Zero-shot Variant Effect Prediction

The core mechanism compares the log-likelihood of the wild-type sequence against the mutated sequence. A mutation that significantly reduces the model's probability of the sequence is flagged as deleterious. This is often computed as the log odds ratio between the mutant and wild-type amino acid probabilities at the mutated position, sometimes incorporating the surrounding sequence context via a sliding window approach.

Unlike traditional variant effect predictors that require training on labeled datasets like Deep Mutational Scans (DMS) or clinical variant databases, zero-shot methods operate purely on the evolutionary information learned during unsupervised pre-training. This eliminates the risk of overfitting to specific protein families or assay conditions and enables generalization to orphan proteins with no prior functional annotations.

A common implementation involves masking the target position in the input sequence and computing the model's predicted probability for the wild-type residue versus the mutant residue. This pseudo-log-likelihood approach, used by models like ESM-1v, isolates the effect of the single substitution from the rest of the sequence context, providing a direct score for the biochemical plausibility of the amino acid change.

The model's score reflects whether a mutation aligns with the evolutionary grammar of proteins. A substitution that is common in homologous sequences (e.g., a leucine to isoleucine in a hydrophobic core) will have a high probability and a low predicted impact. Conversely, a mutation that violates conserved physicochemical constraints (e.g., a glycine to tryptophan in a tight turn) will be scored as highly pathogenic.

Performance is rigorously validated against high-throughput experimental datasets. For example, zero-shot scores from ESM-1v achieve state-of-the-art correlation with DMS data across dozens of proteins, often matching or exceeding supervised methods. Key metrics include Spearman's rank correlation and MCC (Matthews Correlation Coefficient) for binary classification of pathogenic versus benign variants.

Unlike binary classifiers, zero-shot prediction provides a continuous score that can resolve an allelic series—ranking multiple different mutations at the same residue from most to least severe. This granularity is critical for distinguishing between a complete loss-of-function mutation and a mild hypomorphic variant, enabling nuanced genotype-phenotype mapping without training on clinical labels.

A transformer-based protein language model trained on UniRef90 sequences using masked language modeling. ESM-1v predicts variant effects by computing the log-likelihood ratio between the wild-type and mutant amino acid at the position of interest.

• Mechanism: Masks the mutated position and scores how well the model predicts the reference vs. alternative residue given the surrounding sequence context.
• Key Insight: The model learns evolutionary constraints from millions of natural sequences, so a mutation that violates these constraints receives a low likelihood score.
• Benchmark: Achieves state-of-the-art performance on Deep Mutational Scan (DMS) datasets without any experimental training data.

An autoregressive protein language model that scores variants using the conditional probability of each amino acid given all preceding residues in the sequence. Tranception explicitly incorporates Multiple Sequence Alignment (MSA) information through retrieval augmentation.

• Scoring Formula: Computes the log probability ratio P(mutant | context) / P(wild-type | context) across the entire sequence.
• MSA Retrieval: Augments the autoregressive model with evolutionary profiles from homologous sequences, improving accuracy on deep mutational scans.
• Advantage: Captures long-range dependencies and epistatic interactions that masked models may miss due to its left-to-right factorization.

A Variational Autoencoder (VAE) trained on multiple sequence alignments of protein families. EVE learns a continuous latent space that captures the distribution of functional sequences and scores variants based on their deviation from this learned distribution.

• Scoring Mechanism: Computes the Evolutionary Index — a measure of how likely a variant sequence is under the VAE's generative model compared to the wild-type.
• Unsupervised: Requires only an MSA of homologous sequences, not experimental labels.
• Clinical Relevance: EVE scores correlate strongly with human pathogenic annotations from ClinVar, enabling variant interpretation for rare diseases.

A suite of large autoregressive protein language models trained on over one billion sequences with conditioning tags for taxonomic and functional annotations. ProGen2 enables zero-shot variant scoring by computing the probability of a sequence under a specific functional conditioning.

• Conditional Scoring: Evaluates P(sequence | function_tag), allowing assessment of whether a mutation preserves or disrupts a specific functional property.
• Scale Advantage: The massive training corpus provides broad coverage of protein sequence space, improving generalization to novel proteins.
• Applications: Used for zero-shot prediction of mutation effects on thermostability, catalytic activity, and solubility.

A message-passing neural network for inverse protein folding that predicts amino acid probabilities given a 3D protein backbone structure. ProteinMPNN scores variants by evaluating how well a mutation is predicted by the structural context.

• Structure-Based Scoring: Unlike sequence-only models, ProteinMPNN uses spatial information — residue distances, angles, and local geometry — to assess variant plausibility.
• Zero-shot Mechanism: Computes the log probability of the mutant amino acid at a position conditioned on the fixed backbone structure of the wild-type protein.
• Design Integration: Enables simultaneous scoring and redesign, suggesting compensatory mutations that restore structural compatibility.

Combining predictions from multiple zero-shot models often yields more robust variant effect estimates than any single model. Ensemble methods leverage the complementary strengths of different architectures.

• Meta-Predictors: Weighted averages or logistic regression over scores from ESM-1v, EVE, and Tranception improve correlation with experimental DMS data.
• Complementary Signals: Masked models capture local context, autoregressive models capture directional dependencies, and VAE-based models capture family-level constraints.
• Practical Implementation: Tools like VESPA and VARITY integrate multiple zero-shot scores to provide consensus pathogenicity predictions for clinical variant interpretation.

The primary metric for zero-shot variant effect prediction. Perplexity quantifies how 'surprised' a protein language model is by a mutated sequence compared to the wild-type.

• A higher perplexity for the mutant indicates the substitution disrupts the learned grammar of functional proteins
• The delta perplexity (mutant score minus wild-type score) serves as a proxy for functional impact
• This approach requires no labeled training data, making it immediately applicable to any protein

A high-throughput experimental method that assays the functional effect of thousands of single amino acid substitutions across a protein. These datasets serve as the gold-standard benchmark for evaluating zero-shot predictors.

• Generates fitness scores for nearly all possible single mutants
• Provides the ground truth for correlating computational predictions with experimental reality
• Key datasets include β-lactamase, GFP, and TP53

A conceptual mapping of all possible protein sequences to their associated biological fitness or functional activity. Zero-shot predictors attempt to learn this landscape implicitly from evolutionary data.

• Smooth landscapes suggest gradual fitness changes between similar sequences
• Rugged landscapes contain sharp cliffs where single mutations drastically alter function
• Language models capture the local geometry of this landscape through their learned probability distributions

The self-supervised pre-training objective that enables zero-shot variant scoring. During training, random amino acids are masked and the model learns to predict them from surrounding context.

• Analogous to BERT training in natural language processing
• The model develops rich internal representations of residue dependencies
• At inference time, the log-likelihood ratio between wild-type and mutant residues at a position provides the variant effect score

A computational alignment of evolutionarily related protein sequences used to identify conserved regions and inform structural predictions. While modern protein language models reduce MSA dependency, evolutionary information remains critical.

• Deep MSAs reveal co-evolving residue pairs
• Traditional variant effect predictors rely heavily on position-specific conservation scores
• ESM-2 and similar models implicitly learn MSA-like information during pre-training on millions of sequences

The process of navigating a protein language model's latent space to generate novel sequences with altered properties. This extends zero-shot prediction into the generative domain.

• Latent space interpolation between functional sequences can produce chimeric proteins
• Perturbation vectors can steer generation toward desired properties like thermostability
• Demonstrates that zero-shot scoring and generative design are two sides of the same learned representation