Interpretable variant effect prediction from genomic foundation model embeddings

Decoding the Genomic Black Box

Scientists have found a way to peek inside a massive AI model trained on DNA. This allows them to understand why certain genetic mutations cause disease. Instead of just getting a "yes" or "no" on whether a mutation is harmful, this new method explains the specific biological reason. It can identify a broken protein part or a splicing error (a mistake in how DNA is read to make proteins).

Predicting the effect of a genetic variant—a change in a single DNA letter—is a fundamental challenge. When we sequence a patient's genome, we often find thousands of variations. Most are classified as "Variants of Uncertain Significance" (VUS). This means we do not know if they cause disease or are harmless. Current specialized models can predict if a variant is pathogenic (disease-causing) or benign (harmless). However, they are often siloed. Protein-focused models struggle with regulatory DNA (the parts of DNA that control gene activity). Meanwhile, sequence-to-function models rarely explain why a change matters. This leaves clinicians with a score but no mechanism to support a diagnosis.

The fragmentation of variant prediction

Existing approaches suffer from two primary bottlenecks: biological fragmentation and a lack of interpretability. Most current tools are specialized for specific "regimes." For instance, protein-based models focus almost exclusively on missense substitutions (changes that swap one amino acid for another). Regulatory models focus on how DNA affects gene expression. Genome-wide meta-predictors attempt to bridge this. They compress heterogeneous evidence into a single scalar score (a single number). But this process loses the nuance of the underlying biology.

Furthermore, most methods provide a prediction without a mechanism. For clinical use, a score alone is rarely sufficient. Medical guidelines require categorized evidence to justify a classification. Relying solely on output likelihoods is risky. Likelihood is the change in probability a model assigns to a sequence when a variant is introduced. This approach collapses the rich, multi-billion-parameter internal state of a neural network into a single, often unreliable number.

Probing the second-order structure of Evo 2

The authors propose a solution through "structured probing" of Evo 2. This is a 7-billion-parameter genomic foundation model. Rather than treating the model as a black box, they interrogate its internal representations. These are the high-dimensional numerical vectors that encode the model's understanding of a sequence.

The researchers implement a three-stage pipeline:

1. The Covariance Probe: To predict pathogenicity, the authors pass both the reference and mutated sequences through Evo 2. They do not simply average the differences in their embeddings. Averaging is called mean-pooling. Instead, they compute a learned low-rank covariance. This captures "second-order structure," or how different features co-occur across positions. To simplify, mean-pooling is like recording the average temperature of a room. Covariance is like noting how temperature and humidity move together. The latter provides a much deeper sense of the environment's state.
2. Annotation Probes: To extract meaning, the authors train 236 "annotation probes." These probes map the model's latent space (its internal mathematical landscape) onto known biological properties. These include protein domains, splice modules, and histone modifications. By measuring the change ($\Delta$) in these properties, they create a "disruption profile."
3. LLM Synthesis: Finally, these disruption profiles are fed into a large language model (specifically Claude 3.5 Sonnet). The LLM acts as a translator. It synthesizes the mathematical shifts into a natural-language mechanistic hypothesis. For example, it might state that a variant "destroys a splice-acceptor site."

From clinical labels to experimental reality

The authors report that this covariance-based approach achieves high accuracy across many variant types. When evaluated on 833,970 single-nucleotide variants from ClinVar, the probe achieved an aggregate AUROC of 0.997 [Figure 1b]. AUROC is a metric where 1.0 is perfect discrimination. This score means the model is extremely effective at separating pathogenic from benign variants. Notably, the model showed strong zero-shot generalization to indels (insertions or deletions). It outperformed the classical CADD predictor in this category [Figure 1d].

The study validates these predictions against physical biological reality. The covariance probe maintains high performance across the entire spectrum of evolutionary conservation [Figure 2a]. This suggests it captures functional information that simple conservation-based models miss. In direct experimental transfers to Deep Mutational Scan (DMS) datasets, the probe outperformed loss-based scoring. For the LDLR gene, it achieved a Spearman correlation of 0.41, compared to 0.34 for AlphaMissense [Figure 2d]. This correlation measures how well model predictions track actual functional changes.

The mechanistic hypotheses generated by the LLM were also verifiable. For a specific BRCA1 variant, the system predicted the activation of a "cryptic" (alternative and unintended) splice site. The authors validated this using RNA sequencing in ovarian cancer cell lines. They found that the mutant reads matched the model's predicted nucleotide-level resolution exactly [Figure 4b].

Limits of the interpretive interface

The authors identify several important caveats. First, the pathogenicity probe was trained using ClinVar labels. This creates a "home-domain advantage." The model might be exceptionally good at predicting things it saw during training. The authors argue that the DMS and biobank analyses provide stronger evidence of true generalization.

Second, the interpretability depends on the quality of the annotation probes. These probes are limited to currently known, labeled genomic features. If a variant disrupts a mechanism not yet in a database, the profile may fail to capture it.

Finally, the LLM-generated interpretations are "testable mechanistic hypotheses." They are not definitive clinical truths. The authors warn that the LLM might occasionally rely on memorized knowledge of famous variants. These outputs require expert review before clinical use.

A new framework for scientific AI

The verdict on this work is a qualified "yes" for researchers and a "not yet" for autonomous clinical deployment. The authors have shown that the internal representations of a foundation model are organized around biological rules. By using covariance-based probing, they recovered predictive power lost in simpler methods.

For the technical community, the most significant contribution is the "evidence interface." The disruption profile bridges the gap between incomprehensible high-dimensional activations and the reasoning of a language model. This provides a blueprint for interrogating other scientific foundation models. It turns opaque "black box" predictions into actionable, human-readable knowledge. The authors have released these tools through the Evo Variant Effect Explorer (EVEE). This makes 4.2 million interpreted variants available for research.

Figures from the paper