In this post:

TL;DR: a comprehensive benchmark for yeast-focussed sequence-to-expression models

Summary

Sequence-to-expression (S2E) models predict sequencing assays like RNA-seq that measure some part of the regulatory processing of DNA directly from a DNA sequence. We don’t have a standardised way to compare S2E-models for yeast which is the gap we fill with https://github.com/Tom-Ellis-Lab/yeast-seq2expression-benchmark.

In the first release, we benchmark two recent models, Shorkie and Yorzoi, across a set of 10 benchmark tasks measuring everything from single-nucleotide variants to bacterial chromosomes in yeast. One result from this benchmark is that models seem to capture promoter grammar in training conditions (growth on glucose) well but completely ignore the effect of inserting a fixed expression cassette at different locations of the genome or put differently assign too little importance to the flanking sequence context.

We hope you find this collection of tasks useful to build and benchmark your own models on and are very happy to help with any kind of questions or hands-on support with setting it up for you!

In the coming months, we will publish our main findings as a paper - stay tuned!

A benchmark to understand (yeast) S2E-models

Why this benchmark exists

The primary motivation for this benchmark is that yeast is an industrially relevant organism, for which sequence-to-function models offer a route to understanding how changes in DNA sequence alter transcription. It remains unclear, however, how well current models perform and how far they generalise across different aspects of this problem. In the medium term, better S2E models would accelerate DNA design by serving as in-silico oracles for ML-generated sequences.

Yeast is also substantially easier to engineer than mammalian cell lines. This tractability has produced resources with no mammalian counterpart, such as cells carrying exogenous chromosomes or scrambled genomes. Probing S2E models on tasks that are simply unavailable in mammalian systems may therefore reveal general strengths and weaknesses of this model class.

Our benchmark tests models across six diverse tasks on 10 datasets

What this benchmark measures

In the first release, we benchmark two recent models, Shorkie and Yorzoi, across a set of 10 benchmark tasks measuring a range of regulatory mechanisms.

Overview

Benchmark Description Task Primary metric
Caudal eQTL single-nucleotide cis-eQTLs vs matched controls binary classification AUROC / AUPRC (mean ± SEM over 4 negative sets)
Kita eQTL single-nucleotide cis-eQTLs (independent panel) binary classification AUROC / AUPRC
Rafi / deBoer MPRA (promoter) ~71k random 80 bp promoters in a dual-reporter plasmid GPRA regression (marginalized logSED) Pearson / Spearman
Shalem MPRA (terminator) designed 3′-end / terminator variants (cleavage & termination) regression (marginalized logSED) Pearson r (Spearman alongside)
Chen synonymous MPRA synonymous-codon effect on mRNA abundance regression, 3 libraries Pearson r + Spearman ρ on log2(R/D)
Wu RFP insertions fixed RFP cassette across ORF-deletion loci (position effect) regression Pearson r + Spearman ρ (+ tail AUROC)
Hong IGR insertions fixed reporter across intergenic loci (position effect) regression Spearman ρ on IntProp
Brooks SCRaMBLE SCRaMBLE rearrangement (altered neighbour / downstream context) coverage-track LFC direction balanced accuracy, then Spearman / Pearson
Cuperus 5′-UTR random 50 bp 5′-UTRs (Kozak, uORFs, structure) regression (HIS3 reporter) Spearman/Pearson + partial-corr over Kozak features
Meneu foreign DNA whole bacterial chromosomes integrated in yeast (far-OOD sequence) zero-shot coverage-track prediction per-window Pearson + JS divergence (+ fold-change error)

Contact

Feel free to reach out at [email protected]