Variant Effect Prediction

The core mechanism involves computing delta scores—the quantitative difference in a model's prediction between the reference and alternate allele. This is performed by presenting both sequences to a genomic neural network and subtracting the reference logit from the alternate logit. A large absolute delta score indicates a high-impact variant. This approach is foundational to tools like DeepSEA and Enformer, which predict chromatin effects and expression changes directly from DNA sequence.

A systematic perturbation technique where every nucleotide in a given sequence is computationally mutated to all three alternate bases. The model's prediction change is recorded for each substitution, generating a mutational landscape map. This exhaustive approach reveals which base pairs are functionally critical and which are tolerant to change, providing a complete view of sequence constraint without performing costly laboratory experiments.

Model predictions are validated against experimental Deep Mutational Scans (DMS) and massively parallel reporter assays (MPRAs). These high-throughput assays measure the functional impact of thousands of variants simultaneously. Key performance metrics include:

• Spearman correlation between predicted delta scores and measured activity
• Area under the ROC curve (AUROC) for classifying pathogenic vs. benign variants
• Precision-recall curves for imbalanced clinical datasets like ClinVar

Beyond protein-coding changes, models predict the impact of non-coding variants on transcription factor binding, chromatin accessibility, and histone modification profiles. Multi-task architectures like Sei and Enformer output thousands of epigenetic tracks simultaneously. A variant that disrupts a conserved GATA2 binding motif in an enhancer region, for example, would receive a high deleteriousness score due to predicted loss of regulatory activity.

The prediction alone is insufficient; the model must explain why a variant is deleterious. Integrated Gradients and DeepLIFT generate nucleotide-resolution attribution maps that highlight the specific base pairs driving the prediction. If a missense variant disrupts a splice donor site, the attribution map will show high importance concentrated at the exon-intron boundary, providing mechanistic evidence that can be validated by RNA sequencing data.

Advanced models capture non-additive interactions between variants, known as epistasis. A mutation that is benign in isolation may become deleterious when combined with a second variant elsewhere in the sequence. Transformer-based architectures with self-attention mechanisms learn these long-range dependencies, enabling the prediction of compound heterozygous effects and genetic background interactions that simpler additive models miss.

A systematic perturbation technique that computationally mutates every nucleotide in a sequence to quantify its impact on model predictions. ISM is the most direct method for variant effect prediction, creating a comprehensive map of position-specific sensitivity.

• Computes prediction change for all 3 alternate bases at every position
• Produces a mutation sensitivity map directly interpretable as functional constraint
• Computationally expensive but serves as the gold-standard ground truth for faster methods

An axiomatic feature attribution method that computes the path integral of gradients from a baseline input to the actual input, satisfying the completeness axiom. For variant effect prediction, it quantifies exactly how much each nucleotide contributes to the difference between reference and variant predictions.

• Guarantees that attributions sum to the prediction difference
• Requires careful selection of a biological baseline (e.g., all-zero embedding)
• Satisfies sensitivity and implementation invariance axioms

A backpropagation-based attribution algorithm that compares neuron activations to a reference state using rescale and revealcancel rules. DeepLIFT is widely adopted in genomics because it efficiently explains variant effects without requiring gradient integration.

• Computes contribution scores in a single backward pass
• Handles saturation effects that zero gradients miss in deep networks
• Reference sequence choice critically impacts biological interpretability

A unified framework based on Shapley values from cooperative game theory that assigns each genomic feature an importance score for a particular prediction. SHAP values provide a theoretically grounded, additive feature attribution that connects directly to variant effect magnitude.

• KernelSHAP: Model-agnostic but computationally heavy
• DeepSHAP: Combines DeepLIFT rules with Shapley calculations for speed
• Guarantees local accuracy and consistency properties

The quantitative difference in a model's prediction score between a reference and an alternate allele, used to assess the functional impact of genomic variants. This is the most direct operationalization of variant effect prediction, mapping a genetic change to a scalar pathogenicity signal.

• Computed as score(alt) - score(ref) for any prediction task
• Can be derived from any model output: binding affinity, expression, splicing
• Forms the basis for clinical variant classification pipelines

Quantitative measures that evaluate how accurately an attribution map reflects the true decision-making logic of a genomic model through perturbation experiments. These metrics are essential for validating that variant effect predictions are causally grounded, not correlational artifacts.

• ROAR: Retrain after removing top features to measure degradation
• AOPC: Area over the perturbation curve as salient bases are masked
• Infidelity Measure: Expected error between input perturbation and attribution perturbation