Skip to content

Wen Group

Cell-Type-Resolved Chromatin Effects of AD/PD Variants: DNA Language Model vs.

Code Video Cerebral Valley

Core trick

Fine-tunes a DNA language model on accessibility labels and compares its variant-effect correlations with a CNN across four brain cell types.

Tags

science

Description

Cell-Type-Resolved Chromatin Effects of AD/PD Variants: DNA Language Model vs. ChromBPNet. We asked whether a general DNA language model can rival the specialized CNN ChromBPNet at predicting how non-coding variants affect chromatin accessibility. We fine-tuned our model, NucEL, on the same brain DNase-seq and GM12878 ATAC data and ran ChromBPNet's own endpoints. NucEL beat zero-shot ChromBPNet at accessibility classification in all four brain cell types and on GM12878 (0.967 vs 0.899), and independently reproduced ChromBPNet's variant-effect atlas (ρ=0.50; CR1 the shared top AD hit). A 1×→50× scaling curve shows the gap is data-limited, not architectural — suggesting general sequence models can match purpose-built regulatory CNNs given enough data.
FreakingPotato/NucEL_Chromatin_Effects★ 10 forksPythonMITCreated Jul 13, 2026 · last push Jul 13, 2026

README excerpt

# Cell-Type-Resolved Chromatin Effects of AD/PD Variants: DNA Language Model vs. ChromBPNet

**Claude Science Hackathon — Research Track**

Predicting how a non-coding variant reshapes chromatin accessibility in a cell type you care about is the home turf of **ChromBPNet**, a specialized convolutional network (the approach used by the Corces lab). This project treats it as a head-to-head contest: we take our own general-purpose DNA language model, **NucEL** (a ModernBERT architecture pretrained with masked language modeling on genomic sequence), and make it perform ChromBPNet's own cell-type-resolved endpoints on identical benchmarks.

> **Central finding.** Given enough labeled data, a general DNA language model **overtakes the specialized CNN on accessibility classification** and **independently reproduces its variant-effect atlas** (direction + top hits). The gap at low data is **data-limited, not architectural**.

All computation ran locally on RTX PRO 6000 (Blackwell) GPUs.

📄 **Full report:** [`report/report_en.html`](report/report_en.html) · [`report/report_en.pdf`](report/report_en.pdf)

---

## TL;DR scoreboard

![Scoreboard](figures/Figure1_scoreboard_LM_vs_ChromBPNet.png)

**Figure 1.** Head-to-head on ChromBPNet's own cell-type-resolved endpoints. The fine-tuned DNA language model NucEL beats zero-shot ChromBPNet at accessibility classification on GM12878 (0.967 vs 0.899) and on all four brain cell types (mean 0.884 vs 0.744), and reproduces ChromBPNet's variant-e

Team