Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation

Thus far, sequence-based models have focused on assays in which measured activity is proportional to local sequencing read counts.

“Thus far, such models have focused on assays in which measured activity is proportional to local sequencing read counts.”

• cites [4] Alipanahi, B., Delong, A., Weirauch, M. T. & Frey, B. J. Predicting the sequence specificities of DNA- and RNA-binding proteins by deep learning. Nat. Biotechnol. 33, 831–838 (2015). sourceNot checked — no full text of the source
• cites [5] Zhou, J. & Troyanskaya, O. G. Predicting effects of noncoding variants with deep learning-based sequence model. Nat. Methods 12, 931–934 (2015). sourcePartly supported

  The paper describes a sequence-based model trained on TF-binding, DNase I–sensitivity, and histone-mark profiles (Abstract) and uses sequencing reads to assess allelic imbalance (Online Methods, ‘Evaluation of the single-nucleotide sensitivity of chromatin feature prediction’). It does not state that measured activity in these assays is proportional to local read counts, or establish that sequence-based models generally had focused on such assays. Its own training labels are binary peak-overlap labels (Online Methods, ‘Data for training DeepSEA’). (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [6] Kelley, D. R., Snoek, J. & Rinn, J. L. Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks. Genome Res. 26, 990–999 (2016). sourcePartly supported

  The Results section describes Basset as a sequence-based model of accessibility using DNase-seq data. But Methods (“DNase I hypersensitivity data”; page not provided) says its training labels were binary indicators of significant peaks, not activity values proportional to local sequencing read counts. The paper therefore supports the assay focus but not the claimed read-count relationship. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [7] Avsec, Ž. et al. Base-resolution models of transcription-factor binding reveal soft motif syntax. Nat. Genet. 53, 354–366 (2021). sourcePartly supported

  The Introduction (page unavailable) describes existing transcription-factor binding models as predicting binary binding events or continuous binding signals, and the Results describe ChIP-nexus profiles in terms of local read counts. This supports the claim for the transcription-factor binding models discussed, but the paper does not establish that sequence-based models generally had focused on read-count-based assays. (LLM, conf=0.78)

  Experimental — accuracy not yet measured
• cites [8] Kelley, D. R. et al. Sequential regulatory activity prediction across chromosomes with convolutional neural networks. Genome Res. 28, 739–750 (2018). sourceSupported
• cites [9] Kelley, D. R. Cross-species regulatory sequence activity prediction. PLoS Comput. Biol. 16, e1008050 (2020). sourcePartly supported

  The Materials and methods section (Functional genomics data; page not specified) says the models focused on sequencing assays and describes deriving local signal from aligned reads. But it also says the models used log-fold-change signal tracks for many datasets; those targets are not simply proportional to local read counts. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [10] Chen, K. M., Wong, A. K., Troyanskaya, O. G. & Zhou, J. A sequence-based global map of regulatory activity for deciphering human genetics. Nat. Genet. 54, 940–949 (2022). sourcePartly supported

  The Main section says earlier sequence models focused on predicting transcription-factor binding and histone modifications. Results describes Sei’s chromatin-profile targets, but Methods (Deep learning sequence model training; page unavailable) says its training labels are binary peak calls. The paper does not establish that measured activity in these assays is proportional to local sequencing read counts. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Metrics that summarize post-transcriptional regulation inevitably struggle to describe complex splicing outcomes, unannotated de novo events or the intricate and sometimes competitive relationship between transcription, splicing and intronic polyadenylation.

“However, such metrics inevitably struggle to describe complex splicing outcomes, unannotated de novo events or the intricate and sometimes competitive relationship between transcription, splicing and (intronic) polyadenylation24–26.”

• cites [24] Hallier, M., Tavitian, A. & Moreau-Gachelin, F. The transcription factor Spi-1/PU.1 binds RNA and interferes with the RNA-binding protein p54nrb. J. Biol. Chem. 271, 11177–11181 (1996). sourcePartly supported

  The Results and Discussion report that the transcription factor Spi-1/PU.1 interfered with splicing in an in vitro β-globin assay, offering a limited example of a transcription-associated protein affecting splicing. The paper does not address whether post-transcriptional metrics struggle to describe splicing outcomes, unannotated de novo events, or the relationship with intronic polyadenylation; searches for those terms found no supporting passage. The indexed passages provide no page number for the Results and Discussion. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [25] Oksuz, O. et al. Transcription factors interact with RNA to regulate genes. Mol. Cell 83, 2449–2463.e13 (2023). sourceNot checked — no full text of the source
• cites [26] Kwon, B. et al. Enhancers regulate 3′ end processing activity to control expression of alternative 3′UTR isoforms. Nat. Commun. 13, 2709 (2022). sourcePartly supported

  The Introduction (page unavailable) says it is difficult to disentangle transcript production from processing when assessing expression of single-UTR genes. That supports a narrower version of the claimed complexity, but the paper does not establish that summary metrics inevitably fail to describe complex splicing, unannotated de novo events, or competition with intronic polyadenylation. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Predicted tissue-pooled distal-to-proximal polyadenylation coverage ratios of held-out genes were highly correlated with measurements from GTEx (Spearman's R = 0.81) and PolyADB v.3.

“Predicted tissue-pooled distal-to-proximal polyadenylation coverage ratios of held-out genes were highly correlated with measurements from GTEx (Spearman’s R = 0.81; Supplementary Fig. 3f) and PolyADB v.3 (refs. 43,44) (Supplementary Fig. 3g).”

• cites [43] Zhang, H., Hu, J., Recce, M. & Tian, B. PolyA_DB: a database for mammalian mRNA polyadenylation. Nucleic Acids Res. 33, D116–D120 (2005). sourcePartly supported

  The Results (Data access and visualization; page unavailable) identify polyA_DB as a database of polyadenylation sites, supporting the citation’s identification of the resource. Our copy does not identify the database as PolyADB v.3 or establish that this version supplied the cited measurements. The correlation is presented as a result of the citing study, not a finding attributed to this source. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [44] Wang, R., Nambiar, R., Zheng, D. & Tian, B. PolyA_DB 3 catalogs cleavage and polyadenylation sites identified by deep sequencing in multiple genomes. Nucleic Acids Res. 46, D315–D319 (2018). sourceSupported

Recent work on the Enformer model using self-attention has demonstrated a path toward modeling long-range regulatory information.

“Recent work on the Enformer model using self-attention has demonstrated a path toward achieving this goal13.”

• cites [13] Avsec, Ž. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nat. Methods 18, 1196–1203 (2021). sourceSupported

Previous models have only attempted to work with RNA-seq after summarizing gene expression in a single statistic.

“Previous models have only attempted to work with RNA-seq after summarizing gene expression in a single statistic.”

• cites [11] Agarwal, V. & Shendure, J. Predicting mRNA abundance directly from genomic sequence using deep convolutional neural networks. Cell Rep. 31, 107663 (2020). sourceNot checked — no full text of the source
• cites [12] Zhou, J. et al. Deep learning sequence-based ab initio prediction of variant effects on expression and disease risk. Nat. Genet. 50, 1171–1179 (2018). sourcePartly supported

  The Online Methods section on representative transcription start sites says most expression-profiling datasets were quantified at the gene level, including RNA-seq datasets. That supports the idea of summarizing expression, but it does not establish that all previous models worked with RNA-seq only after reducing expression to a single statistic. No page number is available in our copy. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [13] Avsec, Ž. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nat. Methods 18, 1196–1203 (2021). sourcePartly supported

  The Results section (page unavailable) identifies ExPecto as a previous model trained to predict gene-expression levels measured by RNA-seq. It does not say that all previous models used RNA-seq only after reducing expression to a single statistic. The Methods passage about summing CAGE counts describes this paper’s evaluation, not that broader claim about earlier RNA-seq models. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Borzoi enables direct characterization of tissue-specific cis-regulatory TF motifs by applying attribution methods to predicted RNA-seq coverage statistics.

“Borzoi enables direct characterization of tissue-specific cis-regulatory TF motifs by applying attribution methods to the predicted RNA-seq coverage statistics45–50.”

• cites [45] Talukder, A., Barham, C., Li, X. & Hu, H. Interpretation of deep learning in genomics and epigenomics. Brief. Bioinform. 22, bbaa177 (2021). sourceCould not verify

  The Chromatin interactions section describes saliency maps for Basenji, and the Gene expression prediction section discusses RNA-seq as a data source, but neither addresses Borzoi. Searches for Borzoi and for attribution of predicted RNA-seq coverage to tissue-specific TF motifs found no passage supporting the claim in our copy; page numbers were not available. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [46] Simonyan, K., Vedaldi, A. & Zisserman, A. Deep inside convolutional networks: visualising image classification models and saliency maps. In Proc. International Conference on Learning Representations (ICLR) 2014 (ICLR, 2014).Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [47] Sundararajan, M., Taly, A. & Yan, Q. Axiomatic attribution for deep networks. In Proc. 34th International Conference on Machine Learning (PMLR 70) (eds Precup, D. & Teh, Y. W.) 3319–3328 (Curran Associates, 2017). sourceSupported
• cites [48] Smilkov, D., Thorat, N., Kim, B., Viégas, F. & Wattenberg, M. SmoothGrad: removing noise by adding noise. In International Conference on Machine Learning, Workshop on Visualization for Deep Learning (ICML, 2017).Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [49] Shrikumar, A., Greenside, P. & Kundaje, A. Learning important features through propagating activation differences. In Proc. 34th International Conference on Machine Learning (PMLR 70) (eds Precup, D. & Teh, Y. W.) 3145–3153 (Curran Associates, 2017). sourcePartly supported

  Section 4.2 (pp. 7–8) shows that attribution methods can identify regulatory-protein motifs in a simulated DNA classification task. It does not address Borzoi, tissue-specific characterization, or attribution applied to predicted RNA-seq coverage statistics. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [50] Lundberg, S. M. & Lee, S.-I. A unified approach to interpreting model predictions. In Proc. 31st Conference on Neural Information Processing Systems (NIPS 2017) (eds. von Luxburg, U. et al.) 1–10 (Curran Associates, 2017). sourcePartly supported

  Section 2 (p. 2) describes methods that attribute a model’s output to its input features, supporting the reference to attribution methods. The cited paper does not establish that Borzoi applies those methods to predicted RNA-seq coverage statistics or that this characterizes tissue-specific cis-regulatory TF motifs. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Sequence-based models of post-transcriptional regulation rely on genome annotations and transformed measurements extracted from RNA-seq to isolate each regulatory mechanism.

“Similarly, sequence-based models of post-transcriptional regulation rely on genome annotations and transformed measurements extracted from RNA-seq to isolate each regulatory mechanism (for example, percent spliced-in for splicing)14–23.”

• cites [14] Cheng, J. et al. MMSplice: modular modeling improves the predictions of genetic variant effects on splicing. Genome Biol. 20, 48 (2019). sourcePartly supported

  The Background (page unavailable) describes percent spliced-in as an RNA-seq-derived measure used by sequence-based splicing models. The Results, “Modular modeling strategy” (page unavailable), says MMSplice trains separate splicing modules using genome annotations and measured splicing outcomes. These passages support the splicing example, but not the broader claim that models of post-transcriptional regulation generally use those inputs to isolate each regulatory mechanism. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [15] Jaganathan, K. et al. Predicting splicing from primary sequence with deep learning. Cell 176, 535–548.e24 (2019). sourceNot checked — no full text of the source
• cites [16] Zeng, T. & Li, Y. I. Predicting RNA splicing from DNA sequence using Pangolin. Genome Biol. 23, 103 (2022). sourcePartly supported

  The Methods section ('Generating training and test sets'; page unavailable) says Pangolin mapped RNA-seq reads using genome annotations and derived splice-site usage labels from those reads. Its training measurement was a Splice-Site Strength Estimate, not percent spliced-in; the paper discusses percent spliced-in for a separate evaluation using minigene assays. The source supports the splicing example in broad terms, but not the specific measurement example or the claim about post-transcriptional models generally. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [17] Leung, M. K. K., Delong, A. & Frey, B. J. Inference of the human polyadenylation code. Bioinformatics 34, 2889–2898 (2018). sourceSupported
• cites [18] Vainberg Slutskin, I., Weinberger, A. & Segal, E. Sequence determinants of polyadenylation-mediated regulation. Genome Res. 29, 1635–1647 (2019). sourcePartly supported

  The Discussion describes sequence-based models of polyadenylation-mediated regulation, and Methods > Computational analyses describes a transformed RNA-level measurement. But Results > High-throughput measurement of expression levels and cleavage maps says the models use measurements from a designed reporter library, not genome annotations and RNA-seq-derived measurements to isolate regulatory mechanisms. The paper does not address percent spliced-in or splicing. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [19] Arefeen, A., Xiao, X. & Jiang, T. DeepPASTA: deep neural network based polyadenylation site analysis. Bioinformatics 35, 4577–4585 (2019). sourceNot checked — no full text of the source
• cites [20] Li, Z. et al. DeeReCT-APA: prediction of alternative polyadenylation site usage through deep learning. Genomics Proteomics Bioinformatics 20, 483–495 (2022). sourcePartly supported

  The Methods section ('A genome-wide PAS quantification dataset'; page unavailable) describes a sequence-based model trained on PAS usage values calculated from 3′-mRNA sequencing reads, using a reference of identified PASs. This supports the claim for alternative polyadenylation, but the paper does not establish that sequence-based models generally rely on genome annotations and RNA-seq-derived measurements to isolate each post-transcriptional mechanism. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [21] Bogard, N., Linder, J., Rosenberg, A. B. & Seelig, G. A deep neural network for predicting and engineering alternative polyadenylation. Cell 178, 91–106.e23 (2019). sourcePartly supported

  The “APARENT Accurately Predicts Native APA” section (page unavailable) uses annotated APA events and RNA-seq read counts to estimate isoform proportions. But the Abstract and “A Massively Parallel Reporter Assay for APA” section say APARENT itself was trained on sequenced synthetic reporter constructs, not that it relies on genome annotations and transformed RNA-seq measurements to isolate each regulatory mechanism. The paper therefore supports an example of using those data, not the broader claim of reliance. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [22] Linder, J., Koplik, S. E., Kundaje, A. & Seelig, G. Deciphering the impact of genetic variation on human polyadenylation using APARENT2. Genome Biol. 23, 232 (2022). sourcePartly supported

  The Methods sections on endogenous datasets and tissue-specific modeling describe mapping RNA-seq reads to annotated polyadenylation sites and normalizing read counts into isoform proportions. The paper thus supports this approach for modeling polyadenylation, but its APARENT2 training data also come from a massively parallel reporter assay; it does not establish that sequence-based models of post-transcriptional regulation generally rely on genome annotations and RNA-seq transformations to isolate each mechanism. No page numbers are available in our copy. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [23] Agarwal, V. & Kelley, D. R. The genetic and biochemical determinants of mRNA degradation rates in mammals. Genome Biol. 23, 245 (2022). sourcePartly supported

  The Methods sections on gene annotations, half-life data pre-processing, and Saluki training describe a sequence-based model that uses annotated mRNA structure and transformed half-life measurements to study RNA degradation. The paper does not establish that all those measurements were extracted from RNA-seq, and it models mRNA decay rather than the cited sentence’s splicing example (percent spliced-in). (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Models trained on data from multiple species have been shown to improve performance.

“Models trained on data from multiple species have been shown to improve performance9, but chromatin profiling and the CAGE gene expression assays have been performed on far fewer species than RNA-seq.”

• cites [9] Kelley, D. R. Cross-species regulatory sequence activity prediction. PLoS Comput. Biol. 16, e1008050 (2020). sourceSupported

By processing a large region centered on the transcription start site, several models can predict normalized gene counts.

“By processing a large region centered on the transcription start site (TSS), several models can predict normalized gene counts11–13.”

• cites [11] Agarwal, V. & Shendure, J. Predicting mRNA abundance directly from genomic sequence using deep convolutional neural networks. Cell Rep. 31, 107663 (2020). sourceNot checked — no full text of the source
• cites [12] Zhou, J. et al. Deep learning sequence-based ab initio prediction of variant effects on expression and disease risk. Nat. Genet. 50, 1171–1179 (2018). sourceSupported
• cites [13] Avsec, Ž. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nat. Methods 18, 1196–1203 (2021). sourceSupported

Running MoDISco on gradients from tissue-pooled exon-to-intron coverage ratios for genes from the Gasperini set found known splice-regulatory motifs.

“When running MoDISco on gradients from tissue-pooled exon-to-intron coverage ratios for genes from the Gasperini set58, we found known splice-regulatory motifs (Fig. 7b).”

• cites [58] Gasperini, M. et al. A genome-wide framework for mapping gene regulation via cellular genetic screens. Cell 176, 377–390.e19 (2019). sourceSupported

• AssumptionThe authors state this as an assumption: The inverse transforms applied to Borzoi's raw predictions yield a tensor that can be reasonably assumed to reflect counts.

Reads aligning with greater density at the 3′ end of transcripts and other confounders (for example, GC bias) caused false positives when classifying alternatively used splice sites based on predicted coverage.

“For example, reads aligning with greater density at the 3′ end of transcripts83,84 and other confounders (for example, GC bias) caused false positives as we attempted to classify alternatively used splice sites based on predicted coverage.”

• cites [83] Roberts, A., Trapnell, C., Donaghey, J., Rinn, J. L. & Pachter, L. Improving RNA-seq expression estimates by correcting for fragment bias. Genome Biol. 12, R22 (2011). sourcePartly supported

  The Background (page unavailable) describes positional bias that can place fragments preferentially toward the beginning or end of transcripts, supporting the possibility of greater density at the 3′ end. It does not specifically report a 3′-end excess. The false positives in splice-site classification are presented as the citing authors’ own observation, not as a finding attributed to this source. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [84] Roberts, A. & Pachter, L. Streaming fragment assignment for real-time analysis of sequencing experiments. Nat. Methods 10, 71–73 (2013). sourcePartly supported

  The paper’s main text (labeled Abstract in our copy; page unavailable) says sequence bias affects fragment prevalence at specific positions, supporting the general point that RNA-seq coverage can be uneven. It does not establish greater read density at the 3′ end of transcripts. The sentence presents the false-positive splice-site classifications as the citing authors’ own experience, not as a finding of this source. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Similar results are obtained on the data from Gasperini et al. (2019).

“Similar results are obtained on the data from Gasperini et al. (2019)58 (Fig. 4d and Extended Data Fig. 5b).”

• cites [58] Gasperini, M. et al. A genome-wide framework for mapping gene regulation via cellular genetic screens. Cell 176, 377–390.e19 (2019). sourceSupported

When comparing Borzoi, Enformer and a distance-to-TSS baseline on classifying measured positive from negative enhancer-gene interactions, Borzoi has superior AUPRC and AUROC at all distances.

“When comparing Borzoi, Enformer and a distance-to-TSS baseline on their ability to classify measured positive from negative enhancer–gene interactions in data from previous works60–65, we find that Borzoi has superior average precision (AUPRC) and area under the receiver operating characteristic curve (AUROC) at all distances (Fig. 4c and Extended Data Fig. 5a).”

• cites [60] Fulco, C. P. et al. Activity-by-contact model of enhancer–promoter regulation from thousands of CRISPR perturbations. Nat. Genet. 51, 1664–1669 (2019). sourceSupported
• cites [61] Fulco, C. P. et al. Systematic mapping of functional enhancer–promoter connections with CRISPR interference. Science 354, 769–773 (2016). sourceSupported
• cites [62] Klann, T. S. et al. CRISPR–Cas9 epigenome editing enables high-throughput screening for functional regulatory elements in the human genome. Nat. Biotechnol. 35, 561–568 (2017). sourceSupported
• cites [63] Xie, S., Duan, J., Li, B., Zhou, P. & Hon, G. C. Multiplexed engineering and analysis of combinatorial enhancer activity in single cells. Mol. Cell 66, 285–299.e5 (2017). sourceNot checked — no full text of the source
• cites [64] Huang, J. et al. Dissecting super-enhancer hierarchy based on chromatin interactions. Nat. Commun. 9, 943 (2018). sourceSupported
• cites [65] Nasser, J. et al. Genome-wide enhancer maps link risk variants to disease genes. Nature 593, 238–243 (2021). sourceSupported

Pangolin has a slight advantage over Borzoi for classifying causal sQTLs from matched negatives (dAUPRC = 0.01, evaluated on all SNPs within distances of ≤10,000 bp from an annotated splice site).

“When comparing Borzoi to Pangolin16 for the task of classifying the causal sQTLs from matched negatives, Pangolin has a slight advantage (Fig. 7d–e and Supplementary Fig. 12c; dAUPRC = 0.01, evaluated on all SNPs within distances of ≤10,000 bp from an annotated splice site).”

• cites [16] Zeng, T. & Li, Y. I. Predicting RNA splicing from DNA sequence using Pangolin. Genome Biol. 23, 103 (2022). sourceSupported

Borzoi was consistently more accurate than APARENT2 at classifying paQTLs from matched negatives.

“Compared to variant scores predicted by the APARENT2 model22, Borzoi was consistently more accurate (Fig. 6f).”

• cites [22] Linder, J., Koplik, S. E., Kundaje, A. & Seelig, G. Deciphering the impact of genetic variation on human polyadenylation using APARENT2. Genome Biol. 23, 232 (2022). sourceSupported

Borzoi correctly predicts upregulation of SHTN1 expression owing to the creation of a CEBP binding motif.

“Borzoi correctly predicts the upregulation of SHTN1 expression owing to the creation of a CEBP binding motif69–72 (see Supplementary Fig. 9a and Extended Data Fig. 6a,b for additional examples).”

• cites [69] Burda, P., Laslo, P. & Stopka, T. The role of PU.1 and GATA-1 transcription factors during normal and leukemogenic hematopoiesis. Leukemia 24, 1249–1257 (2010). sourceNot checked — no full text of the source
• cites [70] Oikawa, T. et al. The role of Ets family transcription factor PU.1 in hematopoietic cell differentiation, proliferation and apoptosis. Cell Death Differ. 6, 599–608 (1999). sourceNot checked — no full text of the source
• cites [71] Yanai, H., Negishi, H. & Taniguchi, T. The IRF family of transcription factors: inception, impact and implications in oncogenesis. Oncoimmunology 1, 1376–1386 (2012). sourceCould not verify

  The retrieved passages from the References section concern interferon regulatory factors, not Borzoi or SHTN1. Searches for the claim’s wording, SHTN1 and CEBP, and Borzoi found no passage in our copy addressing the predicted upregulation or its proposed motif mechanism. (LLM, conf=0.77)

  Experimental — accuracy not yet measured
• cites [72] Tamura, T., Yanai, H., Savitsky, D. & Taniguchi, T. The IRF family transcription factors in immunity and oncogenesis. Annu. Rev. Immunol. 26, 535–584 (2008). sourceNot checked — no full text of the source

There is a correlation between codon-aggregated gradient saliencies of gene exon coverage and MPRA measurements from a previous publication (Pearson's R = 0.59).

“Although we generally do not find determinants of mRNA half-life in the 3′ UTR attributions, we do observe a correlation between codon-aggregated gradient saliencies of gene exon coverage and MPRA measurements from a previous publication78 (Pearson’ R = 0.59) (Supplementary Fig. 11a).”

• cites [78] Forrest, M. E. et al. Codon and amino acid content are associated with mRNA stability in mammalian cells. PLoS ONE 15, e0228730 (2020). sourceCould not verify

  The Results section describes codon-dependent mRNA stability measured with individual reporter constructs, and the Methods describe their measurement by northern blot. Searches for “MPRA,” “massively parallel,” and “0.59” found no passage identifying MPRA measurements in this paper. The cited paper also cannot verify the citing authors’ correlation with their own gradient saliencies. (LLM, conf=0.75)

  Experimental — accuracy not yet measured

The paper did not find sequence elements related to mRNA half-life in Borzoi's sequence attributions.

“Furthermore, we did not find sequence elements related to mRNA half-life in Borzoi’s sequence attributions23,82.”

• cites [23] Agarwal, V. & Kelley, D. R. The genetic and biochemical determinants of mRNA degradation rates in mammals. Genome Biol. 23, 245 (2022). sourcePartly supported

  The Results section, ‘A deep learning-based genetic model of mRNA half-life,’ identifies sequence motifs associated with changes in half-life. It analyzes Saluki, however, and does not report a failure to find those elements in Borzoi’s sequence attributions; the citing authors’ ‘we did not find’ is not a finding of this source. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [82] Zhao, W. et al. Massively parallel functional annotation of 3′ untranslated regions. Nat. Biotechnol. 32, 387–391 (2014). sourceSupported

Maximizing the PIP-weighted inverse distance outperforms minimizing the PIP-weighted distance for predicting a credible set's target gene.

“Maximizing the PIP-weighted inverse distance outperforms the previously described approach of minimizing the PIP-weighted distance105.”

• cites [105] Mountjoy, E. et al. An open approach to systematically prioritize causal variants and genes at all published human GWAS trait-associated loci. Nat. Genet. 53, 1527–1533 (2021). sourceSupported

Higher-order structures matching GENCODE annotations were generally found in the later self-attention layers of Borzoi.

“After exploring the predicted attention maps for several different loci, we noticed that higher-order structures matching GENCODE annotations104 were generally found in the later self-attention layers.”

• cites [104] Harrow, J. et al. GENCODE: the reference human genome annotation for The ENCODE Project. Genome Res. 22, 1760–1774 (2012). sourceSupported

Including experimental assays studied by Enformer and Basenji in training data helps the model use important sequence features for making RNA coverage predictions.

“To help the model use important sequence features for making its RNA coverage predictions, we also included the experimental assays studied by the Enformer and Basenji models in the training data8,9,13.”

• cites [8] Kelley, D. R. et al. Sequential regulatory activity prediction across chromosomes with convolutional neural networks. Genome Res. 28, 739–750 (2018). sourceSupported
• cites [9] Kelley, D. R. Cross-species regulatory sequence activity prediction. PLoS Comput. Biol. 16, e1008050 (2020). sourceSupported
• cites [13] Avsec, Ž. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nat. Methods 18, 1196–1203 (2021). sourceSupported

By applying attribution methods to derived statistics, the authors extract cis-regulatory motifs driving RNA expression and post-transcriptional regulation in normal tissues.

“By applying attribution methods to the derived statistics, we extract cis-regulatory motifs driving RNA expression and post-transcriptional regulation in normal tissues.”

Using statistics derived from Borzoi's predicted coverage, the authors isolate and accurately score DNA variant effects across multiple layers of regulation, including transcription, splicing and polyadenylation.

“Using statistics derived from Borzoi’s predicted coverage, we isolate and accurately score DNA variant effects across multiple layers of regulation, including transcription, splicing and polyadenylation.”

Borzoi is a model that learns to predict cell-type-specific and tissue-specific RNA-seq coverage from DNA sequence.

“Here, we introduce Borzoi, a model that learns to predict cell-type-specific and tissue-specific RNA-seq coverage from DNA sequence.”

Predicted tissue-specific coverage ratio fold changes showed moderate correlation with measured fold changes between GTEx tissues (Spearman's R = 0.23−0.41).

“Predicted tissue-specific coverage ratio fold changes showed moderate correlation with measured fold changes between GTEx tissues (Spearman’s R = 0.23−0.41; Fig. 2f and Supplementary Fig. 3h).”

The model's attention matrices comprehensively capture gene structure.

“In the Supplementary Information, we show that the model relies on well-known regulatory features to make predictions and that the model’s attention matrices comprehensively capture gene structure (Extended Data Fig. 2 and Supplementary Fig. 2).”

SGK1 harbors an upstream TSS that is highly expressed in brain but not blood.

“For example, SGK1 harbors an upstream TSS that is highly expressed in brain but not blood (Fig. 2c; see Extended Data Fig. 3a for additional examples).”

Pearson's R increased to 0.75 when averaging the predictions across the full ensemble.

“Pearson’s R increased to 0.75 when averaging the predictions across the full ensemble (Fig. 1c).”

Borzoi accurately predicts variation within transcript structure; on the top 20% of test set genes with the highest variance in coverage across exons and introns, the average Pearson's R between predicted and measured RNA coverage at the bin level was 0.88 across all genes and samples.

“Finally, we note that Borzoi accurately predicts variation within the transcript structure; evaluated on the top 20% of test set genes with the highest variance in coverage across the span of exons and introns, the average Pearson’s R value between predicted and measured RNA coverage (at the bin level) was 0.88 across all genes and samples (Supplementary Fig. 1e).”

Borzoi introduced new capabilities to predict variant effects on splicing and polyadenylation that match or exceed the state of the art.

“Our model improved performance relative to Enformer on downstream tasks to identify distal enhancers and predict genetic variant effects on gene expression, and it introduced new capabilities to predict variant effects on splicing and polyadenylation that match or exceed the state of the art.”

After quantile-normalizing predictions across experiments and subtracting each gene's mean expression, the mean Pearson's R is 0.58 (0.55 per replicate).

“After quantile-normalizing the predictions across experiments and subtracting each gene’s mean expression (so that the value represents the residual expression beyond the mean), we observe a mean Pearson’s R of 0.58 (0.55 per replicate) (Fig. 1e), indicating that the model explains a significant amount of variation observed between tracks (such as tissue-specific and cell-type-specific differences”

The model relies on well-known regulatory features to make predictions.

“In the Supplementary Information, we show that the model relies on well-known regulatory features to make predictions and that the model’s attention matrices comprehensively capture gene structure (Extended Data Fig. 2 and Supplementary Fig. 2).”

Borzoi effectively learns several layers of gene regulation.

“We developed a model, named Borzoi, that effectively learns several layers of gene regulation.”

Borzoi predicts exon-intron coverage patterns with striking concordance for even long genes with many exons, as exemplified by the 190 kb gene INSR.

“Despite the challenges involved with modeling RNA-seq coverage from only underlying DNA sequence, Borzoi predicts exon–intron coverage patterns with striking concordance for even long genes with many exons, as exemplified in Fig. 1b by the 190 kb gene INSR.”

Borzoi could accurately predict tissue-specific gene expression coverage on held-out test genes.

“We first noted that Borzoi could accurately predict tissue-specific gene expression coverage on held-out test genes (for example, the blood-specific gene ADGRE1 visualized in Fig. 2a; see also Supplementary Fig. 3a,b).”

Predicted and measured fold change in gene-level coverage of one tissue relative to the average coverage of the four other tissues had a Spearman's R range from 0.52 to 0.75 when using the ensemble of four model replicates.

“We compared the predicted and measured fold change in gene-level coverage of one tissue relative to the average coverage of the four other tissues, observing a Spearman’s R range from 0.52 to 0.75 when using the ensemble of four model replicates (Fig. 2b).”

Borzoi improved performance relative to Enformer on downstream tasks to identify distal enhancers and predict genetic variant effects on gene expression.

“Our model improved performance relative to Enformer on downstream tasks to identify distal enhancers and predict genetic variant effects on gene expression, and it introduced new capabilities to predict variant effects on splicing and polyadenylation that match or exceed the state of the art.”

When comparing predicted to measured gene-level coverage values, the mean Pearson's R is 0.87 across held-out genes (0.86 per model replicate).

“When comparing predicted to measured gene-level coverage values, we observe a mean Pearson’s R of 0.87 across held-out genes (0.86 per model replicate) (Fig. 1d and Supplementary Fig. 1a–d).”

Test set predictions matched RNA-seq coverage with a mean Pearson's R value of 0.74 across human samples when using one model replicate.

“Test set predictions matched RNA-seq coverage with a mean Pearson’s R value of 0.74 across human samples when using one model replicate.”

Combining Borzoi and CADD scores resulted in the highest accuracy (mean AUROC = 0.57).

“Combining their scores resulted in the highest accuracy (mean AUROC = 0.57).”

Motifs shared between tissues generally tend to regulate distinct loci.

“Motifs shared between tissues generally tend to regulate distinct loci (Fig. 3b, inset).”

Borzoi exhibits competitive performance compared to Enformer when predicting non-coding regulatory mutations in promoters and enhancers as measured by massively parallel reporter assays (MPRAs).

“In the Supplementary Information, we show that Borzoi exhibits competitive performance compared to Enformer when predicting non-coding regulatory mutations in promoters and enhancers as measured by massively parallel reporter assays (MPRAs) (Supplementary Fig. 10).”

Motifs for well-known polyadenylation regulators (for example, CFIm, CPSF, CstF) emerge from the attribution scores of the predicted distal polyadenylation ratio.

“Motifs for well-known polyadenylation regulators (for example, CFIm, CPSF, CstF) emerge from the attribution scores of the predicted distal polyadenylation ratio (Fig. 6b).”

Borzoi recapitulates known regulatory motifs for esophagus and K562.

“We similarly recapitulate known regulatory motifs for esophagus and K562 (Supplementary Fig. 5c–e).”

Training on DNase-seq and ATAC-seq data in addition to RNA-seq delivered better concordance with eQTL measurements and enhancer-gene linking data.

“and delivered better concordance with eQTL measurements and enhancer–gene linking data.”

Variant prediction quality was only marginally affected by whether or not the variant occurs in genomic sequences seen during training.

“Variant prediction quality was only marginally affected by whether or not the variant occurs in genomic sequences seen during training”

Aggregated differences in gradient saliency for each pair of tissues among seqlets matching each TF were highly correlated with observed TPM fold changes for the corresponding TFs.

“These scores were highly correlated with observed TPM fold changes for the corresponding TFs (Fig. 3c and Supplementary Fig. 5f,g).”

These predictions were better correlated with expression than LMNB1 DamID-seq.

“These predictions were better correlated with expression than LMNB1 DamID-seq, which measures nuclear lamina interactions and constitutes a strong baseline.”

At closer distances, the average rank of all model predictions (Borzoi, APARENT2 and Saluki) surpasses either model's individual performance.

“At closer distances, the average rank of all model predictions (Borzoi, APARENT2 and Saluki) surpasses either model’s individual performance.”

Borzoi outperforms Enformer with even a single model (mean R = 0.292).

“Borzoi outperforms Enformer with even a single model (mean R = 0.292).”

Borzoi demonstrates competitive performance to state-of-the-art models in classifying fine-mapped QTLs.

“and demonstrates competitive performance to state-of-the-art models in classifying fine-mapped QTLs.”

Training on DNase-seq and ATAC-seq data in addition to RNA-seq, as well as mouse data, substantially improved predictions.

“In ablation experiments, we found that training on DNase-seq and ATAC–seq data in addition to RNA-seq, as well as mouse data, substantially improved predictions (Supplementary Fig. 9d).”

Borzoi detects well-known regulators for each tissue, such as SPI1/B and IRF4/8 for blood, HNF4A/G and HNF1A for liver, SOX9 and REST for brain and MYOD1 and MEF2D for muscle.

“We detect well-known regulators for each tissue, such as SPI1/B and IRF4/8 for blood, HNF4A/G and HNF1A for liver, SOX9 and REST for brain and MYOD1 and MEF2D for muscle.”

Including training data such as DNase-seq and ATAC-seq in addition to RNA-seq improves performance.

“Through ablation experiments, we find that including training data such as DNase-seq and ATAC–seq in addition to RNA-seq improves performance (Supplementary Fig. 7a–c).”

Training on DNase-seq and ATAC-seq data in addition to RNA-seq consistently improved test set accuracies compared to training on RNA-seq alone.

“Through a number of ablation studies, we discovered that training on DNase-seq and ATAC–seq data in addition to RNA-seq consistently improved test set accuracies compared to training on RNA-seq alone”

Restricted to ENCODE candidate cis-regulatory elements, Borzoi and CADD exhibited equal discriminative power (mean AUROC = 0.55).

“Restricted to ENCODE candidate cis-regulatory elements, Borzoi and CADD exhibited equal discriminative power (mean AUROC = 0.55; Fig. 5e and Supplementary Fig. 9f).”

Spearman's R reached 0.77 when comparing TF saliency in blood and muscle.

“For example, Spearman’s R reached 0.77 when comparing TF saliency in blood and muscle.”

Scores derived from DNase tracks were most concordant with measured expression levels (20-fold cross-validation, Spearman's R = 0.58 for promoter ARHGEF9).

“In general, the scores derived from DNase tracks were most concordant with the measured expression levels (Fig. 4e and Supplementary Fig. 8c; 20-fold cross-validation, Spearman’s R = 0.58 for promoter ARHGEF9).”

The model performed, at best, marginally better than a TSS distance baseline at prioritizing true eGenes among other genes surrounding an eQTL.

“The model performed, at best, marginally better than a TSS distance baseline.”

The performance gap between Borzoi and APARENT2 decreased when APARENT2's predictions were scaled by the reference isoform percent from PolyADB.

“However, the performance gap decreased when scaling APARENT2’s predictions by the reference isoform percent from PolyADB, suggesting that context is an important determinant.”

Window-shuffled ISM is a more reliable attribution method in 3′ UTRs because of buffering effects.

“We also note that window-shuffled ISM is a more reliable attribution method in 3′ UTRs because of buffering effects (Supplementary Fig. 11b).”

Using Borzoi's ensemble with an L2 score was superior to Enformer and its original sum aggregation at discriminating eQTLs (mean AUROC = 0.794 vs 0.747 across tissues).

“Using Borzoi’s ensemble with an L2 score was superior to Enformer and its original sum aggregation at discriminating eQTLs (mean AUROC = 0.794 vs 0.747 across tissues; Fig. 5b,c).”

Buffering effects were less problematic when interpreting repeat-like splicing motifs with ISM.

“Buffering effects were less problematic when interpreting repeat-like splicing motifs with ISM (Supplementary Fig. 12a).”

• AssumptionThe authors state this as an assumption: The inverse transforms applied to Borzoi's raw predictions yield a tensor that can be reasonably assumed to reflect counts.

Borzoi correctly predicts the change in site usage caused by each of two paQTLs when compared to RNA-seq tracks of GTEx individuals harboring the alternative allele.

“Compared to RNA-seq tracks of GTEx individuals harboring the alternative allele, Borzoi correctly predicts the change in site usage caused by each variant.”

Variant effect scores derived from Borzoi's predicted RNA-seq tracks discriminated fine-mapped paQTLs from expression-matched negatives with a monotonic increase in accuracy at closer distances to the nearest PAS (AUPRC = 0.64-0.74).

“The variant effect scores derived from the predicted RNA-seq tracks discriminated paQTLs from the matched negatives with a monotonic increase in accuracy at closer distances to the nearest PAS (Fig. 6e; AUPRC = 0.64–0.74).”

Borzoi is better than Pangolin at distances closer to the junction (dAUPRC = 0.02, evaluated on variants ≤200 bp from an annotated junction).

“By contrast, Borzoi is better at distances closer to the junction (Fig. 7d–e; dAUPRC = 0.02, evaluated on variants ≤200 bp from an annotated junction).”

The authors generally do not find determinants of mRNA half-life in the 3′ UTR attributions.

“Although we generally do not find determinants of mRNA half-life in the 3′ UTR attributions, we do observe a correlation between codon-aggregated gradient saliencies of gene exon coverage and MPRA measurements from a previous publication78 (Pearson’ R = 0.59) (Supplementary Fig. 11a).”

Borzoi is a new sequence-based machine-learning model that learns to predict sequencing coverage from a vast set of RNA-seq experiments.

“In this paper, we propose a new sequence-based machine-learning model, Borzoi, that learns to predict sequencing coverage from a vast set of RNA-seq experiments.”

Borzoi predicts fine-mapped causal intronic paQTLs well, with an average AUPRC of 0.725.

“Borzoi predicts fine-mapped causal intronic paQTLs well, with an average AUPRC of 0.725 (Extended Data Fig. 9a,b and Supplementary Fig. 13a).”

Borzoi exhibits greater Spearman correlation than Enformer when comparing effect size predictions to fine-mapped eQTL coefficients (mean R = 0.334 vs 0.227).

“Borzoi also exhibits greater Spearman correlation than Enformer when comparing effect size predictions to fine-mapped eQTL coefficients (mean R = 0.334 across tissues vs R = 0.227; Fig. 5d and Supplementary Fig. 9b,c).”

Gradients at the position of maximal liver-specific saliency for gene CFHR2 highlight motif hits for CEBPA/B and HNF4A/G.

“As an example, gradients at the position of maximal liver-specific saliency for gene CFHR2 highlight motif hits for CEBPA/B and HNF4A/G (Fig. 3a).”

The L2 score achieves greater performance than the SUM score on the fine-mapped eQTL classification task for Borzoi.

“The L2 score extracts more information and achieves greater performance on this task for Borzoi.”

Enformer also benefits from the L2 score on the fine-mapped eQTL classification task, though less than Borzoi.

“All previous Enformer work uses the SUM score, but we observed here that it also benefits from L2, though less than Borzoi.”

Window-shuffled ISM worked better for 3′ UTRs owing to buffering effects.

“we found that window-shuffled ISM worked better for 3′ UTRs owing to buffering effects.”

• AssumptionThe authors state this as an assumption: The inverse transforms applied to Borzoi's raw predictions yield a tensor that can be reasonably assumed to reflect counts.

This approach improved performance for both Enformer and Borzoi compared to a simpler strategy of averaging the absolute-valued gradients in a 2 kb window centered on the enhancer.

“This approach improved performance for both Enformer and Borzoi compared to a simpler strategy of averaging the absolute-valued gradients in a 2 kb window centered on the enhancer.”

After filtering for genes with ≥3 elements tested, the scRNA-seq dataset contains 404 positives of 19,104 tested within 262 kb of the gene's TSS.

“After filtering for genes with ≥3 elements tested, the scRNA-seq dataset contains 404 positives of 19,104 tested within 262 kb of the gene’s TSS.”

When deriving statistics from Borzoi's DNase or histone modification tracks, inserting the full TRIP reporter (rather than only the promoter) made predictions marginally worse.

“By contrast, when deriving statistics from Borzoi’s DNase or histone modification tracks (for example, H3K4me3) we only inserted the promoter, as these predictions became marginally worse when inserting the full reporter.”

The difference in Spearman's R between the 128 bp window and the 4,096 bp window was small (for example, <0.02 for ARHGEF98).

“the difference in Spearman’s R was small (for example, <0.02 for ARHGEF98)”

Test accuracies for Borzoi versus Enformer on overlapping datasets were broadly similar despite modifications.

“Nevertheless, we examined test accuracies for Borzoi versus Enformer (v.3.0) on these overlapping datasets and found them to be broadly similar despite these modifications (Extended Data Fig. 1a–d).”

The cell-type mapping changes led to better performance for Borzoi.

“These changes led to better performance for Borzoi and were reasonable choices with respect to the target genes’ expression patterns.”

Many of the fine-mapped paQTL positives are shared between tissues.

“Given that many of the positives are shared between tissues (there are a total of 1,058 unique paQTLs, each occurring in at least one tissue)”

Input gradients and ISM produced high-quality attributions for splice junctions and enhancer-promoter regions.

“Although input gradients and ISM produced high-quality attributions for splice junctions and enhancer–promoter regions”

• AssumptionThe authors state this as an assumption: The inverse transforms applied to Borzoi's raw predictions yield a tensor that can be reasonably assumed to reflect counts.

RNA-seq reads aligned across a transcript depend on a much larger region of sequence containing the gene's exons and relevant cis-regulatory elements.

“By contrast, the most popular sequencing assay, RNA sequencing (RNA-seq), does not have this property; RNA-seq reads aligned across a transcript will depend on a much larger region of sequence containing the gene’s exons and relevant cis-regulatory elements.”

• AssumptionRests on an assumption: RNA-seq coverage patterns integrate multiple layers of gene regulation: transcription, splicing, termination or polyadenylation and RNA stability.

Borzoi explains a significant amount of variation observed between tracks, such as tissue-specific and cell-type-specific differences.

“After quantile-normalizing the predictions across experiments and subtracting each gene’s mean expression (so that the value represents the residual expression beyond the mean), we observe a mean Pearson’s R of 0.58 (0.55 per replicate) (Fig. 1e), indicating that the model explains a significant amount of variation observed between tracks (such as tissue-specific and cell-type-specific differences”

Chromatin profiling and CAGE gene expression assays have been performed on far fewer species than RNA-seq.

“Models trained on data from multiple species have been shown to improve performance9, but chromatin profiling and the CAGE gene expression assays have been performed on far fewer species than RNA-seq.”

Effective RNA-seq modeling requires working with very large sequences and algorithms that propagate information across large distances.

“Given that mammalian genes often span hundreds of thousands of nucleotides, effective RNA-seq modeling requires working with very large sequences and algorithms that propagate information across large distances.”

• AssumptionRests on an assumption: Mammalian genes often span hundreds of thousands of nucleotides.

Borzoi enables variant scoring and interpretation through multiple layers of regulation, including transcription, splicing and polyadenylation.

“Borzoi enables variant scoring and interpretation through multiple layers of regulation, including transcription, splicing and polyadenylation”

Applying sequence attribution methods to statistics derived from predicted coverage tracks provides tissue-specific interpretations of enhancers driving RNA expression and post-transcriptional regulation within the transcript.

“By applying sequence attribution methods to statistics derived from the predicted coverage tracks, Borzoi provides tissue-specific interpretations of enhancers driving RNA expression and post-transcriptional regulation within the transcript.”

Context is an important determinant of paQTL prediction performance.

“suggesting that context is an important determinant.”

Genetics researchers can ignore whether a variant occurs in genomic sequences seen during training when using the model.

“meaning that genetics researchers can ignore this factor when using the model.”

Borzoi's splice-gain predictions appear less well-calibrated.

“whereas Borzoi’s splice–gain predictions appear less well-calibrated.”

The subsequent analyses are still unbiased because the evaluations are done on out-of-domain measurements not trained on.

“We argue that these are still unbiased analyses, as the evaluations are done on out-of-domain measurements not trained on (for example, the alternative alleles of fine-mapped QTLs and their estimated effects were not part of the training data).”

Modeling coverage patterns around exons at 128 bp resolution can obstruct splice site learning.

“modeling the coverage patterns around these exons at such a coarse resolution can obstruct splice site learning.”

• AssumptionRests on an assumption: Mammalian exons regularly cover fewer than 128 bp.

The worsening of Borzoi's DNase/histone predictions when inserting the full TRIP reporter is attributable to the PiggyBac transposable elements flanking the reporter, which Borzoi inherently does not predict well owing to the clipping of unmappable regions during the original training data processing.

“We attribute this phenomenon to the PiggyBac transposable elements flanking the reporter, which Borzoi inherently does not predict well owing to the clipping of unmappable regions during the original training data processing.”

The alternative alleles of fine-mapped QTLs and their estimated effects were not part of the training data.

“(for example, the alternative alleles of fine-mapped QTLs and their estimated effects were not part of the training data).”