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

Shorkie is a semi-supervised sequence-to-expression model for yeast: a masked DNA language model pretrained on hundreds of closely related fungal genomes and fine-tuned on thousands of epigenomic and transcriptomic profiles—including a large set of transcriptional-regulator induction RNA-seq experiments generated for this study—to predict RNA-seq coverage and variant effects.

This repository lets you run Shorkie on your own yeast sequences and variants — and reproduce the analyses in the Shorkie paper. The model framework lives in the baskerville-yeast and westminster repositories (pinned as submodules under external/); this repo adds an installable helper package (src/shorkie), the released model/data catalogue, runnable examples, and the figure notebooks.

📖 Full documentation: khchao.com/shorkie — installation, the released models, usage guides for Shorkie and Shorkie_LM, a figure-by-figure analysis gallery, a dataset catalogue, and troubleshooting.

Please open a GitHub issue for bugs or questions. For other inquiries, contact drk (at) calicolabs.com, jlinder (at) calicolabs.com, or kuanhao.chao (at) gmail.com.

Requirements

OS: Linux. Python: 3.9. TensorFlow: ~2.15 (exact pins in environment.yml).
CPU is enough to run inference, score variants, and reproduce the figures from released data.
GPU (CUDA-capable) is needed only for training / fine-tuning and the GPU-marked figure panels — those also need tensorrt==8.6.1 and a CUDA-enabled TensorFlow build. containers/ ships a known-good Docker / Apptainer image for a scheduler-free run.

Quickstart

git clone --recurse-submodules https://github.com/calico/shorkie-paper.git   # (or git@github.com:calico/shorkie-paper.git with SSH access)
cd shorkie-paper
conda env create -f environment.yml && conda activate yeast_ml      # env name: yeast_ml
pip install -e external/baskerville-yeast -e external/westminster -e .   # model code + this package
cp config/paths.example.yaml config/paths.yaml                      # then edit `work_root`
bash data/download.sh --minimal                                     # 8 Shorkie folds for the example below

data/download.sh takes --minimal, --models [lm|finetuned|random_init|lm-variants|all] — all three model variants (Shorkie LM, the Shorkie 8-fold ensemble, and the Shorkie_Random_Init 8-fold ablation) are live on the public bucket gs://seqnn-share — plus --genome, --lm-corpus <tier>, --supervised, --eqtl, --mpra (all verified against data/manifest.json). Every filesystem path resolves through config/paths.yaml — there are no hardcoded machine paths.

To actually run the models you also need the R64 reference genome:

bash data/download.sh --genome -u <your-gcp-project>   # FASTA + GTF + .fai -> <dest>/genome/R64/

Use this copy, not a fresh Ensembl download. The chromosome naming is load-bearing and differs between the two files — the FASTA uses chrI…chrXVI, the GTF uses I…XVI. A genome pulled straight from Ensembl or SGD will not match, and every example will fail to find its sequence.

Approximate download sizes: --minimal (8 Shorkie folds) ≈ 0.46 GB; --models all (LM + both 8-fold ensembles) ≈ 0.97 GB; --genome ≈ 22 MB. The LM corpora (--lm-corpus) and --supervised bigwigs/TFRecords are large (the 165_Saccharomycetales tier is ~3.7 GB, the supervised bigwigs ~93 GB); see Data availability below, or the size_bytes / approx_size fields in data/manifest.json, for exact figures.

Then check your install:

pytest -q                       # release-integrity + smoke tests (fast, offline, no weights needed)

Using Shorkie on your own data

examples/ — step-by-step notebooks: load each model, run inference, score variant effects for Shorkie / Shorkie_LM, and fine-tune the LM on your own RNA-seq tracks. Start here.
examples/6_finetune_minidemo.sh — before committing GPU-days to fine-tuning, run this: the real --restore pipeline on a tiny slice of the released data, ~96 MB and about a minute on CPU. Its output is deliberately not a usable model — it proves the mechanics work.
minimal_example/ — a self-contained CLI that scores one SNP end-to-end (see Minimal Example below).
containers/ — Docker / Apptainer image for a scheduler-free run.
src/shorkie/ — the importable helper package: config (path resolution), models.ensemble (8-fold loader + logSED), helpers.yeast_helpers, viz.load_cov.

Stuck? The troubleshooting guide is keyed on the exact error messages people actually hit.

Model Availability

📖 Detailed documentation: khchao.com/shorkie → Released models — what each model is and when to use it, the on-disk layout, the shared unet_small_bert_drop architecture and its (16384, 170) input encoding, and guidance on choosing between the three.

The model weights are downloaded as .h5 files from the URLs below (or with data/download.sh --models all). Shorkie LM, Shorkie (8-fold), and Shorkie_Random_Init (8-fold) are all live on the public bucket gs://seqnn-share and catalogued (with md5s) in data/manifest.json.

(live) Shorkie LM
(live) Shorkie (gs://seqnn-share/shorkie_models/shorkie/)
- f0 | f1 | f2 | f3 | f4 | f5 | f6 | f7
(live) Shorkie_Random_Init (from-scratch ablation, lr 5e-4, 8-fold; gs://seqnn-share/shorkie_models/shorkie_random_init/)
- f0 | f1 | f2 | f3 | f4 | f5 | f6 | f7

See examples/ for runnable notebooks on each model — loading, inference, variant-effect prediction, and fine-tuning the LM on your own RNA-seq tracks.

LM checkpoints for the other corpus tiers

Catalogued as models.lm_variants (fetch with data/download.sh --models lm-variants): the four unet_small runs behind the Figure 1F/G corpus-scaling comparison, plus 1341_Fungus at unet_small_bert_drop.

Variant	Corpus	Architecture	`num_features`
`R64_yeast__unet_small`	R64 (1 genome)	`unet_small`	6
`80_strains__unet_small`	80_strains	`unet_small`	85
`165_Saccharomycetales__unet_small`	165_Saccharomycetales	`unet_small`	170
`1341_Fungus__unet_small`	1341_Fungus	`unet_small`	1366
`1341_Fungus__unet_small_bert_drop`	1341_Fungus	`unet_small_bert_drop`	1366

Not drop-in replacements for Shorkie LM. The released Shorkie LM is 165_Saccharomycetales + unet_small_bert_drop; the four unet_small runs are a different architecture, used for the corpus-scaling figure, and are meaningful compared against each other rather than against the released model. The one genuine alternative is 1341_Fungus__unet_small_bert_drop — same architecture, largest corpus.

num_features is not stored in params.json and must be set at load time; it is 4 (DNA) + num_species + 1. Passing the wrong value raises Error loading weights by name: axes don't match array.

Data availability

Everything we curated for this study, with how to get it. All of it is catalogued — with sizes, MD5s and destinations — in data/manifest.json, which is what data/download.sh reads. A browsable version is at khchao.com/shorkie → Datasets.

Two buckets: model weights are on the public gs://seqnn-share (plain HTTPS works, no account). Datasets are on gs://shorkie-paper, which is requester-pays — it needs gsutil and a billing-enabled GCP project via -u PROJECT; you pay egress only.

Dataset	Size	Get it with
Model weights (LM + Shorkie 8-fold + Random_Init)	0.97 GB	`data/download.sh --models all`
R64 reference genome (FASTA + GTF + `.fai`)	22 MB	`data/download.sh --genome -u PROJECT`
LM corpus — R64 (1 genome)	23 MB	`data/download.sh --lm-corpus R64 -u PROJECT`
LM corpus — 80_strains (80 genomes)	1.5 GB	`data/download.sh --lm-corpus 80_strains -u PROJECT`
LM corpus — 165_Saccharomycetales ⭐ (165 genomes)	3.7 GB	`data/download.sh --lm-corpus 165_Saccharomycetales -u PROJECT`
LM corpus — 1341_Fungus (1,361 genomes)	42.5 GB	`data/download.sh --lm-corpus 1341_Fungus -u PROJECT`
Supervised tracks — BigWigs	~93 GB	`data/download.sh --supervised bigwigs -u PROJECT`
Supervised tracks — 8-fold TFRecords	~10 GB	`data/download.sh --supervised tfrecords -u PROJECT`
cis-eQTL benchmark (scores + DREAM baselines)	~64 MB	`data/download.sh --eqtl -u PROJECT`
MPRA benchmark (ground truth + cached scores)	~1.6 GB	`data/download.sh --mpra all -u PROJECT`

⭐ = the corpus Shorkie_LM was actually pretrained on.

Cannot use Google Cloud? Requester-pays needs a billing-enabled GCP project, which is impractical in some regions. scripts/00_setup/zenodo_upload.py publishes the models and corpora to Zenodo — see data/README.md.

The pretraining corpora

Four tiers of increasing phylogenetic breadth, each with raw genomes and matched 16,384 bp ZLIB TFRecords. Shorkie LM was pretrained on 165_Saccharomycetales; the others back the ablations.

Tier	Genomes	Train seqs	`gs://shorkie-paper/data/unsupervised/…`
R64	1	1,201	`{genome,processed}/R64/`
80_strains	80	102,315	`{genome,processed}/80_strains/`
165_Saccharomycetales ⭐	165	385,551	`{genome,processed}/165_Saccharomycetales/`
1341_Fungus	1,361	625,355	`{genome,processed}/1341_Fungus/`

All four share one held-out split, drawn from S. cerevisiae R64 only and split by whole chromosome — valid: chrXI, chrXIII, chrXV; test: chrXII, chrXIV, chrXVI, with chrXI–XVI excluded from training everywhere. Keep that split if you want numbers comparable to ours. Per-tier species lists, with NCBI accessions, are committed in data/species_lists/.

1341_Fungus nests one level deeper. Its TFRecords live under an extra 1342_Fungus/ subdirectory (yes, a different number), unlike the other three tiers — so --lm-corpus 1341_Fungus yields …/processed/1341_Fungus/1342_Fungus/*.tfr and a fixed-depth glob will miss them. The public label is also historical: the cleaned corpus holds 1,361 assemblies.

Training script: scripts/02_train/shorkie_lm/.

The supervised tracks

5,215 tracks on S. cerevisiae R64 at 16 bp resolution, which Shorkie is fine-tuned on:

Induction Dynamics Gene Expression Atlas (IDEA) — RNA-seq induction time-course samples, generated for this study by Calico Life Sciences (related to IDEA 1.0; Hackett, S.R. et al., Mol Syst Biol, 2020)
Yeast strain RNA-seq across diverse S. cerevisiae isolates (Caudal, É. et al., Nat Genet, 2024)
ChIP-exo and ChIP-MNase (Rossi, M.J. et al., Nature, 2021)

gs://shorkie-paper/data/supervised/{bigwigs,processed}/. The targets sheet is also committed at minimal_example/sheet.txt so you can inspect track metadata without downloading anything. Training script: scripts/02_train/shorkie_finetuned/.

Benchmark data availability

External benchmark datasets used to evaluate Shorkie, with sources and primary references.

Released. The reproduction-critical subsets — the per-SNP eQTL score TSVs (Caudal/Kita/Renganaath, Shorkie / Shorkie_LM / Shorkie_Random_Init) and the MPRA ground-truth expression + cached Shorkie/DREAM scores — are catalogued in data/manifest.json and live on gs://shorkie-paper/{eqtl,mpra}/. Fetch them with data/download.sh --eqtl / --mpra (requester-pays; pass -u PROJECT) and Figures 6–7 reproduce on CPU without re-scoring. The large raw inputs (the 1011-genomes GVCF, the full DREAM Challenge sequences, the DREAM-RNN/PrixFixe weights) are third-party and obtained from their original sources below (not re-hosted).

MPRA (Random Promoter DREAM Challenge)

Dataset: Random Promoter DREAM Challenge MPRA (held-out set; 71,103 promoter sequences spanning eight categories: native promoters, random 80-bp oligos, high-expression, low-expression, “challenging” sequences, SNV perturbations, motif perturbations, and motif tiling).
Primary reference: Rafi, A. M. et al. “A community effort to optimize sequence-based deep learning models of gene regulation.” Nat Biotechnol (2024).
Notes: We evaluated Shorkie by replacing MPRA constructs into genomic context upstream of TSSs (details in the paper).

cis-eQTL benchmarks

We evaluate Shorkie and compare to DREAM models on three independent yeast cis-eQTL resources:

Caudal et al. pan-transcriptome
- Data portal: 1002 Yeast Genomes project
  - GWAS summary stats
  - gVCF (1011 isolates)
- Primary reference:
  - Caudal, É. et al. “Pan-transcriptome reveals a large accessory genome contribution to gene expression variation in yeast.” Nat Genet 56, 1278–1287 (2024).
  - Peter, J. et al. “Genome evolution across 1,011 Saccharomyces cerevisiae isolates.” Nature 556, 339–344 (2018).
- Notes: We benchmarked 1,901 local cis-eQTLs from ~1,000 isolates; negative controls were noncoding SNPs matched by allele, TSS distance, and MAF.
Kita et al. high-resolution eQTLs
- Supplementary table
- Primary reference: Kita, R. et al. “High-resolution mapping of cis-regulatory variation in budding yeast.” PNAS 114 (2017).
- Notes: We benchmarked 683 variants, stratified into Promoter, UTR5, UTR3, and ORF categories.
Renganaath et al. MPRA-validated cis-regulatory variants
- Article (eLife 2020)
- Primary reference: Renganaath, K., Chong, R., Day, L., Kosuri, S., Kruglyak, L. & Albert, F. W. “Systematic identification of cis-regulatory variants that cause gene expression differences in a yeast cross.” eLife 9, e62669 (2020).
- Notes: We benchmarked 142 core-promoter variants (Figure 7, panel G).

Minimal Example: Variant Effect Prediction with Shorkie

The minimal_example/ directory contains a self-contained script that demonstrates how to load Shorkie and compute a logSED (log₂ Sequence Effect Difference) score for a single SNP — no fine-tuning required.

Setup

Download model weights (8 folds). Easiest: bash data/download.sh --minimal (fetches into the my_shorkie/train/f{i}c0/train/model_best.h5 layout below and verifies MD5s against data/manifest.json). Or manually:

mkdir -p my_shorkie/train
for i in 0 1 2 3 4 5 6 7; do
  mkdir -p my_shorkie/train/f${i}c0/train
  wget -O my_shorkie/train/f${i}c0/train/model_best.h5 \
    https://storage.googleapis.com/seqnn-share/shorkie_models/shorkie/f${i}/model_best.h5
done

Fetch the R64 reference genome: bash data/download.sh --genome -u <your-gcp-project> (FASTA + GTF + .fai). See the naming caveat under Quickstart — use this copy rather than a fresh Ensembl download.

Run

python minimal_example/run_shorkie_variant.py \
  --model_dir  my_shorkie \
  --params_file  minimal_example/params.json \
  --targets_file minimal_example/sheet.txt \
  --fasta_file   /path/to/genome.fasta \
  --gtf_file     /path/to/genome.gtf \
  --chrom chrI --pos 124373 --ref T --alt C --gene YAL016C-B

Output

==================================================
  Variant  : chrI:124373 T>C
  Gene     : YAL016C-B
  logSED   : +0.0643
==================================================
  logSED > 0 → alt increases predicted expression
  logSED < 0 → alt decreases predicted expression

See minimal_example/README.md for full documentation.

Reproducing the paper figures

Each main-text figure has one notebook in notebooks/ (fig01–fig07). A notebook either runs end-to-end from released data (data/download.sh) or loads a gated intermediate produced by the cited scripts/ stage, then renders the panels by calling that figure's builders under reproduction/figure_NN/. See notebooks/README.md for the full figure → artifact → config-key index, and reproduction/ for the per-figure panel builders, published crops, and reproduced-vs-published checks (verify_figNN.csv).

The end-to-end pipelines live in scripts/, staged 00_setup → 01_data_build → 02_train → 03_eval → 04_analysis, for all three model variants:

Variant	Train	Analysis
Shorkie LM (masked DNA LM)	`02_train/shorkie_lm/`	`04_analysis/shorkie_lm/`
Shorkie (fine-tuned)	`02_train/shorkie_finetuned/`	`04_analysis/shorkie/`
Shorkie_Random_Init (random-init ablation, lr 5e-4, 8-fold)	`02_train/shorkie_scratch/`	`04_analysis/shorkie_scratch/`

The only difference between finetuned and random-init is the --restore flag + learning rate (see scripts/02_train/README.md).

Citation

If you use Shorkie, please cite our preprint:

Chao, K.-H., Magzoub, M. M., Stoops, E. H., Hackett, S. R., Linder, J., & Kelley, D. R. (2025). Predicting dynamic expression patterns in budding yeast with a fungal DNA language model. bioRxiv. https://doi.org/10.1101/2025.09.19.677475

@article{chao2025shorkie,
  title   = {Predicting dynamic expression patterns in budding yeast with a fungal DNA language model},
  author  = {Chao, Kuan-Hao and Magzoub, Majed M. and Stoops, Emily H. and Hackett, Sean R. and Linder, Johannes and Kelley, David R.},
  journal = {bioRxiv},
  year    = {2025},
  doi     = {10.1101/2025.09.19.677475},
  url     = {https://www.biorxiv.org/content/10.1101/2025.09.19.677475v1}
}

License

What	Terms
Code in this repository	Apache License 2.0 — see `LICENSE`
Model weights (`gs://seqnn-share/shorkie_models/`) and the data we derived (TFRecords, cached scores)	CC BY 4.0 — free to use and redistribute with attribution; please cite the preprint
Genome assemblies in the LM corpora	Third-party public data (Ensembl Fungi release 59 / NCBI GenBank), redistributed here for reproducibility, under their original providers' terms. Accessions are in `data/species_lists/`
Benchmark datasets (DREAM Challenge MPRA, the eQTL studies, 1002 Yeast Genomes)	Their own terms — see the original sources cited above

Please cite both the Shorkie preprint and the original data sources when you use the released data.

Navigation Menu

Repository navigation

README.md

README.md

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

Requirements

Quickstart

Using Shorkie on your own data

Model Availability

LM checkpoints for the other corpus tiers

Data availability

The pretraining corpora

The supervised tracks

Benchmark data availability

MPRA (Random Promoter DREAM Challenge)

cis-eQTL benchmarks

Minimal Example: Variant Effect Prediction with Shorkie

Setup

Run

Output

Reproducing the paper figures

Citation

License

FilesExpand file tree

README.md

Latest commit

History

README.md

File metadata and controls

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

Requirements

Quickstart

Using Shorkie on your own data

Model Availability

LM checkpoints for the other corpus tiers

Data availability

The pretraining corpora

The supervised tracks

Benchmark data availability

MPRA (Random Promoter DREAM Challenge)

cis-eQTL benchmarks

Minimal Example: Variant Effect Prediction with Shorkie

Setup

Run

Output

Reproducing the paper figures

Citation

License