Leveraging genomic deep learning models for the prediction of non-coding variant effects
summary
The gist
Characterizing non-coding variant function remains an important challenge in human genetics with important applications to rare disease, cancer…
In short
Genomic deep learning models are being used to predict how non-coding genetic variants affect biological functions like gene expression or chromatin state. The review examines sequence-based models, including supervised and self-supervised approaches, and discusses strategies for comparing variant effects using various data sources and evaluation tasks.
Key concepts
- Supervised Sequence-to-Activity Models
- These models are trained to directly predict functional activity levels or measurements from DNA sequences. They learn to map specific DNA sequences to experimental outcomes like chromatin accessibility, allowing them to predict the effect of a variant on these molecular phenotypes.
- Self-Supervised Genomic Language Models
- These models use tasks similar to natural language processing, such as predicting masked parts of a DNA sequence from its context. They learn rich representations of DNA sequences without needing labeled functional data initially, which can then be fine-tuned on specific variant effect measurements.
- Variant Effect Prediction Strategies
- When using supervised models, predictions for the reference and alternate alleles are compared using metrics like log fold-change or absolute differences. For self-supervised models, prediction involves comparing internal sequence embeddings or predicting the probability of each allele at a masked location based on surrounding context.
- Validation Data Sources
- Ground truth data comes from two main areas: observational studies (like GWAS) correlating variants with phenotypes, and experimental assays (like CRISPR or MPRAs) that test functional activity in lab settings. Both are used to validate the predictions made by the deep learning models.
Terminology used across episodes
This episode discusses
- Leveraging genomic deep learning models for the prediction of non-coding variant effects · Paper Radio
- Caduceus: Bi-Directional Equivariant Long-Range DNA Sequence Modeling
The paper
Leveraging genomic deep learning models for the prediction of non-coding variant effects · Read on arXiv
Pooja Kathail, Ayesha Bajwa, Nilah M. Ioannidis
Center for Computational Biology, University of California Berkeley · Department of Electrical Engineering and Computer Sciences, University of California Berkeley · Chan Zuckerberg Biohub
Characterizing non-coding variant function remains an important challenge in human genetics. Genomic deep learning models have emerged as a promising approach to enable in silico prediction of variant effects. These include supervised sequence-to-activity models, which predict molecular phenotypes such as genome-wide chromatin states or gene expression levels directly from DNA sequence, and self-supervised genomic language models. Here, we review progress in leveraging these models for non-coding variant effect prediction. We describe practical considerations for making such predictions and categorize the types of ground truth data used to evaluate variant effect predictions, providing insight into the settings in which current models are most useful. Our Review highlights key considerations for practitioners and opportunities for improvement in model development and evaluation.
DOI: 10.1038/s41588-026-02769-7
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Leveraging genomic deep learning models for the prediction of non-coding variant effects".
Marcus: Here's the gist of the paper. Characterizing non-coding variant function remains an important challenge in human genetics with important applications to rare disease, cancer, and complex traits Genomic deep learning models have emerged as a promising approach to enable in silico prediction of variant effects Here, we review progress in leveraging these models for non-coding variant effect prediction Our Review highlights key considerations for practitioners and opportunities for improvement in model development and evaluation Modeling Paradigms Sequence-based deep learning approaches, adapted from computer vision and natural language processing, have shown success in modeling genomic sequences and functional genomics data Two modeling paradigms, supervised and self-supervised, are commonly used (Fig.
Ines: 1a,b) Supervised sequence-to-activity models (also called sequence-to-function models) learn to predict functional activity levels, or the measurements obtained from genome-wide experimental assays of molecular phenotypes such as chromatin accessibility or histone modifications directly from DNA sequence In contrast, self-supervised genomic language models (Fig.
Marcus: 1b, Box 2) use tasks adapted from natural language processing, such as learning to predict a masked portion of the input DNA sequence from just the surrounding sequence context to learn representations of DNA sequences Both types of models can subsequently be trained (see Box 3 for probing and fine-tuning approaches) on labeled data from supervised tasks of interest, such as additional genome-wide experimental assays or variant effect measurements Variant Effect Prediction Strategies To make a variant effect prediction using a supervised model, two separate functional activity predictions—one for the sequence containing the reference allele and one for the alternate allele—are generated and compared (Fig.
Ines: 1c) Different distance metrics can be used for this comparison, including absolute (e.g.
Title and authors: Ines: So we're looking at this paper now, "Leveraging genomic deep learning models for the prediction of non-coding variant effects". It’s basically about how these deep learning tools can help us guess what a change in DNA sequence actually does when it's not in a protein-coding region.
Marcus: Right. It tackles the challenge of characterizing non-coding variant function, which is tough because you don't have easy experimental data for all those sequences, especially the non-coding ones. The authors point out that genomic deep learning models are starting to look like a promising way to do in silico prediction of variant effects.
Yuki: From a population genetics standpoint, I'm curious what these models are actually capturing about the biology when they make these predictions across all those sequences. What biological signals does the model think it's seeing?
Ines: The paper introduces two main modeling paradigms that they review: sequence-based deep learning approaches adapted from computer vision and natural language processing, which we see used in functional genomics data. They focus on supervised sequence-to-activity models and self-supervised genomic language models.
Marcus: The supervised ones are the ones that learn to predict functional activity levels, like chromatin accessibility or histone modifications directly from the DNA sequence itself. That’s a direct measurement approach for a given DNA sequence, right?
Yuki: And the self-supervised models, they use tasks borrowed from natural language processing, like learning to predict a masked portion of the input DNA sequence based on its surrounding context. That suggests they're trying to build representations of the DNA grammar itself.
Ines: Exactly. The point is that both types can then be trained further using labeled data from other functional assays or variant effect measurements, which is where the real prediction power comes from, as described in the paper’s abstract Characterizing non-coding variant function remains an important challenge in human genetics.
Marcus: And how they actually make a specific variant effect prediction with those models involves generating two separate activity predictions—one for the reference allele and one for the alternate allele—and then comparing them using metrics like absolute or relative differences, such as log fold-change.
Yuki: That comparison step is crucial because it’s how you get a directional prediction, whether the variant increases or decreases something, which is a key question for any genetic study.
Ines: And many of these supervised models use multi-task learning to make simultaneous predictions for hundreds or thousands of cell types and experimental assays. When you have multiple tasks, they aggregate those individual effect predictions into a single score per variant using methods like averaging or taking the maximum prediction.
Title and authors: Marcus: That aggregation step is where they try to get a more robust summary of the variant's impact across different biological contexts, which helps manage the noise from having so many different experimental assays involved.
Yuki: But I wonder about those limitations. If you aggregate predictions over hundreds of cell types, are you losing the specific signal for a particular disease-relevant cell type? What about rare variants that only occur in one small population?
Ines: That’s a valid concern, and it leads into the paper’s discussion on validation data sources. They discuss using two broad categories of ground truth data: observational data from naturally occurring human genetic variation, and experimental data from introduced variation in laboratory contexts.
Marcus: The observational side involves things like GWAS or quantitative trait locus studies, which correlate variant dosages with organism-level phenotypes or molecular traits like gene expression levels, drawing from large human cohorts. But the paper notes a limitation there: these studies identify both causal and non-causal variant associations because of correlations between nearby variants or linkage disequilibrium.
Yuki: So, even when you use big cohort data to find an association, you can’t always be sure which specific non-coding variant is driving the effect because of that linkage disequilibrium issue.
Ines: Exactly. And they also mention that association studies have power to detect effects only for relatively common variants and are biased by the genetic ancestry of those cohorts, which is another constraint. Related types of observational data include rare variants associated with extreme or outlier phenotypes, plus databases like ClinVar or the Human Gene Mutation Database that give categorical labels like pathogenic or benign.
Marcus: Then you have the experimental side, using things like massively parallel reporter assays and CRISPR-based perturbation experiments to test functional activity levels of synthetic sequences or make direct perturbations to native DNA. That allows them to collect functional data from sequences that aren't even present in observed populations.
Yuki: So, when you combine the observational strength of large cohorts with the experimental capability of targeted perturbation experiments, what kind of picture are we building about these non-coding effects?
Ines: The paper moves into constructing evaluation tasks. They suggest three main types: functional variant identification, which is a binary classification task to see if models can distinguish a positive set of variants from negative ones; effect direction prediction, which is another binary task to predict whether a variant increases or decreases a phenotype; and effect size prediction, which is a regression task to quantitatively predict the magnitude of the effect.
Title and authors: Marcus: The paper also mentions that when using genome-wide summary statistics from GWAS or QTL studies to evaluate predicted effect sizes, you can use approaches like stratified linkage disequilibrium score regression or signed LD profile regression to explicitly account for that LD.
Yuki: That’s interesting because it suggests we might need statistical methods specifically designed to handle the complexity of the human genome structure when interpreting these predictions.
Ines: And they point out that scale and phenotype specificity matter a lot, saying that successful prediction at a given scale may require specific modeling choices, like understanding proximal sequence modulation for chromatin states. They stress that predicting effects on epigenetic or chromatin state needs a causal understanding of relevant sequence features, like transcription factor binding motif disruption.
Marcus: Then they highlight the utility of supervised models for predicting variant effects on local chromatin state—including identification and direction prediction—using datasets involving chromatin accessibility, methylation, histone modification, and transcription factor binding variant effect data.
Yuki: So, what is the consistent finding here about model accuracy across these tasks? The paper claims that predicted variant effects are generally more accurate when the magnitude of the predicted difference between reference and alternate alleles is larger.
Ines: They back that up with examples like DeepSEA for chromatin accessibility and histone modification variants, CpGenie for methylation QTLs, and DeepFIGV for chromatin accessibility QTLs. That suggests a strong correlation between the size of the predicted effect and model accuracy in those specific contexts.
Marcus: Moving on to transcriptional regulation predictions, they note that many genomic deep learning models perform better than non-deep-learning baselines when predicting effects on gene expression, particularly for variants near genes.
Yuki: But there's still room for improvement in predicting distal variant effects and quantitative effect sizes, and also in making those predictions personalized to specific gene expression profiles.
Ines: They specifically test supervised models trained to predict RNA splicing patterns from DNA sequence on their ability to predict variant effects on splicing itself. For five’ UTR variants, a model called Optimus five-Prime was tested; it accurately predicted MPRA-measured effects of five’ UTR SNVs from ClinVar.
Marcus: That's a concrete example of where these sequence models are showing tangible success, linking the DNA sequence directly to an observed molecular phenotype measured in an assay.
Title and authors: Yuki: So, what does this mean for the future? If we focus on incorporating biologically motivated inductive biases into those supervised models—making them interpretable-by-design—could that help capture higher order cis-regulatory rules that current data might miss?
Ines: That’s a key improvement suggestion. It means building in known biological constraints so the models don't just learn patterns blindly, but learn rules that are already biologically plausible, which could increase accuracy on local chromatin state predictions.
Marcus: For the self-supervised language models, they suggest using multi-species alignment to boost unlabeled training data. This could help them learn more robust sequence conservation patterns relevant to regulatory elements across different species, which is useful for finding rare human variants.
Yuki: I agree, having more diverse unlabeled data from evolutionary history should give those models a better foundation for understanding the broader function of these regulatory regions.
Ines: And finally, they suggest adopting formal variant effect benchmarks and competitions to standardize data formats and model comparisons. That standardization would allow for a systematic comparison of modeling and variant scoring decisions, which addresses a gap in reproducibility in this field.
Marcus: So to wrap up on "Leveraging genomic deep learning models for the prediction of non-coding variant effects", the paper shows how sequence deep learning models are being used to predict various molecular phenotypes from DNA sequences, covering everything from chromatin states to gene expression.
Yuki: It seems like the main implication is that we have a new toolkit—these sequence models—that can generate predictions for non-coding variants, but it still depends heavily on the quality of the training data and how we design our evaluation tasks.
Ines: Exactly. The success seems tied to the magnitude of the predicted difference between reference and alternate alleles, as seen in examples like DeepSEA and CpGenie. We see promise for predicting effects on local chromatin state when we can feed these models biologically informed constraints.
Marcus: And for practical implementation, they are pushing for standardization through benchmarks so that when we compare different model architectures or different scoring methods, the comparison is fair and systematic.
Yuki: For those of us looking at the wider context, it means we’re getting better tools to start prioritizing likely causal variants in traits by using these sequence predictions as a filter before we even look at complex population statistics.
Ines: That's what it boils down to for now. We have these powerful predictors, but the next step is integrating them into a workflow that uses those biological constraints and standardizes how we score the results to get reliable insights.
The paper's summary: Ines: So we're looking at how deep learning models are trying to figure out what happens when you change DNA sequences that aren't in protein-coding regions.
Marcus: Right, it boils down to using these sequence models—the deep learning stuff—to make predictions about the functional effect of non-coding variants.
Yuki: The core idea is taking that massive amount of sequence data and training AI to predict how that DNA change impacts things like chromatin accessibility or histone modifications.
Ines: They look at two main ways to build these models: supervised models, which learn directly from experimental data on what those activities actually are, and self-supervised language models, which use tasks borrowed from natural language processing on the DNA sequence itself.
Marcus: The supervised ones try to predict activity levels right away, using the DNA sequence as input for a functional assay. But they also have these multi-task learning approaches where one model tries to predict hundreds of different things at once.
Yuki: And for those self-supervised models, they're basically teaching the AI how sequences are structured by learning tasks like masking parts of the DNA and guessing what’s missing based on the surrounding context.
Ines: Then, to actually get a variant effect prediction, they take those two types of models and compare them directly. They generate an activity prediction for the reference allele and another one for the alternate allele.
Marcus: And they compare those two predictions using different metrics, like absolute differences or log fold-change between the reference and alternate predictions.
Yuki: That comparison is how they figure out if a variant will increase or decrease some biological activity, which is what we really want to know about its function in the cell.
Ines: Many of these models then aggregate those different task predictions into one final score for the variant, using things like averaging or taking the maximum prediction across all those different experiments.
Marcus: That aggregation step is important because it tries to give you a single sense of how much that variant impacts multiple biological processes simultaneously.
Yuki: But they also look at how we get this truth—the validation data. They use observational data from big human studies, like GWAS, and experimental data from things like CRISPR experiments to check their predictions against real-world results.
Ines: They test three main tasks: identifying the functional variants, predicting the direction of the effect—whether it's an increase or decrease—and predicting the actual size of that effect.
Marcus: And they’re using techniques like stratified linkage disequilibrium score regression to try and account for those tricky correlations between nearby variants when looking at those GWAS statistics.
Yuki: The paper stresses that the accuracy of these predictions really depends on the scale you're looking at; for instance, predicted effects are usually more accurate when the difference between the reference and alternate alleles is larger.
Ines: For chromatin states specifically, they find that supervised models work better when they also incorporate knowledge about sequence features, like where transcription factor binding motifs are disrupted.
Marcus: So, looking at what these models can actually do—predicting effects on gene expression or splicing—they're doing better than simpler deep learning baselines, but there’s still work left on predicting very distant effects.
Yuki: It suggests that these sequence-based tools are a powerful new lens for understanding non-coding regions, but they need to be paired with careful data validation and biologically informed constraints to really nail the causal signal.
The paper's improvements: Marcus: So now we’re looking at how the authors suggest they can make these deep learning models better for predicting those non-coding variant effects.
Ines: They focus on two main ways to improve the models: adding biological rules and increasing the training data available.
Marcus: For those supervised sequence-to-activity models, they suggest incorporating biologically motivated inductive biases, meaning building in known biological constraints so the model learns rules that are already plausible.
Yuki: That makes sense because current data might just be showing patterns without the underlying biology there to guide the learning process.
Ines: Exactly, they want interpretable-by-design models that can better capture those higher order cis-regulatory rules that we haven't fully mapped yet.
Marcus: And for those self-supervised language models, they propose using multi-species alignment to get more unlabeled training data.
Yuki: That could be huge because it means the AI would learn more robust sequence conservation patterns relevant to regulatory elements across different species, which is useful when we’re looking at rare human variants.
Ines: It helps the model understand evolution better, which should improve its performance on specific human tasks.
Marcus: Then they talk about standardizing everything with formal variant effect benchmarks and competitions.
Yuki: That standardization would help everyone compare models fairly, which addresses the issue of reproducibility that’s a big problem right now.
Ines: It lets us systematically compare how different modeling choices or different variant scoring decisions actually perform on the same task.
Marcus: So, the implication here is that we need to move beyond just feeding these AI massive amounts of sequence data and start designing models with more biological intelligence built in.
Yuki: And for those of us looking at population genetics, it means we can get more reliable predictions for traits even when the sample size is smaller because the model will be learning from a broader evolutionary context.
Conclusion: Ines: So to wrap things up, this paper on "Leveraging genomic deep learning models for the prediction of non-coding variant effects" shows how sequence deep learning is moving beyond just reading DNA to actually predicting functional consequences.
Marcus: It summarizes the main modeling paradigms and the different strategies they use to compare reference versus alternate allele predictions.
Yuki: And it lays out what’s needed from validation—the observational data versus the experimental assays—to make those in silico predictions meaningful.
Ines: The big implication is that we have these tools now that can generate a prediction for variant effects across many different molecular phenotypes, like chromatin states or gene expression.
Marcus: But they also point out that accuracy really hinges on the magnitude of the predicted difference between the two alleles, and how well you handle those linkage disequilibrium scores in your statistics.
Yuki: From a population standpoint, this suggests we can start using these sequence models as a filter to prioritize variants that actually have a biological impact before we look at complex traits in large cohorts.
Ines: It’s about using these AI tools not just to find associations, but to predict the actual physical effect on the DNA structure itself.
Marcus: The authors are pushing for improvements by adding biological rules and standardizing the way we evaluate these models through formal benchmarks.
Yuki: I think having that kind of standardization is crucial so that when we look at these predictions across different studies, we actually know what’s reliable.
Ines: It really shows the path forward: sequence-based deep learning is a powerful new way to probe non-coding function, provided we keep building in biological context and rigorous testing.
Marcus: We're moving from just finding correlations to predicting causal directions with these models, which is a big step for genomics data science.
Yuki: It’s promising stuff that moves us closer to understanding the regulatory architecture of the human genome, even if there’s still work to do on those distal effects.
More episodes
- 2610.33881-Sensorimotor behavior from an untrained brain-and-nerve-cord connectome
- 2610.40344-Morphological Features of Murine Pulmonary Arterial Networks for Diagnosing Vascular Disease
- Daily Summary for 2026-11-19
- 2609.10193-Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets
- 2610.38236-Generative discovery of chemical reaction networks
- 2507.11176-An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2504.03590-Towards a Unified Framework for Determining Conformational Ensembles of Disordered Proteins
- 2603.15217-A multiscale discrete-to-continuum framework for structured population models