Borzoi

Summary¶

Borzoi predicts cell-type- and tissue-specific RNA-seq coverage from 524 kb of DNA sequence at 32 bp resolution, then derives variant effect scores for transcription, splicing and polyadenylation from statistics computed on that one predicted coverage track.[1] This is a deliberately different design philosophy from models that predict each regulatory layer with a dedicated head: Borzoi's premise is that RNA-seq coverage already integrates every post-transcriptional layer, so a model that reproduces the coverage shape has implicitly learned all of them. It is the direct predecessor and principal comparator for AlphaGenome.

Architecture and training¶

Input: 524 kb DNA. Output: coverage in 32 bp bins.[1]
Backbone: the Enformer core — a convolution-and-subsampling tower followed by self-attention blocks with relative positional encodings operating at 128 bp — extended with a U-Net decoder that upsamples back to 32 bp in two 2× steps, each combining upsampled attention outputs with the matching feature map from the convolution tower via residual convolution.[1]
Training data: uniformly processed ENCODE RNA-seq (866 human, 279 mouse datasets) plus two to three replicates per GTEx tissue from recount3, paired with the Enformer training set (CAGE, DNase-seq, ATAC-seq, ChIP-seq).[1]
Ensembling: four randomly initialised replicate models, averaged.[1]

Ablations showed that including DNase-seq and ATAC-seq alongside RNA-seq, and including mouse data, both substantially improved eQTL concordance and enhancer–gene linking — i.e. the auxiliary chromatin data is not decoration, it teaches the model where the regulatory elements are.[1]

Track prediction performance¶

Metric	Value
Bin-level Pearson R, human RNA-seq (single model / ensemble)	0.74 / 0.75
Gene-level Pearson R (exon-aggregated, log2)	0.87
Gene-level R after quantile normalisation and mean-subtraction (cross-track specificity)	0.58
Bin-level R within transcript structure, top-20%-variance genes	0.88

Against Enformer on overlapping datasets, accuracies are broadly similar with two exceptions: Borzoi is lower on DNase and higher on CAGE.[1]

For tissue-resolved isoform structure across five GTEx tissues (blood, liver, brain, muscle, oesophagus):[1]

tissue-specific gene expression fold change: Spearman R 0.52–0.75;
alternative TSS usage ratios: Spearman R 0.85 against measurements, but only 0.29–0.50 for tissue-specific fold changes;
alternative polyadenylation distal/proximal ratios: Spearman R 0.81 against GTEx, 0.23–0.41 for tissue-specific fold changes.

Variant effect prediction¶

Expression (eQTL). Classifying fine-mapped GTEx eQTLs from TSS-distance-matched negatives, the Borzoi ensemble with an L2 coverage statistic reached mean AUROC 0.794 versus 0.747 for Enformer with its original sum aggregation; a single Borzoi model still reached 0.788, and Borzoi with the sum statistic 0.772.[1] For effect-size magnitude, mean Spearman R was 0.334 versus Enformer's 0.227.[1]

Limitation the authors state plainly: at prioritising the true eGene among the genes surrounding an eQTL, Borzoi performed "at best, marginally better than a TSS distance baseline".[1] Ranking variants and identifying their target gene are different problems, and this model class solves the first much better than the second.

Enhancer–gene linking. Borzoi scores elements up to 262 kb from a gene — roughly twice Enformer's reach — and uses exon rather than TSS annotations, which is more robust to alternative isoforms. It beat Enformer and a distance-to-TSS baseline on average precision and AUROC at all distances, on both the CRISPR datasets of Fulco/Gschwind-type screens and Gasperini et al. (2019).[1]

Polyadenylation (paQTL). On 1,058 fine-mapped 3′ QTLs from the eQTL Catalogue against expression-matched, PAS-distance-matched negatives, AUPRC rose monotonically with proximity to the polyadenylation site (0.64–0.74), consistently beating APARENT2.[1] For intronic polyadenylation QTLs specifically — where the PAS competes with the flanking splice junctions — average AUPRC was 0.725, evidence that the model has learned the competition between two distinct regulatory processes.[1]

Splicing (sQTL). On 4,105 fine-mapped sQTLs, Pangolin held a slight overall advantage (ΔAUPRC = 0.01 for SNPs within 10 kb of a splice site), because most distant variants are de novo splice-gain mutations that a dedicated splice-site model classifies easily from the local effect, and Borzoi's splice-gain predictions are less well calibrated. Borzoi was better close to the junction (ΔAUPRC = 0.02 within 200 bp). Critically, the average rank of both models beat either alone (ΔAUPRC > 0.02).[1]

Negative selection. Distinguishing gnomAD singletons from common variants within ENCODE cCREs, Borzoi and CADD were equal (mean AUROC 0.55 each) and their combination was best (0.57) — sequence-model and conservation signal are not redundant.[1]

Recommended variant scores¶

The authors give explicit guidance, worth following because the scores are not interchangeable:[1]

Question	Statistic
Does the variant change abundance of gene X?	exon-aggregated coverage log fold change for X
Does it change polyadenylation?	maximum difference in coverage log ratio between any 3′ cleavage site
Does it change splicing?	maximum normalised difference in any coverage bin within the gene body
Target gene unknown	gene-agnostic total L2 norm of differential coverage

Attribution method also matters: input gradients and standard ISM give good attributions at splice junctions and enhancer–promoter regions, but window-shuffled ISM is more reliable in 3′ UTRs because of buffering effects.[1]

Interpretability¶

Applying TF-MoDISco to tissue-specific exon-coverage gradients recovers the expected master regulators de novo — SPI1/B and IRF4/8 for blood, HNF4A/G and HNF1A for liver, SOX9 and REST for brain, MYOD1 and MEF2D for muscle — and the aggregated per-TF saliency differences between tissue pairs correlate with measured TF expression fold changes (Spearman R up to 0.77 for blood versus muscle).[1] The same procedure applied to polyadenylation and splice-centric coverage ratios recovers CFIm/CPSF/CstF and known splice-regulatory motifs respectively.[1]

Limitations¶

Stated by the authors:[1]

Tissue-specific splicing is not captured; the model "rather tended to predict the average RNA-seq shape". This is the single most important caveat for splicing work.
No mRNA half-life determinants were recoverable from 3′ UTR attributions.
Sequencing bias contaminates splice inference: 3′-end read pile-up and GC bias produced false positives when classifying alternatively used splice sites from predicted coverage.
eGene prioritisation barely beats a distance baseline (above).
32 bp bins blur base-resolution features; the authors name megabase receptive fields and finer resolution as the next goal — which is what AlphaGenome subsequently delivered.

Citations¶

[1] Linder, J., Srivastava, D., Yuan, H., Agarwal, V. & Kelley, D. R. (2025). Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation. Nature Genetics 57:949–961. Supports: architecture, training data and ensembling; all track-prediction and QTL benchmark figures; the eGene-prioritisation and tissue-specific-splicing limitations; recommended variant scores; attribution findings; the gnomAD/CADD comparison. Location: Results ("RNA-seq model design"; "Borzoi accurately predicts RNA-seq and other assays"; "Inference of tissue-specific expression and isoform usage"; "Improved context use for gene expression prediction"; "Borzoi prioritizes genetic variants that influence expression"; "Functional polyadenylation variant interpretation"; "Functional splicing variant interpretation"; "Intronic polyadenylation variant interpretation"); Discussion; Figs. 1–7. Source paper: s41588-024-02053-6.pdf