Yorzoi

Yorzoi is a sequence-to-expression model for Saccharomyces cerevisiae that predicts strand-specific RNA-seq coverage directly from DNA sequence. It was developed at Imperial College London with the University of British Columbia and posted as a preprint in September 2025. The motivating problem is practical: budding yeast is the standard chassis for testing heterologous DNA, but there is no cheap way to know whether a multi-gene construct will be transcribed as designed before it is built.

Existing yeast regulatory models mostly operate on a single gene and return one scalar per sequence, which cannot capture the failure modes that matter in real constructs — readthrough past a weak terminator occluding a downstream promoter, converging promoters colliding, or a bidirectional promoter driving antisense transcription. These reshape the coverage profile without necessarily changing a per-sequence readout such as fluorescence or TPM. Yorzoi instead predicts a positionally and strand-resolved profile over a 4,992 bp window spanning one to three genes, at 10 bp resolution.

The model adapts the architecture of Borzoi, the human and mouse RNA-seq coverage model, to a genome roughly 500 times smaller, shrinking the input window from Borzoi's 524 kb so that tiling the yeast genome still yields enough training samples. It arrives alongside Shorkie, a contemporaneous yeast sequence-to-expression model built on cross-species self-supervised pretraining; Yorzoi's distinguishing move is to broaden its training data beyond native sequence space rather than beyond a single species.

#Key Features

• Multi-gene, base-resolved output: Predicts 162 tracks — forward and reverse strand for each of 81 RNA-seq experiments — over the central 3 kb of the input window, so transcriptional interference between neighboring genes is visible in the prediction.
• Trained on non-native sequence: The corpus includes 200 kb–2 Mb human DNA inserts carried on yeast artificial chromosomes and SCRaMbLE-rearranged synthetic chromosome arms, extending the model past the native yeast genome that generative design tools routinely leave behind.
• Zero-shot transfer to MPRAs: Without any fine-tuning on assay data, summed coding-region coverage correlates with measured fluorescence at Pearson R = 0.52 on a random core-promoter library and R = 0.65 on a 13,000-member terminator library.
• Homology-aware data splits: Roughly 128 million pairwise BLAT scores build a homology graph used to select per-chromosome hold-out windows that minimize similarity to the training set, guarding against leakage-inflated performance.
• Open and immediately usable: MIT-licensed code, Apache-2.0 weights on HuggingFace, a PyPI package, a hosted HTTP inference API, and a no-code web application at yorzoi.eu.

#Technical Details

Yorzoi's architecture derives from Flashzoi, a PyTorch reimplementation of Borzoi using FlashAttention-2 and rotary positional embeddings. Convolutional layers downsample the one-hot sequence while widening channels, eight transformer blocks attend over the resulting 156-position representation at 32 bp resolution, upsampling layers with skip connections restore resolution, and a head emits per-track coverage. Removing downsampling layers relative to Borzoi yields a compact 19.3M-parameter model. Training used a shape-and-magnitude decomposed objective — multinomial negative log-likelihood plus 0.1× Poisson negative log-likelihood — over 250 epochs at batch size 30 on a single NVIDIA RTX A6000, taking about 1.5 hours.

Training data comprises about 21 Mb of sequence paired with coverage: 12 Mb from the native genome and its SCRaMbLE rearrangements, profiled by 71 direct Nanopore RNA-seq runs, and 9 Mb of human DNA in YACs, profiled by 10 new stranded Illumina experiments, tiled into roughly 16,000 samples. On held-out windows Yorzoi reaches a median per-sequence Pearson R of 0.69, with Nanopore-derived tracks predicted better (mean R = 0.73) than Illumina tracks (0.61). Magnitude error, as log2 fold change of summed counts, has σ = 1 for native yeast sequence and σ = 0.8 for human sequence.

#Applications

The model targets design-stage screening in yeast synthetic biology. The authors demonstrate three uses: comparing gene orders and orientations in a five-gene violacein pathway, where placing vioD divergently from vioA depresses its predicted coverage; benchmarking four codon optimizers, which substitute up to 60% of codons synonymously and are predicted to swing transcript levels roughly 15-fold; and driving the Ledidi editing framework to propose non-coding edits that double predicted GFP transcript levels. Inference runs through a web app and an HTTP API, so wet-lab groups without GPUs or ML expertise can prioritize constructs before committing to synthesis.

#Impact

Yorzoi is the first model to predict RNA-seq coverage profiles from sequence for S. cerevisiae at near-base-pair resolution across multiple genes, and its corpus of human DNA expressed in yeast is a novel source of non-native regulatory signal. The authors are candid about where it falls short: predicting how SCRaMbLE rearrangements change expression reaches only Pearson R = 0.33 and 0.62 balanced accuracy on the direction of change, and an ablation found that adding human DNA improved coverage prediction on human inserts without consistently improving downstream MPRA or novel-junction performance. The work remains a preprint awaiting peer review, and the training data covers a single growth condition, so conditional or inducible promoter behavior is outside its current scope.