13  Variant Effect Prediction

Warning

TODO:

• …
• …

13.1 From Handcrafted Scores to Foundation Models

Variant effect prediction (VEP) sits at the heart of modern genomics. Most variants discovered in clinical sequencing are rare and lack direct experimental evidence; yet clinicians still need to decide whether they’re benign, pathogenic, or somewhere in between. Earlier in this book we saw:

• Conservation and heuristic scores (e.g., traditional tools like SIFT (Ng and Henikoff (2003)), PolyPhen (Adzhubei et al. (2010)), CADD (Rentzsch et al. (2019))), which combine evolutionary constraint and manually engineered features.
• Sequence-to-function CNNs like DeepSEA and ExPecto (Chapters 5–6), which predict chromatin and expression to estimate regulatory effects.
• Specialized architectures like SpliceAI (Chapter 7), which target specific mechanisms such as splicing.
• Protein language models (Chapter 9), which learn rich representations from large-scale protein sequences and can be adapted for missense VEP.

The frontier today is shaped by foundation models that combine:

• Massive pretraining (proteome- or genome-scale),
• Long-range context (from kilobases to megabases),
• Multiple sources of information (sequence, structure, multi-species alignments, multi-omic outputs).

This chapter surveys four landmark systems:

• AlphaMissense – proteome-wide missense pathogenicity predictions.
• GPN-MSA – a DNA language model over multi-species alignments for genome-wide VEP.
• Evo 2 – a generalist genomic language model spanning all domains of life.
• AlphaGenome – a unified, megabase-scale sequence-to-function model with state-of-the-art regulatory VEP.

Together, they preview what “Genomic Foundation Models” look like when specialized for variant interpretation.

---

13.2 AlphaMissense: Proteome-Wide Missense Pathogenicity

AlphaMissense, developed by DeepMind, provides precomputed pathogenicity scores for ~71 million possible human missense variants, covering almost every single–amino acid change in the proteome.

13.2.1 Inputs: Combining Sequence and Structure

AlphaMissense builds on two pillars:

1. Protein language modeling
   • A transformer-based model is trained on massive multiple sequence alignments (MSAs), learning which amino acids tend to appear at each position across evolution.
     
   • From this, the model infers how “surprising” a given amino acid substitution is in its evolutionary context.
2. Predicted 3D structure from AlphaFold2
   • Structural context (packing, secondary structure, local interactions) helps distinguish tolerated changes (e.g., on solvent-exposed loops) from disruptive ones (e.g., in tightly packed cores or active sites).

For each variant, AlphaMissense ingests:

• The wild-type sequence,
• The substitution position and amino-acid change,
• Sequence context from the MSA,
• Structural environment derived from AlphaFold2.

These features are fed into a neural network that outputs a pathogenicity probability between 0 and 1.

13.2.2 Training and Calibration

AlphaMissense’s training is hybrid:

• Self-supervised pretraining learns general sequence and structural representations from evolutionary data.
• Supervised calibration uses:
  • ClinVar and similar databases for labeled pathogenic/benign variants,
  • Population frequencies (e.g., gnomAD) under the assumption that common variants are more likely benign.

The model’s raw scores are calibrated so that:

• Scores near 0 behave like “likely benign,”
  
• Scores near 1 behave like “likely pathogenic,”
  
• Intermediate scores capture uncertainty and ambiguous cases.

In practice, AlphaMissense adopts score cutoffs that approximately map to “likely benign,” “uncertain,” and “likely pathogenic” categories used in clinical interpretation frameworks.

13.2.3 Performance and Clinical Utility

Across diverse benchmarks—ClinVar, curated expert panels, and multiplexed assays of variant effect (MAVEs)—AlphaMissense:

• Achieves state-of-the-art AUROC and AUPRC for missense VEP.
• Generalizes across many genes, including those with little prior annotation.
• Produces scores that are more consistent with experimental functional readouts than many earlier predictors.

As a result, AlphaMissense scores have already been integrated into:

• Clinical re-annotation of exomes,
  
• Reclassification of variants of uncertain significance (VUS),
  
• Gene-specific studies where high-throughput functional assays are impractical.

13.2.4 Limitations and Caveats

Despite its impressive performance, AlphaMissense has important limitations:

• Missense-only: It does not natively handle nonsense, frameshift, regulatory, or deep intronic variants.
• Single-variant focus: It scores one substitution at a time, ignoring combinations of variants (compound heterozygosity, epistasis).
• Dependent on training labels: Any biases in ClinVar or population data (e.g., ancestry representation) can propagate into scores.
• Interpretability: While attention maps and feature attributions can be examined, the reasoning for a particular score is often opaque.

For these reasons, guidelines recommend treating AlphaMissense as supporting evidence to be combined with segregation, functional data, and population frequencies—not as a standalone decision-maker.

---

13.3 GPN-MSA: Genome-Wide Variant Effect Prediction from MSAs

While AlphaMissense focuses on proteins, GPN-MSA tackles the harder problem of genome-wide variant effect prediction in complex genomes such as human, directly at the DNA level.

13.3.1 Alignment-Based DNA Language Model

GPN-MSA extends earlier Genomic Pre-trained Network (GPN) models by operating on multi-species genome alignments:

• Input: a stack of aligned sequences from multiple species (e.g., human plus dozens of mammals).
• Representation: the model sees both:
  • The reference sequence (e.g., human), and
    
  • Auxiliary features encoding how each aligned species matches, mismatches, or gaps at each base.

The model is trained with a masked language modeling (MLM) objective:

• Randomly mask nucleotides in the reference sequence,
• Predict the masked base given the surrounding context and the aligned sequences.

This encourages the model to learn evolutionary constraints: positions where substitutions are strongly disfavored across species get very confident predictions; unconstrained positions allow more flexibility.

13.3.2 Variant Scoring Strategies

GPN-MSA supports several ways to derive variant effect scores:

• Likelihood-based scoring: compare the model’s log-likelihood (or probability) of the reference vs. alternate allele at the variant position.
• Embedding distance: compute embeddings for reference and alternate sequences and use their difference (e.g., Euclidean distance) as an effect magnitude.
• Influence scores: quantify how much a variant perturbs the model’s outputs across the surrounding genomic context.

Because the model operates on whole-genome alignments, it can score:

• Coding and noncoding variants,
• Regulatory elements, introns, UTRs, and intergenic regions,
• Variants in regions with complex conservation patterns, where simple phyloP-like scores Siepel et al. (2005) struggle.

13.3.3 Benchmarking and Applications

GPN-MSA demonstrates strong performance on:

• Genome-wide pathogenic vs. benign classification datasets,
• Variant sets from genome-wide association studies,
• Functional readouts from high-throughput reporter assays.

Practically, GPN-MSA is useful for:

• Genome-wide prefiltering: prioritizing candidate causal variants in regulatory regions.
• Complementing protein-focused tools: supplying information where AlphaMissense is blind (deep noncoding, intronic, intergenic).

Its key limitation is dependency on high-quality multi-species alignments; coverage and quality drop in repetitive, structurally complex, or poorly aligned regions.

---

13.4 Evo 2: A Generalist Genomic Language Model

Evo 2 pushes the foundation-model paradigm to the extreme: it is a genome-scale language model trained across all domains of life—bacteria, archaea, eukaryotes, and phages—on >9 trillion DNA tokens.

13.4.1 Scale and Architecture

Key features of Evo 2 include:

• Autoregressive training on DNA: predict the next base given the preceding context, analogous to next-token prediction in text LLMs.
• A StripedHyena 2 architecture, blending convolutional and attention mechanisms to support:
  • Context windows up to 1 million base pairs,
    
  • Efficient long-range modeling.
• Multiple model sizes (e.g., 7B and 40B parameters) with open-source weights, training code, and the OpenGenome2 dataset.

Evo 2 is designed as a generalist: it is not trained specifically for VEP, but rather to model genomic sequences broadly.

13.4.2 Zero-Shot Variant Effect Scoring

Remarkably, Evo 2 can be used for zero-shot variant interpretation:

• For a given locus, compute the model’s sequence likelihood (or log-probability) for the reference allele.
• Then compute the likelihood for the alternate allele (or sequence containing it).
• The difference in likelihood provides a variant effect score—variants that strongly reduce probability are inferred to be more disruptive.

In benchmarks reported in the preprint and follow-up analyses:

• Evo 2 achieves competitive or state-of-the-art accuracy for pathogenic vs. benign classification across multiple variant types (coding and noncoding), even without variant-specific supervised training.
• A simple supervised classifier built on Evo 2 embeddings reaches state-of-the-art performance on tasks like BRCA1 VUS classification.

13.4.3 Cross-Species Variant Interpretation

Because Evo 2 is trained across diverse species:

• It naturally supports variant effect prediction in non-model organisms (e.g., livestock, crops).
• It can help quantify mutation load, prioritize variants for breeding programs, and guide genome editing designs across species.

However, its generality comes with trade-offs:

• Domain-specific models (like AlphaMissense for human missense or AlphaGenome for regulatory variants) may still outperform Evo 2 on certain human-centric tasks.
• Careful calibration and benchmarking are required before clinical use.

---

13.5 AlphaGenome: Unified Megabase-Scale Regulatory Modeling

Where Evo 2 is generalist and sequence-only, AlphaGenome is explicitly designed as a multimodal regulatory model of the human genome, with a focus on variant effect prediction across many functional readouts.

13.5.1 Architecture: CNNs + Transformers over 1 Mbp

AlphaGenome takes as input 1 megabase (1 Mb) of DNA sequence and produces predictions at single-base resolution for a large set of genomic “tracks,” including:

• Chromatin accessibility and histone marks,
• Transcription factor binding,
• Gene expression (e.g., CAGE-like signals),
• 3D genome contacts,
• Splicing features (junctions and splice-site usage).

Architecturally:

• Convolutional layers detect local sequence motifs.
• Transformer blocks propagate information across the full megabase context.
• Task-specific heads output different experimental modalities across many tissues and cell types.

This design generalizes earlier models like Basenji/Enformer (for regulatory tracks) and SpliceAI (for splicing) into a single, unified model.

13.5.2 Variant Effect Prediction Across Modalities

Given a reference sequence and a candidate variant, AlphaGenome scores variant effects by:

1. Predicting genome-wide functional tracks for the reference sequence.

2. Predicting the same tracks for the sequence bearing the variant.

3. Comparing predictions to obtain delta signals across:

   • Regulatory elements (promoters, enhancers, insulators),
   • Splicing patterns (gain/loss of splice junctions),
   • Gene expression levels,
   • 3D contact maps affecting enhancer–promoter communication.

On extensive benchmarks, AlphaGenome:

• Achieves state-of-the-art accuracy in predicting unseen functional genomics tracks.
• Shows strong performance on diverse variant effect tasks (e.g., noncoding disease variants, splicing disruptions, regulatory MPRA data).
• Provides mechanistic hypotheses (which tracks/tissues are disrupted) rather than only a single scalar risk score.

An API makes AlphaGenome accessible to the research community, enabling large-scale variant scoring without local training infrastructure.

---

13.6 Comparing Design Choices Across Modern VEP Models

The models in this chapter span different points in the design space:

Model	Input Modality	Context Length	Pretraining Data	Variant Types	Primary Outputs
AlphaMissense	Protein sequence + structure	Protein-length	MSAs + structural environment	Missense only	Pathogenicity probability
GPN-MSA	Multi-species DNA alignments	kb-scale windows	Whole-genome MSAs (multiple species)	Coding + noncoding	Likelihood / embedding-based scores
Evo 2	Raw DNA sequence	Up to ~1 Mb	OpenGenome2 (all domains of life)	All variant types	Zero-shot likelihood-based scores
AlphaGenome	Raw DNA sequence	1 Mb	Human genome + multi-omic tracks	All variant types	Multi-omic tracks + delta effects

Key contrasts:

• Scope
  • AlphaMissense is human-missense-specific, with deep clinical calibration.
    
  • GPN-MSA and AlphaGenome are human-genome-centric, spanning coding and regulatory variants.
    
  • Evo 2 is cross-species and general-purpose.
• Context and long-range effects
  • AlphaMissense operates at protein scale.
    
  • GPN-MSA uses modest windows centered on the variant.
    
  • Evo 2 and AlphaGenome support megabase-scale context, capturing long-range regulatory interactions.
• Outputs
  • AlphaMissense and GPN-MSA primarily output scalar scores.
    
  • Evo 2 outputs likelihoods/embeddings that require task-specific postprocessing.
    
  • AlphaGenome outputs rich functional profiles, enabling mechanistic hypotheses.

---

13.7 Practical Use: Choosing and Interpreting Modern VEP Tools

In realistic workflows, these models are complementary rather than competing.

13.7.1 Coding Missense Variants

For human missense variants:

• Use AlphaMissense as a high-coverage, clinically calibrated score.
• Complement with:
  • Protein language model embeddings (Chapter 9) for gene- or domain-specific modeling,
  • Conservation and population data (e.g., GPN-MSA in coding regions, gnomAD frequencies),
  • Gene-level context (constraint metrics, disease association).

13.7.2 Noncoding and Regulatory Variants

For regulatory variation (promoters, enhancers, introns, intergenic):

• Use AlphaGenome to obtain:
  • Tissue-specific changes in chromatin and expression,
  • Splicing consequences (especially for intronic and exonic variants),
  • Potential disruption of long-range enhancer–promoter interactions.
• Use GPN-MSA when:
  • You want a conservation-grounded score,
  • High-quality multi-species alignments are available,
  • You’re scanning broad regions genome-wide.

13.7.3 Cross-Species and Large-Scale Modeling

For non-human organisms, or when building general-purpose genomic tools:

• Leverage Evo 2 for:
  • Zero-shot variant scoring in poorly annotated species,
  • Designing or screening edits (e.g., CRISPR designs),
  • Serving as a feature extractor feeding downstream supervised models.

13.7.4 Score Interpretation and Calibration

Regardless of the model:

• Treat scores as probabilistic evidence, not binary labels.
• Consider:
  • Calibration (does a score of 0.9 truly correspond to ~90% pathogenic variants?),
  • Distribution of scores within a gene (outliers are more suspect),
  • Consistency across tools (agreement between AlphaMissense, GPN-MSA, AlphaGenome, Evo 2, and simpler conservation metrics strengthens confidence).

Where possible, tie predictions back to:

• Mechanistic hypotheses (splice site disruption, enhancer–promoter rewiring),
• Experimental follow-up (targeted assays, MPRA, CRISPR screens).

---

13.8 Open Challenges and Future Directions

Even these state-of-the-art systems leave major gaps:

• Ancestry and population bias
  Training data and labels remain skewed toward certain ancestries, raising concerns about performance and calibration in underrepresented populations.

• Complex variant patterns
  Most models focus on single-base or single-amino-acid changes. Systematic handling of:

  • Haplotypes,
    
  • Indels and structural variants,
    
  • Epistatic interactions across distant loci
    is still in its infancy.

• Integrating multi-omics and longitudinal data
  AlphaGenome marks a step toward unified multi-omic prediction, but dynamic phenomena (developmental trajectories, environment, time-series responses) are only lightly modeled.

• Interpretability and clinical communication
  Translating high-dimensional predictions into explanations that clinicians and patients can understand—and that map onto emerging guidelines for AI-assisted variant interpretation—remains a human-factor challenge.

• Safe deployment and continual learning
  As more functional datasets and clinical labels accumulate, models will need continual updating without catastrophic forgetting, along with governance frameworks to track model versions and provenance.

In the next chapters, we will connect these VEP systems to broader issues in evaluation, bias, and multi-omics integration, positioning them within the broader landscape of Genomic Foundation Models. This chapter’s models illustrate how the building blocks from earlier chapters—NGS, functional genomics, CNNs, transformers, protein and DNA language models—coalesce into powerful, end-to-end systems for variant interpretation.

Adzhubei, Ivan A., Steffen Schmidt, Leonid Peshkin, Vasily E. Ramensky, Anna Gerasimova, Peer Bork, Alexey S. Kondrashov, and Shamil R. Sunyaev. 2010. “A Method and Server for Predicting Damaging Missense Mutations.” Nature Methods 7 (4): 248–49. https://doi.org/10.1038/nmeth0410-248.

Ng, Pauline C., and Steven Henikoff. 2003. “SIFT: Predicting Amino Acid Changes That Affect Protein Function.” Nucleic Acids Research 31 (13): 3812–14. https://doi.org/10.1093/nar/gkg509.

Rentzsch, Philipp, Daniela Witten, Gregory M Cooper, Jay Shendure, and Martin Kircher. 2019. “CADD: Predicting the Deleteriousness of Variants Throughout the Human Genome.” Nucleic Acids Research 47 (D1): D886–94. https://doi.org/10.1093/nar/gky1016.

Siepel, Adam, Gill Bejerano, Jakob S. Pedersen, Angie S. Hinrichs, Minmei Hou, Kate Rosenbloom, Hiram Clawson, et al. 2005. “[PhastCons] Evolutionarily Conserved Elements in Vertebrate, Insect, Worm, and Yeast Genomes.” Genome Research 15 (8): 1034–50. https://doi.org/10.1101/gr.3715005.