Biology subjects

Linder, J.

Publications and source records attributed to Linder, J..

16 recordsLinked to original sources

bioRxiv2026

A high-resolution protein-RNA binding atlas for structure-aware interaction prediction and post-transcriptional regulatory modeling

RNA-binding proteins (RBPs) control post-transcriptional gene expression by regulating RNA processing, localization, stability, and translation. However, our understanding of RNA regulation remains limited by the resolution, scale, and uniformity of available RBP binding maps. Here we introduce eclair CLIP (ecCLIP), a high-throughput CLIP-seq platform that combines endogenous protein tagging for uniform affinity capture with ligation-free library preparation to generate high-resolution RNA-protein binding profiles. We applied ecCLIP to 67 endogenously tagged RBPs in HEK293T cells, generating a high-quality transcriptome-scale atlas that recovered known binding preferences and captured post-transcriptional regulons linked to RNA half-life. To test the utility of the atlas, we incorporated ecCLIP into Borzuki, a Borzoi-derived sequence-to-function model trained on complementary RNA degradation, expression, CLIP, and Perturb-seq data. Inclusion of ecCLIP data improved RNA half-life prediction and assessment of 3' UTR variant effects on RNA stability, demonstrating that our atlas helps capture RNA dynamics. To directly model the cis-regulatory grammar underlying these RBP-RNA interactions, we developed Pythia, a structure-aware neural network architecture that incorporates context-free grammar-inspired fixed convolutions to explicitly encode RNA base-pairing features. Pythia predicted RBP binding from sequence and structural context and decomposed predictions into sequence- and structure-dependent components. We tested this decomposition for five RBPs using massively parallel reporter CLIP (MPRC) binding assays in cells, using synthetic RNA libraries containing mutations designed to preserve or disrupt predicted RNA structures across nominated binding sites. We used the same MPRC strategy to confirm allele-specific binding effects for variants prioritized by Pythia. Together, ecCLIP and Pythia provide a resource for structure-aware decoding of RNA-protein interaction grammar and interpretation of post-transcriptional regulatory variation.

genomics↗

bioRxiv2026

Cerberus: bidirectional state space blocks improve accuracy and efficiency of regulatory sequence models

Sequence-to-function models predict regulatory activity and variant effects from DNA sequence alone, and leading models such as Borzoi compute long-range interactions with self-attention. The reference genome caps unique training sequences, and new assays add labels, not sequences. This constraint favors architectures that learn more from limited sequence diversity. We compared long-range blocks with the rest of the architecture and the training data held fixed, replacing Borzoi's transformer blocks with Hydra, a bidirectional state space block built on Mamba-2. Hydra blocks ran 5.7-fold faster on 8,192-token inputs, and models built on them predicted binned coverage on held-out sequences and classified fine-mapped eQTLs more accurately. Hydra's linear cost in sequence length let us run the blocks at 32 bp resolution and remove the U-net decoder, leaving a convolution-SSM model smaller and more accurate than the one it replaced. Scaling this architecture up on an augmented track collection produced Cerberus, an ensemble of eight models over 786 kb of input. We evaluated it on GTEx v11 fine-mapped eQTL, sQTL, and paQTL benchmarks with negatives matched on allele frequency, gene expression, and phenotype-specific positional annotations to reduce confounding by these properties. Cerberus is more accurate than Borzoi on all three phenotypes, 0.692 versus 0.668 mean eQTL AUPRC across 48 tissues, and estimates eQTL effect sizes better, 0.379 versus 0.321 Spearman $\rho$. Against AlphaGenome, it trades leads in eQTL classification, ahead on variants 3 to 100 kb from the transcription start site and behind in coding sequence and within 3 kb. Interpretation of the trained blocks shows interaction range growing with block depth, from a 0.9 kb median half-life in the first block to 79 kb in the seventh, and the deepest heads anchoring on promoters, enhancers, and CTCF peaks. We release the models, the training data, the training and evaluation code, and the benchmark sets, supporting variant scoring and transfer learning.

genomics↗

bioRxiv2026

Scalable saturation mutagenesis reveals gene regulatory architecture and rare variant effects

Long-context sequence-to-function models enable nucleotide-resolution prediction of regulatory variant effects, motivating comprehensive mutational interrogation across broad genomic contexts. Yet conventional in silico saturation mutagenesis (ISM) scores mutations one at a time, requiring millions of model evaluations for a single gene and billions to trillions at genome scales. Here, we introduce Multi-ISM, a scalable framework that reformulates ISM as a sparse recovery problem. Multi-ISM generates mutational maps using 45-fold fewer model evaluations than exhaustive single-variant ISM and matches or exceeds its accuracy on variant-effect benchmarks. Multi-ISM is architecture-agnostic and transfers across large-scale sequence-to-function models. We applied Multi-ISM to 5,000 protein-coding genes, including 3,317 OMIM disease genes, generating base-pair-resolution, tissue-resolved attribution maps across 500-kb windows. These maps supported enhancer--gene prioritization and identification of cell-type-specific regulatory elements. Gene-level summaries of the maps captured regulatory complexity and showed that more constrained genes had smaller predicted mutational effects. Aggregating Multi-ISM predictions into gene-level rare-variant burdens improved personalized expression prediction over a common-variant elastic net, with the largest gains at expression outliers. Multi-ISM makes nucleotide-resolution interpretation of long-context sequence models a routine computation rather than a dedicated effort, so that new architectures, functional readouts, and cellular contexts can be mapped as they appear.

genomics↗

bioRxiv2025

Massively parallel assay of human splice variants reveals cis-regulatory drivers of disease-associated and cell type-specific splicing regulation

Splice-disrupting variants (SDVs) underlie many human diseases, yet systematic functional maps of their effects across cell types remain limited. We developed Cell-type Oriented Massively Parallel reporter Assay of Splicing Signatures (COMPASS) to measure splicing outcomes for 87,546 single and double variants across more than 1,700 genes in five human cell lines of diverse tissue origin. COMPASS targets disease relevant gene sets, including ACMG actionable genes and SFARI autism-associated genes, enabling systematic dissection of splicing impacts in health and disease. Our measurements reveal numerous SDVs, including ClinVar variants currently classified as variants of uncertain significance. Using prime editing, we validate variant effects in the genome for ClinVar variants in BIN1, a gene implicated in Alzheimers disease, cancer, and cardiac pathology. Benchmarking our MPRA against state-of-the-art predictive models shows that while these approaches capture broad variant effects, important gaps remain. Our data reveal putative RNA-binding protein motifs whose disruption drives splicing changes, providing mechanistic insight into variant impact. Finally, cross-cell line comparisons reveal a distinct subset of variants that drive cell type-specific splicing programs. This study delivers the largest cell type-resolved, base-resolution atlas of splicing variant effects to date, providing a resource to support variant reclassification in clinical genomics and aid in the selection of therapeutic targets. Graphical abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=200 SRC="FIGDIR/small/681955v1_ufig1.gif" ALT="Figure 1"> View larger version (60K): org.highwire.dtl.DTLVardef@81765dorg.highwire.dtl.DTLVardef@181e55dorg.highwire.dtl.DTLVardef@f96eaeorg.highwire.dtl.DTLVardef@1d16d4a_HPS_FORMAT_FIGEXP M_FIG C_FIG

synthetic biology↗

bioRxiv2025

Predicting dynamic expression patterns in budding yeast with a fungal DNA language model

Predicting gene expression from DNA sequence remains challenging due to complex regulatory codes. We introduce a masked DNA language model pretrained on 165 fungal genomes closely related to budding yeast that captures conserved regulatory grammar. Fine-tuning the LM on yeast RNA-seq data--including high-resolution transcriptional regulator induction time courses generated in this study--yielded Shorkie, a model that substantially improves gene expression prediction compared to baselines trained without self-supervision. Shorkie identified canonical transcription factor (TF) binding motifs and tracked their usage across induction experiments. Furthermore, Shorkie accurately predicted variant effects, outperforming leading sequence-to-expression models in cis-eQTL classification and achieving high concordance with massively parallel reporter assays. Interpretability analyses revealed Shorkies ability to resolve promoter dynamics, splicing signals, and temporal changes in regulatory motif usage. This framework demonstrates that evolutionary-scale pretraining combined with transfer learning substantially improves our ability to decode gene regulation from sequence, providing insights into noncoding variants and regulatory networks.

genomics↗

bioRxiv2025

Predicting cell type-specific coverage profiles from DNA sequence

Predicting expression profiles from RNA-seq experiments provides a powerful approach for universal sequence-based variant effect prediction, enabling researchers to score variants that affect total gene expression and relative isoform abundances. These models can be repurposed for new prediction tasks through transfer learning. However, current base models train primarily on bulk RNA-seq profiles derived from tissues and cell lines, overlooking the wealth of single-cell 3-seq data that captures cell type-specific gene regulation. Here, we extend the capabilities of our recently developed Borzoi model by training on single-cell 3-seq expression profiles from the Tabula Sapiens, Tabula Muris, and the Adult Brain Atlas, aggregated by cell type. This new model, Borzoi Prime, enables accurate variant interpretation across diverse cells, spanning erythrocytes to microglia. Training on 3-seq profiles improves the models ability to predict cell- and tissue-specific alternative polyadenylation, even in the original bulk RNA-seq data. Through UTR-wide mutagenesis experiments of alternatively polyadenylated genes, we highlight determinants of cell type-specific 3 UTR regulation learned by the model. This cell type-resolved approach opens new possibilities for understanding genetic variant effects via multiple layers of regulation in specific cellular contexts.

genomics↗

bioRxiv2025

Parameter-Efficient Fine-Tuning of a Supervised Regulatory Sequence Model

1DNA sequence deep learning models accurately predict epigenetic and transcriptional profiles, enabling analysis of gene regulation and genetic variant effects. While large-scale training models like Enformer and Borzoi are trained on abundant data, they cannot cover all cell states and assays, necessitating new model training to analyze gene regulation in novel contexts. However, training models from scratch for new datasets is computationally expensive. In this study, we systematically evaluate a transfer learning framework based on parameter-efficient fine-tuning (PEFT) approaches for supervised regulatory sequence models. We focus on the recently published state-of-the-art model Borzoi and fine-tune to new RNA-seq datasets, enabling accurate and efficient analysis of gene regulation across different biological systems. Our results demonstrate that PEFT substantially improves memory and runtime efficiency while achieving high accuracy. The transferred models effectively predict held-out gene expression changes, identify regulatory drivers of differentially expressed genes, and predict cell-type-specific variant effects. Our findings underscore the potential of PEFT to enhance the utility of large-scale sequence models like Borzoi for broader application to study genetic variants and regulatory drivers in any functional genomics dataset. Code for efficient Borzoi transfer is available in the Baskerville codebase https://github.com/calico/baskerville.

bioinformatics↗

bioRxiv2025

Selective State Space Models Outperform Transformers at Predicting RNA-Seq Read Coverage

Transformers are the basis for many state-of-the-art machine learning tools, including those for predicting gene expression data from DNA sequence. The considerable time and cost of training transformer models has motivated development of alternative approaches inspired by ideas from the signal-processing literature, such as state-space models (Mamba), Fourier transforms (Hyena), and wavelet transforms (MultiResNet). To evaluate these methods as potential replacements (or complements) for attention, we developed a software library bilby, implemented using Python and Jax/Flax, providing convolutional, attention, bidirectional Hyena, bidirectional Mamba, and striped-architecture models for supervised multi-task learning in functional genomics. We report a comparison of these architectures, testing several hyperparameters and variations, and reporting performance statistics for the withheld test set as well as downstream SNP classifiers. Relative to models comprising convolution and attention layers (implemented in Python and TensorFlow via the Baskerville library used by the Borzoi software), models comprising convolutional, bidirectional Mamba, and (optionally) attention layers achieve small but consistent improvements in prediction accuracy, for roughly comparable training times and parameter counts, when averaged across all output tracks and data splits (a proportional increase of 3-4% in Pearson R, and 1-2% in r2, with the highest gains achieved when Mamba and attention layers were combined in a striped architecture). In contrast, Hyena (when reimplemented as described in the literature) was not competitive with attention-based models at these tasks, while MultiResNet proved too slow to be practical. The gains in prediction accuracy of the Mamba-based models do not yet translate to significantly improved performance on downstream SNP classification tasks: benchmarks using a GTEx eQTL dataset yield roughly similar results for Mamba- and attention-based classifiers, with attention marginally outperforming Mamba in one metric (a difference of +0.007 in area under ROC) and slightly underperforming by another metric (a difference of -0.006 in Spearman rank correlation). We argue that these results suggest selective state-space models (such as Mamba and Striped Mamba) warrant further exploration for functional genomics tasks. Our code and trained models are publicly available at https://github.com/ihh/bilby.

genomics↗

bioRxiv2023

Rewriting regulatory DNA to dissect and reprogram gene expression

Regulatory DNA sequences within enhancers and promoters bind transcription factors to encode cell type-specific patterns of gene expression. However, the regulatory effects and programmability of such DNA sequences remain difficult to map or predict because we have lacked scalable methods to precisely edit regulatory DNA and quantify the effects in an endogenous genomic context. Here we present an approach to measure the quantitative effects of hundreds of designed DNA sequence variants on gene expression, by combining pooled CRISPR prime editing with RNA fluorescence in situ hybridization and cell sorting (Variant-FlowFISH). We apply this method to mutagenize and rewrite regulatory DNA sequences in an enhancer and the promoter of PPIF in two immune cell lines. Of 672 variant-cell type pairs, we identify 497 that affect PPIF expression. These variants appear to act through a variety of mechanisms including disruption or optimization of existing transcription factor binding sites, as well as creation of de novo sites. Disrupting a single endogenous transcription factor binding site often led to large changes in expression (up to -40% in the enhancer, and -50% in the promoter). The same variant often had different effects across cell types and states, demonstrating a highly tunable regulatory landscape. We use these data to benchmark performance of sequence-based predictive models of gene regulation, and find that certain types of variants are not accurately predicted by existing models. Finally, we computationally design 185 small sequence variants ([≤]10 bp) and optimize them for specific effects on expression in silico. 84% of these rationally designed edits showed the intended direction of effect, and some had dramatic effects on expression (-100% to +202%). Variant-FlowFISH thus provides a powerful tool to map the effects of variants and transcription factor binding sites on gene expression, test and improve computational models of gene regulation, and reprogram regulatory DNA.

genetics↗

bioRxiv2023

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

Sequence-based machine learning models trained on genome-scale biochemical assays improve our ability to interpret genetic variants by providing functional predictions describing their impact on the cis-regulatory code. Here, we introduce a new model, Borzoi, which learns to predict cell- and tissue-specific RNA-seq coverage from DNA sequence. Using statistics derived from Borzois predicted coverage, we isolate and accurately score variant effects across multiple layers of regulation, including transcription, splicing, and polyadenylation. Evaluated on QTLs, Borzoi is competitive with, and often outperforms, state-of-the-art models trained on individual regulatory functions. By applying attribution methods to the derived statistics, we extract cis-regulatory patterns driving RNA expression and post-transcriptional regulation in normal tissues. The wide availability of RNA-seq data across species, conditions, and assays profiling specific aspects of regulation emphasizes the potential of this approach to decipher the mapping from DNA sequence to regulatory function.

genomics↗

bioRxiv2023

Optimizing 5'UTRs for mRNA-delivered gene editing using deep learning

mRNA therapeutics are revolutionizing the pharmaceutical industry, but methods to optimize the primary sequence for increased expression are still lacking. Here, we design 5UTRs for efficient mRNA translation using deep learning. We perform polysome profiling of fully or partially randomized 5UTR libraries in three cell types and find that UTR performance is highly correlated across cell types. We train models on all our datasets and use them to guide the design of high-performing 5UTRs using gradient descent and generative neural networks. We experimentally test designed 5UTRs with mRNA encoding megaTALTM gene editing enzymes for two different gene targets and in two different cell lines. We find that the designed 5UTRs support strong gene editing activity. Editing efficiency is correlated between cell types and gene targets, although the best performing UTR was specific to one cargo and cell type. Our results highlight the potential of model-based sequence design for mRNA therapeutics.

synthetic biology↗

bioRxiv2023

The anti-cancer compound JTE-607 reveals hidden sequence specificity of the mRNA 3' processing machinery

JTE-607 is a small molecule compound with anti-inflammation and anti-cancer activities. Upon entering the cell, it is hydrolyzed to Compound 2, which directly binds to and inhibits CPSF73, the endonuclease for the cleavage step in pre-mRNA 3' processing. Although CPSF73 is universally required for mRNA 3' end formation, we have unexpectedly found that Compound 2- mediated inhibition of pre-mRNA 3' processing is sequence-specific and that the sequences flanking the cleavage site (CS) are a major determinant for drug sensitivity. By using massively parallel in vitro assays, we have measured the Compound 2 sensitivities of over 260,000 sequence variants and identified key sequence features that determine drug sensitivity. A machine learning model trained on these data can predict the impact of JTE-607 on poly(A) site (PAS) selection and transcription termination genome-wide. We propose a biochemical model in which CPSF73 and other mRNA 3' processing factors bind to RNA of the CS region in a sequence-specific manner and the affinity of such interaction determines the Compound 2 sensitivity of a PAS. As the Compound 2-resistant CS sequences, characterized by U/A-rich motifs, are prevalent in PASs from yeast to human, the CS region sequence may have more fundamental functions beyond determining drug resistance. Together, our study not only characterized the mechanism of action of a compound with clinical implications, but also revealed a previously unknown and evolutionarily conserved sequence-specificity of the mRNA 3' processing machinery.

molecular biology↗

bioRxiv2022

Deciphering the Impact of Genetic Variation on Human Polyadenylation

Genetic variants that disrupt polyadenylation can cause or contribute to genetic disorders. Yet, due to the complex cis-regulation of polyadenylation, variant interpretation remains challenging. Here, we introduce a residual neural network model, APARENT2, that can infer 3-cleavage and polyadenylation from DNA sequence more accurately than any previous model. This model generalizes to the case of alternative polyadenylation (APA) for a variable number of polyadenylation signals. We demonstrate APARENT2s performance on several variant datasets, including functional reporter data and human 3 aQTLs from GTEx. We apply neural network interpretation methods to gain insights into disrupted or protective higher-order features of polyadenylation. We fine-tune APARENT2 on human tissue-resolved transcriptomic data to elucidate tissue-specific variant effects. Finally, we perform in-silico saturation mutagenesis of all human polyadenylation signals and compare the predicted effects of >44 million variants against gnomAD. While loss-of-function variants were generally selected against, we also find specific clinical conditions linked to gain-of-function mutations. For example, using APARENT2s predictions we detect an association between gain-of-function mutations in the 3-end and Autism Spectrum Disorder.

genomics↗

bioRxiv2021

Interpreting Neural Networks for Biological Sequences by Learning Stochastic Masks

Sequence-based neural networks can learn to make accurate predictions from large biological datasets, but model interpretation remains challenging. Many existing feature attribution methods are optimized for continuous rather than discrete input patterns and assess individual feature importance in isolation, making them ill-suited for interpreting non-linear interactions in molecular sequences. Building on work in computer vision and natural language processing, we developed an approach based on deep generative modeling - Scrambler networks - wherein the most salient sequence positions are identified with learned input masks. Scramblers learn to generate Position-Specific Scoring Matrices (PSSMs) where unimportant nucleotides or residues are scrambled by raising their entropy. We apply Scramblers to interpret the effects of genetic variants, uncover non-linear interactions between cis-regulatory elements, explain binding specificity for protein-protein interactions, and identify structural determinants of de novo designed proteins. We show that interpretation based on a generative model allows for efficient attribution across large datasets and results in high-quality explanations, often outperforming state-of-the-art methods.

genomics↗

bioRxiv2019

Deep exploration networks for rapid engineering of functional DNA sequences

Engineering gene sequences with defined functional properties is a major goal of synthetic biology. Deep neural network models, together with gradient ascent-style optimization, show promise for sequence generation. The generated sequences can however get stuck in local minima, have low diversity and their fitness depends heavily on initialization. Here, we develop deep exploration networks (DENs), a type of generative model tailor-made for searching a sequence space to minimize the cost of a neural network fitness predictor. By making the network compete with itself to control sequence diversity during training, we obtain generators capable of sampling hundreds of thousands of high-fitness sequences. We demonstrate the power of DENs in the context of engineering RNA isoforms, including polyadenylation and cell type-specific differential splicing. Using DENs, we engineered polyadenylation signals with more than 10-fold higher selection odds than the best gradient ascent-generated patterns and identified splice regulatory elements predicted to result in highly differential splicing between cell lines.

genomics↗

bioRxiv2018

Predicting the Impact of cis-Regulatory Variation on Alternative Polyadenylation

Alternative polyadenylation (APA) is a major driver of transcriptome diversity in human cells. Here, we use deep learning to predict APA from DNA sequence alone. We trained our model (APARENT, APA REgression NeT) on isoform expression data from over three million APA reporters, built by inserting random sequence into twelve distinct 3UTR contexts. Predictions are highly accurate across both synthetic and genomic contexts; when tasked with inferring APA in human 3UTRs, APARENT outperforms models trained exclusively on endogenous data. Visualizing features learned across all network layers reveals that APARENT recognizes sequence motifs known to recruit APA regulators, discovers previously unknown sequence determinants of cleavage site selection, and integrates these features into a comprehensive, interpretable cis-regulatory code. Finally, we use APARENT to quantify the impact of genetic variants on APA. Our approach detects pathogenic variants in a wide range of disease contexts, expanding our understanding of the genetic origins of disease.

genomics↗