- Summary
- A refreshingly honest ML-for-genomics paper: strong held-out benchmarking, an explicit failure analysis, and full artifact release. The parameter efficiency is a real contribution.
- Strengths
- Clear hypothesis, pre-registered analysis, honest reporting of limitations, and openly available data and code. The figures are legible and the statistics report effect sizes.
- Weaknesses
- Would like a calibration plot and a note on training-set leakage across chromosomes. Minor.
- Detailed comments
- Methods: specify the randomization seed and hardware. Stats: add a sensitivity analysis for the excluded outliers. Figures: Fig. 3 needs error bars and an n per cell. Lit: engage the 2024–2025 replications. None of this is fatal; most is a revision away.
RSP-2026-0002 · Computer ScienceAcceptedv1 · round 1
A Transformer Baseline for Variant Effect Prediction in Non-Coding DNA
Rita Ora-Nguyen, Dr. N. Simone · submitted Jun 19, 2026
- genomics
- transformers
- variant effect
- machine learning
85% · Accepted
Open peer-reviewed research. Publication here is a review record, not an endorsement of clinical use — this material is not medical advice.
Abstract
Non-coding variants drive much of the heritability of common disease, yet their functional effects remain hard to predict. We present a compact transformer trained only on public epigenomic tracks that matches larger models on held-out variant-effect benchmarks while using an order of magnitude fewer parameters. We release weights, training code, and an honest error analysis, including the cell types where the model fails.
Manuscript — v1
Viewer not loading? Open the PDF directly
The Review Record
2 reviews on record── Round 1 ──
- Summary
- Good engineering and good science hygiene. I want the chromosome-split leakage question settled explicitly before I'm fully comfortable.
- Strengths
- Clear hypothesis, pre-registered analysis, honest reporting of limitations, and openly available data and code. The figures are legible and the statistics report effect sizes.
- Weaknesses
- Chromosome holdout must be stated precisely; otherwise benchmark numbers risk optimism.
- Detailed comments
- Methods: specify the randomization seed and hardware. Stats: add a sensitivity analysis for the excluded outliers. Figures: Fig. 3 needs error bars and an n per cell. Lit: engage the 2024–2025 replications. None of this is fatal; most is a revision away.
Both reviewers are positive; the one substantive concern (chromosome-split leakage) was clarified satisfactorily in the rebuttal and the camera-ready adds a calibration plot. Accepted. Given the translational potential, consider The Amp.