational autoencoder (VAE) that compresses genome-wide genotypes into low-dimensional latent representations. The decoder models each genotype as one of three states, which lets it represent non-additive effects. GRMs built from the latent representations are used to estimate SNP heritability with GREML. They are compared with standard linear GRMs across WTCCC disease cohorts and GTEx cis-eQTL windows.

Code: github.com/sashachernenkoff/herl

MPS I & II PerkinElmer Genomics
At PerkinElmer Genomics, I developed semi-quantitative LC-MS/MS assays that detect analytes in dried blood spots. These analytes are present only in individuals with MPS I & II, and the assays run together as a panel. The markers are virtually undetectable in the normal population. This gives a larger minimum differential factor and higher confidence in results than traditional MPS tests. The tests also distinguish pseudodeficient alleles from pathogenic variants, which reduces the false positives common in MPS I testing. We then began investigating trends between variant type and marker concentration in MPS I. These trends could help interpret variants of unknown significance, which currently make clinical status difficult to predict.

Pompe CRIM-status determination PerkinElmer Genomics
At PerkinElmer Genomics, I developed a qualitative assay to detect the presence of endogenous acid α-glucosidase (GAA) within human peripheral blood mononuclear cells (PBMCs) isolated from whole blood using Western blotting. The assay allows for rapid CRIM-status determination in patients with Pompe disease in order to more quickly initiate immune tolerance induction therapy.

KDM6B 'bump-and-hole' research University of Pittsburgh Undergraduate Research
As an undergraduate at the University of Pittsburgh, I utilized the 'bump-and-hole' method to develop variants of KDM6B with 'hole-modified' active sites which were sensitive to selective inhibition by α-ketoglutarate engineered with complementary steric 'bumps' to help elucidate KDM6B's specific epigenetic function.