Modeling gene regulation from DNA sequence

Cells read the same genome in different ways, and much of that difference is written in DNA: promoters, transcription-factor binding sites, and arrangements of elements whose effect depends on context. Sequence-to-function models such as Enformer (Avsec et al., 2021) and Borzoi (Linder et al., 2025) predict regulatory activity directly from long stretches of sequence, and they predict it well.

Prediction is the easier half. I build these models to learn what they have learned: whether it transfers to genes, conditions, and species outside training, and whether it identifies the bases that actually control expression.

Questions I’m pursuing

Does the model generalize, or does it remember? A model can fit abundant measurements through correlations that fail on a new gene or a new condition. I evaluate on held-out chromosomes, genes, and conditions so that memorization does not pass for biology.

Where should prior knowledge come from? Functional labels are scarce and concentrated in a few organisms; genome sequence is abundant, and evolution has already tested which variation a genome tolerates. DNA language models can learn from that record (Karollus et al., 2024). Which genomes to learn from is an empirical question, not a matter of using as many as possible.

When is an explanation evidence? Attribution and in silico mutagenesis highlight bases, but a highlighted base is a hypothesis. I check explanations against known motifs, measured variant effects, and condition-specific responses before I trust them.

What I have built

Shorkie: evolutionary pretraining for yeast expression

Shorkie (Chao et al., 2026) began with a supervised yeast model that fit the genes it had seen but did not generalize to unseen genes as strongly as we wanted. Budding yeast has a 12 Mb genome, so we asked whether related fungi could supply what one genome could not. We pretrained the same masked-DNA model on four corpora of increasing breadth (one Saccharomyces cerevisiae genome, 80 strains, 165 Saccharomycetales genomes, and 1,341 fungal genomes), fine-tuned each identically, and compared them. The final model was fine-tuned on 5,215 RNA-seq, ChIP-exo, and histone tracks, including time courses after transcription-factor induction.

The 165-genome order transferred best; the 1,341-genome kingdom did not beat it. Against the same architecture trained from random initialization, pretraining raised mean gene-level correlation on held-out induction tracks from 0.74 to 0.88. Without motif labels, the model recovered Reb1, Rap1, Abf1, Cbf1, the PAC motif, and TATA boxes, and it ranked first among the compared models on an MPRA-validated set of regulatory variants. The published checkpoints can be explored on Shorkie Lab.

Observation and intervention

Sequence models learn from natural variation; perturbations ask what happens when a gene is changed on purpose. I contributed to a genome-wide Perturb-seq effort spanning roughly 100 million cells (You et al., 2026), the intervention side of the question Shorkie approaches from sequence.

Future directions

I want sequence models that earn trust the way experiments do, by predicting something before it is measured. That means evaluating natural variation, functional assays, and controlled perturbations in one frame and reading the disagreements closely. A predicted effect that a perturbation does not show, or a constrained element that no assay marks, points either to a flaw in the model or to an experiment worth running. Telling the two apart is the work.