Shorkie

Predicting dynamic expression patterns in budding yeast with a fungal DNA language model.

Shorkie is a sequence-to-expression model for Saccharomyces cerevisiae: a masked DNA language model pretrained on hundreds of fungal genomes, then fine-tuned on 5,215 epigenomic and transcriptomic tracks — including a large set of transcriptional-regulator induction RNA-seq experiments generated for this study — to predict RNA-seq coverage and variant effects.

License Preprint

What you can do with Shorkie

🧬 Predict RNA-seq coverage

Run the 8-fold ensemble over any 16 kb window and get predicted coverage across all 5,215 tracks.

Using Shorkie
🔬 Score variant effects

Compute logSED for a SNP — the metric used in the paper’s eQTL and MPRA benchmarks — from the command line or a notebook.

Using Shorkie
🧠 Use the DNA language model

Masked-token prediction and self-attention embeddings from the fungal LM that Shorkie is built on.

Using Shorkie_LM
🎛️ Fine-tune on your own tracks

Transfer the LM trunk to your own RNA-seq / ChIP data, including a small demo you can actually run.

Fine-tuning on your own RNA-seq
📊 See the analysis

A figure-by-figure tour of the paper, each linked to an executed notebook that regenerates it and re-checks the published numbers.

Analysis gallery
💾 Get the data

Every dataset curated for this study — models, four pretraining corpora, reference genome, and the eQTL/MPRA benchmarks — with sizes and commands.

Datasets

Quick start

git clone --recurse-submodules https://github.com/calico/shorkie-paper.git
cd shorkie-paper
conda env create -f environment.yml && conda activate yeast_ml
pip install -e external/baskerville-yeast -e external/westminster -e .
cp config/paths.example.yaml config/paths.yaml

bash data/download.sh --minimal                     # 8 Shorkie folds (~0.46 GB)
bash data/download.sh --genome -u <your-gcp-project> # R64 FASTA + GTF

python minimal_example/run_shorkie_variant.py \
    --model_dir ./my_shorkie

See Installation for details and Quick start for a walkthrough of that command’s output.

The three released models

Model

What it is

Use it for

Shorkie_LM

Masked DNA language model, pretrained on 165 Saccharomycetales genomes

Masked-base prediction, sequence embeddings, and as the starting point for fine-tuning

Shorkie

8-fold supervised ensemble, fine-tuned from Shorkie_LM on 5,215 tracks

Coverage prediction and variant-effect scoring — this is the main model

Shorkie_Random_Init

Same architecture and data, trained from scratch (no LM pretraining, lr 5e-4)

The ablation that isolates what LM pretraining contributes

All three are public and catalogued with checksums — see Released models.

Getting started

Using Shorkie

Background

Reference