the problem: protein abundance is shaped by a complex interplay of
transcription, post-transcriptional control, and translation efficiency - factors
that are difficult to learn from scratch with limited labeled data. mRNA levels
alone are a weak proxy for protein output.
what i built: TL-Prot fine-tunes Enformer - a state-of-the-art
pretrained genomic foundation model trained on vast regulatory genomics data - as
a backbone for protein abundance prediction. rather than training from scratch,
it leverages Enformer's learned representations of DNA regulatory context to
bootstrap the prediction task.
what i found: [quantitative results in progress - will update]
why it matters: a central open question in computational biology
is whether pretrained genomic FMs can generalize beyond their original training
tasks. TL-Prot is a concrete test of that - using a sequence-to-function model
trained on chromatin and expression signals to predict a related but distinct
molecular phenotype.
DiploZoi
the problem: most genomic models operate on haploid reference
sequences, ignoring the personal genome variation that drives differences in gene
regulation between individuals. for these models to be clinically useful, they
need to handle diploid, patient-specific inputs.
what i built: DiploZoi takes diploid personal genome sequences
as input and predicts patient-specific gene expression patterns. it's built on
a frozen Borzoi trunk - preserving the pretrained FM representations intact -
with a trainable prediction head adapted for the personalized task.
what i found: a per-bin cross-patient Pearson correlation showed
DiploZoi substantially improves per-patient coverage shape accuracy over base
Borzoi (r ≈ 0.86 vs r ≈ 0.63). however, neither model
showed strong genotype-driven personalization at the per-bin level - suggesting
the gain is mostly tissue-level calibration rather than individual variant effects.
why it matters: fine-tuning on a small patient cohort doesn't
necessarily add personalization signal - it may just recalibrate scale. DiploZoi
gives a controlled test of this: the pretrained representations dominate, and
fine-tuning adds calibration rather than genotype sensitivity. that's a useful
finding for how the field should think about deploying these models clinically.
MPS I & II
at PerkinElmer Genomics, i developed semi-quantitative assays to detect the
presence of analytes (found to only be present in individuals with MPS I & II)
from dried blood spots using LC-MS/MS, which can be run simultaneously as a panel.
these markers are virtually undetectable within the normal population, yielding a
larger minimum differential factor and higher confidence in results compared to
traditional tests for MPS. in addition, the tests also allowed for differentiation
of pseudodeficient allele variants from pathogenic variants, reducing the number
of false positives commonly seen in MPS I testing.
we then began to investigate trends between variant types and MPS marker
concentrations, specifically for MPS I, as we believe it could prove clinically
useful in understanding variants of unknown significance, which currently yield
ambiguity in predicting clinical status.
pompe CRIM-status determination
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.
kdm6b 'bump-and-hole' 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.
