What is AlphaMissense?

How Is the AlphaMissense Score Calculated Differently from Older Methods?

AlphaMissense stands out from earlier variant predictors like REVEL and DANN because of how it learns to identify harmful genetic changes.

Key Difference: What the Models Learn From

AlphaMissense uses information about protein structure and evolutionary patterns. It learns by looking at:

 - How a protein’s 3D shape might change if a single amino acid is altered

 - How similar changes appear (or don’t appear) across different species.

 Importantly, AlphaMissense does not train directly on human clinical databases or labels. This means it avoids biases that can come from existing human data (reference).

REVEL and DANN – older tools rely heavily on clinical and population databases. They learn from previously labeled “pathogenic” or “benign” variants in resources like ClinVar and gnomAD, from combinations of existing prediction tools. This means their predictions can reflect the biases and gaps in current human knowledge.

Why Does This Matter?

AlphaMissense can better predict the effects of rare or previously unseen variants because it isn’t limited by what’s already in clinical databases. It also uses the actual 3D structure of proteins, giving it more biological context for each variant.

REVEL and DANN are strong for common variants but may struggle with rare or novel changes.

AlphaMissense predicts whether a missense variant is likely harmful by analyzing protein structure and evolution, without relying on existing clinical labels. This makes it a powerful, less-biased tool for interpreting genetic changes—especially those that haven’t been seen before.

Practical Use of the score

Based on the AlphaMissense study, the model classifies missense variants into three categories using thresholds applied to the AlphaMissense pathogenicity score.

The thresholds are set to achieve  90% precision for both benign and pathogenic classifications, as estimated from ClinVar data.

The classification ranges are:

These thresholds were chosen to ensure that variants classified as "likely pathogenic" or "likely benign" have a 90% probability of correct classification based on ClinVar benchmarks.

The "ambiguous" category includes variants where confidence is insufficient for a definitive call.

In Gene Inspector Pro, this score is displayed as "AMS: 0.99933" in the column where other predictors are shown:

There is also a dedicated panel called "High AlphaMissense Score" which contains variants that have score >= 0.5.

History

AlphaMissense, developed by DeepMind (Google), represents a significant leap forward in predicting the pathogenicity of missense variants—genetic mutations that alter single amino acids in proteins.

Built on AlphaFold2’s protein structure prediction framework, this AI model leverages evolutionary data and structural context to classify variants as benign or pathogenic with unprecedented accuracy. Unlike traditional methods, it doesn’t require explicit training on clinical databases, reducing biases from human curation.

The paper describing AlphaMissense algorithm was published in September 2023.