Target Audience

  • Researchers, engineers, and readers interested in AlphaMissense, genomic AI, and the historical development of machine learning for biological prediction.
  • Those working in or exploring variant pathogenicity prediction, rare disease diagnosis, precision medicine, and drug discovery.
  • Readers interested in understanding what current AI systems can and cannot predict about biological mechanisms, particularly in the context of whole-body and multiscale biological simulation.

Summary

  • AlphaMissense's major innovation was to combine evolutionary and protein structural knowledge to predict the pathogenicity of approximately 71 million human missense variants.
  • Its fundamental assumption is that variants poorly tolerated by evolution are more likely to be pathogenic in humans. However, evolutionary fitness, protein function, and clinical pathogenicity are related but distinct concepts.
  • Its impressive benchmark performance comes with important qualifications. An AUROC of 0.940 on ClinVar does not guarantee equivalent diagnostic performance on previously unseen patient variants. Independent evaluations published in 2025–2026 have revealed important limitations.
  • Subsequent research has developed in several directions. popEVE focuses on identifying candidate disease-causing variants in patients, VESM on high-performance sequence-based prediction, FuncVEP on experimental functional evidence, and AlphaGenome on regulatory and splicing effects.
  • The central challenge ahead is to move beyond classifying variants as pathogenic or benign toward explaining how genetic changes alter molecular functions, affect cells and tissues, and ultimately contribute to disease.

1. What Did AlphaMissense Change in the History of Genomic AI?

In September 2023, Jun Cheng, Žiga Avsec, and colleagues at Google DeepMind published a paper in Science titled "Accurate proteome-wide missense variant effect prediction with AlphaMissense" [1].

AlphaMissense is an AI model that builds on technology developed for AlphaFold, a system that achieved major advances in protein structure prediction, and applies that knowledge to predicting the pathogenicity of human genetic variants.

The researchers released predictions for approximately 71 million missense variants across 19,233 canonical human proteins.

They also provided predictions for approximately 216 million possible single-amino-acid substitutions [1].

What makes AlphaMissense important is not simply that it achieved higher classification performance than earlier models.

Its historical significance lies in bringing together three previously developing areas of research into a large-scale, practical system for variant interpretation:

  1. Evolutionary biology and population genetics: Inferring how well genetic variants are tolerated from evolutionary conservation and variants observed in populations.
  2. Deep learning for proteins: Learning complex relationships among amino acids, including dependencies related to protein sequence and three-dimensional structure.
  3. Clinical genomics: Prioritizing variants that may contribute to disease among the large numbers identified through patient genome sequencing.

By transferring knowledge learned from protein evolution and structure to clinical variant interpretation, AlphaMissense connected these three fields.

However, understanding its success requires examining a fundamental question:

Are variants that evolution tends not to tolerate necessarily variants that cause disease in humans?

This question is the starting point for understanding both AlphaMissense's place in research history and its limitations.

2. What Does AlphaMissense Actually Predict?

To understand AlphaMissense's prediction target, we first need to distinguish the biological stages between a DNA variant and the development of disease.

2.1 Multiple Biological Processes Connect DNA Variants to Disease

A change in DNA sequence can affect biological systems through several different mechanisms.

For example, replacing one nucleotide, such as A with G, may result in:

  • A change in the amino acid sequence of a protein: A missense variant may substitute one amino acid for another.
  • A change in RNA processing: Splicing may be altered, producing different RNA transcripts or protein products.
  • A change in gene expression: The amount of gene expression, or the cells and tissues in which a gene is expressed, may change.

These changes can affect molecular functions, cellular processes, and tissues, sometimes contributing to disease.

Conceptually, the relationship can be summarized as:

DNA variant → Molecular functional change → Cellular and tissue changes → Organismal phenotype or disease

This is a simplified framework for understanding biological causation. Real biological systems involve interacting pathways, feedback, and effects that are not strictly unidirectional.

AlphaMissense primarily addresses missense variants that change the amino acid sequence of proteins.

2.2 Protein Functional Effects and Pathogenicity Are Not the Same

One of the most important distinctions is between three concepts:

Concept Meaning Central question
Evolutionary fitness The effect of a variant on survival and reproduction Is the variant likely to persist through evolution?
Protein function Effects on protein activity, stability, binding, localization, and other properties How does the variant change the protein's behavior?
Clinical pathogenicity Whether a variant contributes causally to a particular disease Does the variant cause or contribute to disease in a patient?

These concepts are strongly related, but they are not equivalent.

For example, a variant may reduce enzyme activity without causing disease.

Conversely, abnormally increased enzyme activity may cause disease.

In some cases, a variant that reduces a protein's function may even protect against a particular disease.

Therefore, predicting that a variant affects protein function does not automatically establish its clinical pathogenicity.

2.3 AlphaMissense Does Not Directly Predict Disease Penetrance

AlphaMissense produces a pathogenicity score between 0 and 1 for each missense variant.

This score is calibrated using existing clinical variant classification data.

However, an AlphaMissense score of 0.95 does not mean that an individual carrying that variant has a 95% probability of developing a disease.

Estimating disease risk requires additional information, such as:

  • Which disease is being evaluated.
  • Whether the inheritance pattern is autosomal dominant, autosomal recessive, or another mode.
  • What variants are present in the same gene and elsewhere in the genome.
  • The patient's age, sex, environment, and other relevant characteristics.
  • The variant's penetrance and expressivity.

AlphaMissense does not take all these factors as input to calculate an individual's probability of developing a specific disease.

This distinction becomes especially important when considering clinical applications and biological simulation.

3. The Historical Development of Machine Learning for Variant Effect Prediction

To understand why AlphaMissense was a breakthrough, it is useful to examine how successive generations of variant prediction models learned biological information.

The major developments can be summarized as follows.

Period Representative models and research Central approach
2003–2016 SIFT, PolyPhen-2, CADD, REVEL Predict variant effects using evolutionary conservation and engineered features
2017–2021 DeepSequence, PrimateAI, EVE Learn evolutionary constraints and population genetic patterns using deep learning
2021–2023 AlphaFold2, ESM-1b, ESM-2, ESM-1v Develop large-scale representations of protein sequences and structures
2023 AlphaMissense, PrimateAI-3D Integrate evolutionary, structural, and population variation information
2024–2025 Large-scale MAVE studies, popEVE Advance experimental functional assessment and patient-level variant prioritization
2026 VESM, FuncVEP, PATHOS, AlphaGenome Expand knowledge distillation, functional evidence, model integration, and regulatory prediction

3.1 First Generation: Predicting Variant Effects from Evolutionary Conservation

Early representative methods included SIFT (2003) and PolyPhen-2 [2][3].

Their basic premise was:

"If an amino acid has remained highly conserved throughout evolution, it is likely important for protein function, and a mutation at that position may be harmful."

For example, if the same amino acid is conserved across many species, evolutionary constraints may be acting on that position.

This approach proved highly useful.

However, conservation alone cannot fully capture complex relationships between amino acids.

Two residues that are far apart in the primary sequence may be adjacent in the three-dimensional protein structure.

Furthermore, whether a particular substitution is tolerated may depend on the amino acids at other positions.

In 2016, REVEL improved pathogenicity prediction by combining multiple predictors in an ensemble model [4].

3.2 Second Generation: Deep Learning from Evolution's "Natural Experiments"

Beginning around 2017, researchers increasingly used large collections of evolutionarily related protein sequences to learn variant effects.

Representative examples include DeepSequence and EVE [5][6].

These models learn the distribution of protein sequences observed through evolution and estimate how unusual a particular amino acid substitution would be under that distribution.

For example, they can evaluate whether substituting amino acid A with amino acid B is compatible with the amino acids at other positions.

Unlike simple conservation scores, these approaches can learn evolutionary constraints involving combinations of multiple amino acids.

Meanwhile, PrimateAI (2018) used variants observed in human and other primate populations as supervisory signals for training [7].

AlphaMissense strongly inherits this research philosophy of learning from evolutionary information.

3.3 Third Generation: The Emergence of Protein Foundation Models

In 2021, AlphaFold2 substantially advanced protein structure prediction [8].

Protein language models, including the ESM family, also developed rapidly [9].

Just as language models learn relationships between words from large text corpora, protein language models learn biologically plausible sequence patterns from enormous collections of amino acid sequences.

During this period, two major directions emerged:

  • AlphaFold-based approaches: Learning three-dimensional protein structure and spatial relationships among amino acids.
  • ESM-based approaches: Learning general-purpose representations from large protein sequence datasets and applying them to various downstream tasks.

AlphaMissense represents an important example of adapting an AlphaFold-based model to pathogenicity prediction.

PrimateAI-3D also appeared in 2023, combining primate genetic diversity with protein structural information [10].

AlphaMissense was therefore not the only project to explore the integration of evolution and structure.

Multiple research groups were developing related ideas during the same period.

3.4 Did AlphaMissense Replace Earlier Research Paradigms?

An important point is that AlphaMissense did not fundamentally reject the evolutionary-conservation paradigm.

Instead, its development can be understood as:

Prediction from evolutionary conservation → Learning sequence-wide evolutionary constraints → Integrating structural and evolutionary knowledge → Large-scale pathogenicity prediction

AlphaMissense improved the accuracy and scale of this research tradition.

It was a major technological advance, but it retained the fundamental assumption that evolutionary constraints provide useful evidence about pathogenicity.

Recognizing this continuity is essential for understanding the model's implicit assumptions.

4. The Technical Innovations of AlphaMissense

The central design of AlphaMissense is illustrated in Figure 1 of the original paper [1].

4.1 Adapting AlphaFold for Pathogenicity Prediction

AlphaMissense uses an AlphaFold-derived network and is trained in two principal stages.

Stage 1: Protein structure prediction and masked amino acid prediction

An AlphaFold-style network is trained to predict protein structures while also performing a task in which selected amino acids are masked and their original identities must be inferred.

This allows the model to learn not only structural information but also evolutionary sequence constraints.

Stage 2: Fine-tuning using human and primate population variants

Variants observed at relatively high frequencies in human and primate populations are used primarily as benign supervisory signals.

Variants not observed in those populations are used as weak pathogenic supervisory signals.

However, the training weights assigned to rare observed variants are reduced according to their frequencies.

This approach allows the model to learn pathogenicity-related patterns without relying primarily on existing clinical pathogenicity labels as direct training targets.

4.2 The Significance of Learning Without Direct Clinical Labels

Clinical databases such as ClinVar contain variants classified by human experts as pathogenic or benign.

These databases are important resources, but they also have limitations:

  • Well-studied genes tend to accumulate more clinical classifications.
  • Clinical classifications may incorporate findings from previous research or existing computational predictors.
  • Overlap between training and evaluation data can inflate reported model performance.

AlphaMissense attempts to reduce these problems by learning from population variants rather than using ClinVar pathogenicity annotations as its main fine-tuning labels.

However, this does not mean that the model is entirely independent of clinical data.

ClinVar is used for validation, model selection, and score calibration.

Moreover, population allele frequency is itself used in clinical variant classification, creating the possibility of indirect circularity. This issue is discussed in Section 7.

4.3 Reducing Label Noise Through Self-Distillation

The training strategy presents an obvious problem.

Not every variant absent from population databases is pathogenic.

The unobserved variant set includes many variants that are actually benign.

AlphaMissense therefore uses self-distillation.

Predictions from preliminary models are used to filter out unobserved variants considered likely benign, and training is then repeated using the filtered dataset [1].

This reduces label noise, but it does not eliminate the biological distinction between being unobserved and being pathogenic.

The ablation studies in the original paper found that self-distillation improved performance on ProteinGym but not on ClinVar.

By contrast, both AlphaFold-derived pretraining and population-variant fine-tuning were important for achieving strong overall performance.

4.4 AlphaMissense Does Not Simulate Mutant Protein Structures

A crucial characteristic of AlphaMissense is that it does not directly predict the structural changes caused by a mutation.

For example, it does not directly calculate:

  • How many ångströms a protein structure changes after a mutation.
  • How much protein stability changes in kcal/mol.
  • How much enzyme activity or reaction rate changes.
  • How molecular binding affinity changes.

Instead, AlphaMissense transfers structural and evolutionary knowledge learned by AlphaFold into a pathogenicity scoring task.

This is an important example of using representations learned by a large foundation model for a different biological application.

However, using structural knowledge and simulating molecular changes are not the same.

4.5 Major Results from the Original Paper

The principal results reported in the original AlphaMissense paper are summarized below [1].

Evaluation AlphaMissense result Comparison or context
ClinVar classification AUROC 0.940 EVE: 0.911
Mean per-gene ClinVar AUROC 0.950 EVE: 0.921
AUROC for de novo variants in the DDD cohort 0.809 PrimateAI: 0.797
Cancer hotspot AUROC 0.907 VARITY: 0.885
Mean Spearman correlation on ProteinGym 0.514 MAVE data from 72 proteins
Mean Spearman correlation on an additional MAVE dataset 0.450 20 human proteins
Number of missense variant predictions Approximately 71 million Canonical human proteome
Number of single-amino-acid substitution predictions Approximately 216 million 19,233 canonical proteins
Fraction of missense variants assigned a non-ambiguous classification Approximately 89% Likely pathogenic or likely benign

These values reflect the evaluation conditions used in the 2023 paper.

One particularly important aspect is that the authors evaluated the model not only against ClinVar classifications but also against experimental variant effect data from Multiplexed Assays of Variant Effect (MAVEs).

This supports the conclusion that AlphaMissense captures information related to actual protein function, rather than merely reproducing existing clinical annotations.

However, a mean Spearman correlation of 0.514 indicates only partial agreement in the ranking of experimental functional effects.

It does not establish that the model accurately reproduces the magnitude of functional changes or their underlying disease mechanisms.

4.6 What Does Classifying Approximately 89% of Variants Actually Mean?

The original paper classified approximately 32% of human missense variants as likely pathogenic and approximately 57% as likely benign [1].

Together, these categories account for approximately 89% of variants.

The classification thresholds were selected to target 90% precision for each class on ClinVar.

Therefore, the 89% figure means that the model could assign either a likely pathogenic or likely benign label to most variants using the specified thresholds.

It does not mean that 89% of all variants have been experimentally or clinically confirmed.

Likewise, the approximately 90% precision estimate comes from specific ClinVar evaluation conditions.

It is not a guarantee of equivalent precision in every real-world patient population.

This distinction is essential when assessing the practical value of AlphaMissense.

5. Seven Implicit Assumptions Behind AlphaMissense

We can now examine the biological assumptions underlying AlphaMissense's success.

The following seven assumptions are not all explicitly stated as formal assumptions in the original paper.

Rather, they are analytical interpretations of the model's training design, outputs, and evaluation strategy.

Implicit assumption Why it matters
1. Evolutionary constraints reflect human pathogenicity Evolutionary fitness and clinical disease are different concepts
2. Variants absent from populations are more likely to be harmful Rarity or absence does not establish pathogenicity
3. Sequence and structural information can predict pathogenicity Gene expression, cellular state, and environmental factors are not directly modeled
4. Protein functional disruption leads to disease Both loss and gain of function can cause disease, while some functional changes are harmless
5. Individual variants can be evaluated independently Interactions among multiple variants are not directly modeled
6. Pathogenicity can be summarized by a single score Disease identity, inheritance, and penetrance are compressed into one value
7. High performance on known variants generalizes to unknown variants Population biases, disease-specific mechanisms, and evaluation circularity remain

5.1 Assumption 1: Evolutionarily Disadvantageous Variants Are More Likely to Be Pathogenic in Humans

This is the most fundamental assumption.

Amino acids conserved by natural selection are often functionally important, so substitutions at those positions may be harmful.

However, evolutionary fitness and clinical pathogenicity differ in important ways.

For example, diseases that develop after reproductive age may be subject to weaker selection than diseases with severe effects before reproduction.

Furthermore, reducing a particular protein function does not necessarily harm human health.

It may even reduce susceptibility to a particular disease.

Evolutionary constraint is therefore a powerful source of evidence for predicting pathogenicity, but it is not pathogenicity itself.

5.2 Assumption 2: Variants Absent from Populations Are More Likely to Be Pathogenic

AlphaMissense uses the frequency of variants observed in human and primate populations.

This reflects the idea that commonly observed variants tend to be tolerated, while unobserved variants may be more harmful.

However, a variant can be unobserved for several reasons:

  • Strong negative selection.
  • Insufficient population sample size.
  • Recent origin of the variant.
  • Population-specific allele frequencies.
  • A low probability of the mutation occurring in the first place.

Thus, absence from population databases is not equivalent to pathogenicity.

AlphaMissense uses self-distillation to reduce incorrect training labels, but the limitations of using population frequency as a proxy for pathogenicity remain.

5.3 Assumption 3: Protein Sequence and Structure Contain Sufficient Information to Predict Pathogenicity

AlphaMissense uses amino acid sequences, evolutionary alignments, and learned structural representations.

However, the clinical impact of a variant may depend on more than the intrinsic properties of the protein.

Relevant factors include:

  • Which cells and tissues express the gene.
  • The level of gene expression.
  • Which splice isoforms are present.
  • Which other proteins interact with the affected protein.
  • Cellular differentiation and metabolic state.
  • Environmental factors and additional genetic variants.

Another challenge comes from intrinsically disordered regions (IDRs), which do not adopt a single stable three-dimensional structure.

Such regions may still perform essential functions in signaling, molecular interactions, and transcriptional regulation.

AlphaMissense is not designed to evaluate all these conditions using patient-specific or cell-specific inputs.

5.4 Assumption 4: Disruption of Protein Function Leads to Disease

Abnormal protein function causes many genetic diseases.

However, the underlying mechanisms differ substantially.

For example:

  • Loss of Function (LoF): Protein activity is reduced or eliminated.
  • Gain of Function (GoF): Protein activity is increased or a new function is acquired.
  • Dominant Negative: A mutant protein interferes with the function of the normal protein.
  • Haploinsufficiency: A single functional gene copy cannot provide sufficient activity.

These mechanisms may all produce clinically pathogenic variants, but their molecular consequences differ.

A single AlphaMissense score does not explicitly distinguish among them.

5.5 Assumption 5: Individual Variants Can Be Evaluated Independently

AlphaMissense primarily predicts the pathogenicity of individual amino acid substitutions.

However, protein function and disease risk may depend on combinations of variants.

For example, the effect of one mutation may be strengthened or suppressed by another mutation.

Such interactions are known as epistasis.

In clinical genetics, it may also matter whether variants affect both copies of a gene or whether modifier variants are present elsewhere in the genome.

AlphaMissense was not designed to directly evaluate these combinations.

5.6 Assumption 6: Pathogenicity Can Be Represented by a Single Score

AlphaMissense primarily produces a single pathogenicity score.

However, clinically relevant questions are multidimensional.

For a particular variant, we might want to know:

  • Which disease does it cause?
  • What is the probability that a carrier will develop the disease?
  • At what age might symptoms appear?
  • How severe might the disease become?
  • Which tissues are affected?
  • Which treatments might be effective?

Compressing these questions into a single score necessarily omits disease-specific information.

A pathogenicity score can be useful for prioritizing candidate variants, but it does not fully represent disease causation.

5.7 Assumption 7: High Performance on Known Variants Generalizes to Unknown Variants

In machine learning, evaluation requires careful separation of training and test data.

In biology, additional questions arise:

Does performance generalize to previously unstudied genes, unknown disease mechanisms, different ancestry groups, and different experimental conditions?

AlphaMissense was evaluated on multiple datasets in its original publication.

Nevertheless, subsequent research has shown that its performance depends on clinical context and disease mechanism.

Its high benchmark scores are therefore important, but they should not be equated with a complete understanding of biological causation.

6. Limitations Already Visible in the Original AlphaMissense Paper

Some limitations were not discovered only in later studies.

The original AlphaMissense paper already documented several important examples.

6.1 GCK: Missing Pathogenic Gain-of-Function Variants

Figures 3F and 3G of the original paper analyze variants in GCK, the gene encoding glucokinase [1].

Glucokinase is an enzyme involved in glucose metabolism.

Variants that reduce GCK activity can cause maturity-onset diabetes of the young (MODY).

Conversely, variants that abnormally increase GCK activity can cause hyperinsulinemic hypoglycemia.

Thus, different functional effects in the same gene can lead to different diseases.

AlphaMissense performed relatively well on pathogenic GCK variants associated with reduced activity.

However, some pathogenic hyperactive variants, including T65I, were classified as ambiguous or benign.

This is an important failure case.

Predicting whether a variant is evolutionarily unfavorable is different from identifying the direction of functional change and the resulting disease mechanism.

6.2 Reduced Performance in Intrinsically Disordered Regions

The original paper also reported reduced prediction performance for variants located in intrinsically disordered regions [1].

These protein regions do not adopt a single stable structure.

Nevertheless, they may be essential for cellular signaling, molecular interactions, and transcriptional regulation.

The absence of a stable structure does not imply a lack of biological importance.

The reduced performance of AlphaMissense in these regions illustrates a limitation of approaches that benefit strongly from structural information.

6.3 Experimental Functional Effects Are Not Fully Reproduced

AlphaMissense achieved a mean Spearman correlation of 0.514 on ProteinGym MAVE data from 72 proteins.

It also achieved a mean correlation of 0.450 on an additional dataset of 20 human proteins [1].

The original paper examined SHOC2 variants in detail.

SHOC2 forms a complex with MRAS and PP1C and is involved in Ras-MAPK signaling.

AlphaMissense successfully captured important functional constraints in regions such as the RVxF motif and leucine-rich repeats.

These findings support the value of evolutionary and structural information for predicting functional constraints.

However, agreement with experimental measurements remains incomplete.

Moreover, MAVE results can depend on the cell system, experimental design, and measured phenotype.

A high correlation with one functional assay does not automatically establish accurate prediction of clinical pathogenicity.

6.4 Predicting the Importance of Short Genes

Figure 4 of the original paper examines the relationship between average AlphaMissense scores and gene-level functional importance [1].

Traditional methods often infer gene importance by measuring the depletion of potentially damaging variants in human populations.

For short genes, however, the small number of expected variants can limit statistical power.

Because AlphaMissense predicts the effects of possible substitutions across each gene, it can potentially estimate functional importance even when population data are insufficient.

For short genes where LOEUF had limited statistical power, the model achieved an AUROC of approximately 0.88 for distinguishing essential from nonessential genes, compared with approximately 0.81 for LOEUF.

The authors also demonstrated this capability using short proteins in the SF3b spliceosome complex.

This is an important achievement.

However, identifying a gene as essential for cell survival is not the same as determining whether an individual variant causes a particular human disease.

7. Limitations Revealed by Subsequent Studies

Between 2024 and 2026, independent studies examined AlphaMissense using functional experiments and real patient cohorts.

These studies showed that its strong performance on curated benchmarks cannot automatically be generalized to every gene, disease, or patient population.

7.1 IRF6: Benign Variants Can Be Predicted as Pathogenic

A critical review published in 2024 discussed disagreements between AlphaMissense predictions and experimental functional assessments of IRF6 variants [21].

In one example, 15 of 18 variants considered benign based on functional experiments were predicted as pathogenic by AlphaMissense.

This result comes from a limited experimental set involving a particular gene.

It does not imply that AlphaMissense has an overall false-positive rate of 83.3%.

However, it shows that systematic errors can occur for particular genes or functional mechanisms.

7.2 CPA1: Reduced Protein Function Does Not Necessarily Imply Pathogenicity

The same review discussed CPA1 [21].

For certain disease mechanisms involving CPA1, reduced catalytic activity alone does not explain pathogenicity.

Nevertheless, AlphaMissense may classify some variants with reduced catalytic activity as highly pathogenic.

This illustrates the distinction between detecting a functional change and establishing that the change causes a specific disease.

7.3 Performance Differs in Real Rare Disease Cohorts

In 2025, Kong and colleagues evaluated AlphaMissense using a cohort of 7,454 patients and family members covering more than 45 rare diseases [22].

The cohort included 3,383 patients with suspected genetic disorders.

The study compared AlphaMissense predictions with expert clinical classifications.

The main reported results included:

Metric Result
Precision 32.9%
Recall 57.6%
Cohort size 7,454 individuals
Disease coverage More than 45 rare diseases

These results appear very different from the original ClinVar AUROC of 0.940.

However, the metrics and evaluation populations differ, so they cannot be directly compared as evidence that the model itself became less accurate.

The key finding is that distinguishing known clinically classified variants is not the same as identifying causal variants in real patients with rare diseases.

The study also reported problems associated with intrinsically disordered regions.

It further emphasized that gene-level essentiality scores require careful interpretation for genes containing such regions.

7.4 Indirect Circularity Between Population Frequency and Clinical Classification

One important design feature of AlphaMissense is that it learns from population variation rather than primarily from ClinVar pathogenicity annotations.

However, a subtle issue remains.

Population allele frequency is itself used as evidence in clinical variant classification.

For example, a variant that is common in the general population may be unlikely to cause a severe, highly penetrant rare genetic disease.

Thus, some of the population genetic information used in AlphaMissense training overlaps conceptually with information used to assign clinical labels in ClinVar.

This creates the possibility of indirect circularity even when training and evaluation variants do not directly overlap.

The 2026 VESM paper examined related concerns, reporting that AlphaMissense performance depends on allele frequency [12].

VESM, by contrast, explores ways of improving discrimination between rare benign and pathogenic variants without directly using population allele frequency as a prediction feature.

This raises an important question:

How much of AlphaMissense's high performance reflects generalizable knowledge of protein evolution and structure, and how much is supported by population genetic information shared with clinical classification practices?

However, using population frequency is not inherently inappropriate.

Population frequency is a valuable biological signal for variant interpretation.

The methodological issue is how independently a model is evaluated relative to the information used during training.

7.5 Different Disease Mechanisms Favor Different Predictors

A study published in Genetics in Medicine in May 2026 compared AlphaMissense, PrimateAI-3D, and REVEL using a framework based on ACMG/ClinGen clinical classification recommendations [16].

The study reported several differences:

  • AlphaMissense showed strengths for autosomal dominant disease genes associated with haploinsufficiency.
  • REVEL performed better for autosomal recessive variants and certain autosomal dominant genes that tolerate loss of function.
  • Predictor performance and appropriate calibration varied by disease mechanism.

These results indicate that AlphaMissense remains a useful predictor in 2026, but it is not universally optimal across all variants and disease mechanisms.

7.6 Clinical Calibration Requires Additional Validation

AlphaMissense scores are calibrated against ClinVar.

However, clinical interpretation of variants requires integrating multiple evidence types under ACMG/AMP and ClinGen frameworks [19][23].

Computational predictions may contribute evidence under criteria such as PP3 and BP4.

To use them appropriately, their scores must be calibrated against independent data to determine the strength of the evidence they provide.

A 2025 study also examined the calibration of AlphaMissense, ESM1b, and VARITY scores under this framework [24].

Again, classification performance and the strength of evidence suitable for clinical decision-making are not the same.

8. What Has AlphaMissense Not Yet Validated or Solved?

The limitations of AlphaMissense extend beyond the need to improve prediction accuracy.

It is also necessary to distinguish which biological phenomena the model can predict and which fall outside its scope.

Research question Current status Main reason
Can it accurately predict structural changes after mutation? Outside AlphaMissense's direct scope It does not output mutant protein structures
Can it predict changes in enzyme activity or binding affinity quantitatively? Not fully solved It does not directly output experimental functional measurements
Can it distinguish gain-of-function from loss-of-function mechanisms? Documented failure cases exist It misclassifies some variants, including those in GCK
Can it determine whether a variant causes a particular disease? Not generally solved It does not explicitly take the disease or patient phenotype as input
Can it predict disease penetrance or severity? Outside its direct scope It does not model patient-specific conditions
Can it predict interactions among multiple variants? Outside its direct scope It primarily evaluates individual amino acid substitutions
Can it predict cell-type- or tissue-specific functional effects? Outside its direct scope Cell types and cellular states are not explicit inputs
Does it maintain equal performance across ancestry groups? Generalizability is not fully established Population frequency and dataset representation may introduce differences
Does it maintain performance on previously unstudied genes? Generalizability remains uncertain Evolutionary information and available evidence vary among genes
How much does it improve real-world diagnostic yield? Not prospectively demonstrated in the original paper Evaluation of the complete diagnostic workflow is needed
Can it predict treatment response or drug efficacy? Outside its direct scope Pathogenicity and therapeutic response are different outcomes

8.1 The Gap Between Pathogenicity Prediction and Causal Explanation

When AlphaMissense assigns a high pathogenicity score, it suggests that a variant has characteristics associated with known pathogenic variants.

However, drug discovery and biological simulation require answers to more detailed questions:

  • Which molecular function is affected?
  • Is the change a gain or loss of function?
  • In which cells and tissues does the effect occur?
  • Which signaling or metabolic pathways are affected?
  • How does the molecular change produce clinical symptoms?
  • Can a drug or genetic intervention reverse the effect?

These questions go beyond variant classification.

The last question, in particular, concerns causal prediction: what would happen to a biological system if a therapeutic intervention were applied?

AlphaMissense is not designed to predict such intervention effects.

An important direction for future research is therefore to connect pathogenicity prediction with molecular functional changes and causal explanations of disease.

9. How Are Subsequent Studies Changing the Field?

Research following AlphaMissense has not simply focused on improving accuracy on the same classification task.

Instead, it has expanded into different types of training data, prediction targets, and clinical applications.

The principal studies can be compared as follows.

Model Year Main contribution Relationship to AlphaMissense
PrimateAI-3D 2023 Integrates primate genetic diversity with structural information A closely related contemporary approach
popEVE 2025 Develops proteome-wide scores for patient-level causal variant prioritization Extends variant prediction toward patient-level interpretation
VESM 2026 Distills knowledge from ESM models for high-performance sequence-only inference Emphasizes efficient representation and generalization
FuncVEP 2026 Incorporates experimentally measured functional effects Focuses on functional evidence
PATHOS 2026 Integrates multiple protein language models and biological features Combines different pretrained representations
AlphaGenome 2026 Predicts regulatory and splicing effects from DNA sequence Addresses different molecular layers beyond missense variants

9.1 popEVE (2025): Toward Identifying Causal Variants in Patients

Paper: Proteome-wide model for human disease genetics [11]

popEVE is particularly important when comparing subsequent research with AlphaMissense.

Whereas AlphaMissense primarily provides pathogenicity scores for individual variants, popEVE focuses on comparing variants across different genes and prioritizing those most likely to explain a patient's disease.

To achieve this, it accounts for differences in evolutionary constraints among genes while constructing scores that are comparable across the proteome.

The paper reported candidate variants across 442 genes in a severe developmental disorder cohort, including 123 novel candidate genes.

It also demonstrated the potential to prioritize causal candidates using only a child's exome data, even when parental genome information was unavailable.

This research can be understood as a step toward connecting variant-level prediction with patient-level causal variant discovery.

However, correctly ranking candidate variants and accurately classifying individual variants as pathogenic or benign are different tasks.

The models therefore cannot be compared adequately using a single AUROC value.

9.2 VESM (2026): High Performance Without Explicit Structural or MSA Inputs

Paper: Compressing the collective knowledge of ESM into a single protein language model [12]

VESM was published in Nature Methods in March 2026.

AlphaMissense combines structural information, multiple sequence alignments (MSAs), and population genetic data.

VESM instead uses co-distillation to consolidate knowledge from multiple ESM models, enabling high-performance variant effect prediction using protein sequences alone at inference time.

The important finding is that strong prediction performance may be achieved without explicitly providing structural information or MSAs during inference.

The paper reported performance exceeding that of complex models, including AlphaMissense, on several benchmarks.

It also examined performance differences associated with variant allele frequency and the challenge of distinguishing rare benign variants from pathogenic ones.

The researchers additionally developed VESM3, which transfers VESM knowledge to a model incorporating structural information, and VESM++, which integrates sequence and structural representations.

These results do not mean that structural information has become unnecessary.

Instead, they suggest a new research question: How can evolutionary and structural knowledge be learned, compressed, and represented most effectively?

9.3 FuncVEP (2026): From Evolutionary Constraints to Experimental Functional Evidence

Paper: Prediction of human missense variant effects from functional evidence [13]

FuncVEP was published in Nature Genetics on August 26, 2026.

While AlphaMissense emphasizes evolutionary and population variant data, FuncVEP uses diverse experimental functional datasets to predict the functional effects of missense variants.

The paper compared FuncVEP with 48 existing predictors and reported the following improvements:

Evaluation Comparison performance FuncVEP performance
Functional benchmark 78.8% 84.6%
Clinical benchmark 90.1% 92.4%

These values come from the evaluation conditions defined in the FuncVEP study and cannot be compared directly with the AUROC of 0.940 reported in the original AlphaMissense paper.

Using data from UK Biobank and the Mount Sinai Million Health Discoveries Program, the authors also analyzed 494 genes associated with inborn errors of immunity and reported 210 new gene–phenotype associations.

The research direction is clear.

Rather than relying only on the evolutionary intolerance of variants, FuncVEP learns more directly from experimentally measured changes in protein function.

However, functional measurements depend on the cell system and experimental conditions.

Even a model trained on experimental functional evidence cannot necessarily provide a complete explanation of pathogenicity or disease mechanisms in patients.

9.4 PATHOS (2026): Integrating Multiple Foundation Models

Paper: PATHOS: Predicting variant pathogenicity by combining protein language models and biological features [14]

PATHOS combines embeddings from protein language models such as ESM-C 600M and Ankh2 Large with phylogenetic information, allele frequencies, and protein annotations.

The paper compared PATHOS with 65 existing predictors, reporting an MCC of 0.591 on an independently curated clinical dataset and 0.826 on ClinVar data.

Whereas AlphaMissense is centered on an integrated AlphaFold-derived network, PATHOS combines features extracted from multiple pretrained foundation models.

This suggests that integrating the representations learned by different models can be a valuable research direction, even without developing an entirely new foundation model from scratch.

9.5 AlphaGenome (2026): From Protein Variants to DNA Regulatory Mechanisms

Paper: Advancing regulatory variant effect prediction with AlphaGenome [15]

AlphaGenome was published in Nature on January 28, 2026.

It is not a direct successor to AlphaMissense, but it represents an important development in machine learning for genomics.

Whereas AlphaMissense primarily addresses missense variants, AlphaGenome takes DNA sequences of up to approximately one million base pairs as input and predicts various molecular functions.

Its main prediction targets include:

  • Gene expression
  • Chromatin accessibility
  • Transcription factor binding
  • Histone modifications
  • RNA splicing
  • Chromatin contacts

The paper reported performance equal to or better than the strongest existing external models on 25 of 26 variant effect prediction evaluations.

Jun Cheng and Žiga Avsec, who contributed to AlphaMissense, also participated in the AlphaGenome research.

The two models can be compared as follows.

Comparison AlphaMissense AlphaGenome
Main inputs Protein sequence, MSA, missense variant DNA sequence
Main prediction target Missense variant pathogenicity Sequence-dependent molecular functions and regulatory effects
Primary strengths Protein amino acid substitutions Gene regulation, noncoding regions, splicing
Main limitations Does not directly explain disease specificity or direction of functional change Not a dedicated model for clinical pathogenicity of missense variants
Potential applications Prioritizing candidate variants in rare diseases Investigating regulatory and splicing abnormalities

The two models are more complementary than competitive.

A possible future research direction is to integrate predictions across DNA sequence, RNA processing, protein function, and cellular states.

However, this does not mean that a unified, fully validated biological simulation model already exists.

10. Which Models Should Be Used in 2026?

Newer models are not necessarily better than AlphaMissense for every application.

The appropriate choice depends on the biological question and the information available.

Main objective Relevant models or methods Selection considerations
Broad assessment of human missense variants AlphaMissense Provides precomputed scores across the human proteome
Prioritizing causal variants in rare disease patients popEVE with clinical information Focuses on patient-level variant prioritization
Predicting variant effects from sequence alone VESM, VespaG Reduces input requirements and computational burden
Predicting functional effects on proteins FuncVEP, MAVE datasets Emphasizes experimentally measured functional evidence
Supporting clinical missense variant classification REVEL, AlphaMissense, PrimateAI-3D Requires calibration appropriate to disease mechanisms
Evaluating noncoding variant effects AlphaGenome and related models Addresses gene regulation and splicing

There is also an important caveat when combining multiple predictors.

Different models may use overlapping evolutionary information or population frequency data.

Their predictions are therefore not necessarily statistically independent.

Agreement among several models does not automatically mean that several independent pieces of evidence support the same conclusion.

Correlations among predictors and overlap in training information need to be considered.

11. How Is AlphaMissense Being Incorporated into Practical Tools?

AlphaMissense is increasingly used not only as a research model but also as a source of information in genomic analysis workflows.

11.1 Integration into Ensembl

In June 2026, Ensembl release 116 added AlphaMissense and ESM1b prediction scores to variant annotations, replacing MutationAssessor and MetaLR [17].

The information is provided through dbNSFP v5.2c.

This update illustrates how AlphaMissense has moved from a published research model to an information resource used in practical genomic variant analysis.

In addition, AlphaMissenseR was released in 2025 for the analysis, visualization, and validation of AlphaMissense predictions within the R/Bioconductor ecosystem [18].

11.2 Clinical Applications Require Integration of Multiple Evidence Types

In clinical practice, an AlphaMissense score alone is not sufficient to establish the pathogenicity of a variant.

The ACMG/AMP guidelines recommend integrating several types of evidence, including population frequencies, familial segregation, functional experiments, and clinical phenotypes [19].

Computational predictions are one component of this broader evidence framework.

Appropriate calibration and independent validation are needed before using model scores as evidence in clinical classification.

11.3 Toward Databases Integrating Multiple Predictors

The 2026 version of MissenseDB integrates predictions from AlphaMissense, VESM, and VESPA, allowing users to compare variant-level scores and structural locations across approximately 19,400 human proteins [20].

Resources of this kind illustrate a shift toward analyzing variants using multiple sources of information rather than relying on a single prediction model.

However, using multiple models does not by itself guarantee correctness.

12. What Research Is Needed Next?

Overcoming AlphaMissense's limitations will require more than developing larger or more accurate models.

It is also essential to define precisely which biological capabilities are being evaluated.

Five directions are particularly important.

12.1 Disease-Mechanism-Specific Prediction

Future models need to distinguish among loss of function, gain of function, dominant-negative effects, haploinsufficiency, and other mechanisms.

Genes such as GCK, where decreased and increased activity can cause different diseases, provide valuable test cases.

Prediction performance should be evaluated separately for different disease mechanisms.

12.2 Connecting Predictions to Experimental Functional Evidence

Researchers need to integrate multiple forms of experimental evidence, including MAVEs, enzyme activity measurements, protein stability, binding affinity, and cellular localization.

It is equally important to establish how each experimental measurement relates to the relevant disease mechanism.

Models such as FuncVEP represent progress in this direction.

12.3 Incorporating Patient Genotypes and Phenotypes

The clinical significance of a variant may depend on a patient's genetic background and specific disease context.

Future approaches should integrate variant-level predictions with phenotypes, family history, inheritance patterns, and combinations of variants.

popEVE represents one direction toward patient-level causal variant prioritization.

12.4 Independent Evaluation with Reduced Circularity

To understand what a model has genuinely learned, researchers must identify possible overlap between training information and evaluation criteria.

Useful evaluations include:

  • Performance stratified by population allele frequency.
  • Performance stratified by evolutionary conservation.
  • Evaluation by gene and protein domain.
  • Separate assessment of loss-of-function and gain-of-function variants.
  • Temporal evaluation using variants classified after model development.
  • Independent MAVE datasets not used during training.
  • Evaluation across different ancestry groups.

Metrics should include more than AUROC.

AUPRC, false-positive rates, false-negative rates, calibration error, and patient-level causal variant identification rates may also be relevant.

In particular, when a computational predictor has contributed to a clinical variant classification, that classification should not automatically be treated as independent ground truth for evaluating the same predictor.

12.5 Multiscale Prediction from Molecules to Cells and Organisms

The longer-term challenge is to predict the relationships between DNA variants, molecular functions, cellular processes, tissue functions, and organismal phenotypes.

However, simply connecting models from different biological layers is not sufficient.

Each model introduces errors and uncertainty, and researchers must understand how these propagate across layers.

Furthermore, explaining disease mechanisms requires more than correlation.

Causal models capable of predicting the consequences of interventions will become increasingly important.

13. Implications for Biological Simulation, Precision Medicine, and Drug Discovery

The relationship between AlphaMissense and subsequent research suggests several directions for future applications.

13.1 Implications for Precision Medicine

Precision medicine requires identifying which variants among the many present in an individual's genome are most likely to contribute to disease.

Models such as AlphaMissense can help prioritize candidate variants.

Research such as popEVE brings these predictions closer to patient-level causal variant discovery.

However, variant scores alone cannot directly establish disease penetrance or severity.

Integrating computational predictions with clinical phenotypes, inheritance patterns, and other patient information remains essential.

13.2 Implications for Drug Discovery

Drug discovery requires more than knowing that a variant may be pathogenic.

It is also necessary to understand which molecular functions the variant changes.

For example, diseases driven by loss of function and those driven by gain of function may require different therapeutic strategies.

Understanding changes in protein activity, molecular interactions, signaling pathways, and cellular functions is therefore important.

Research using experimental functional evidence, such as FuncVEP, represents one direction toward this goal.

However, functional predictions alone cannot directly establish a drug's efficacy or safety.

13.3 Implications for Biological Simulation

From the perspective of biological simulation, AlphaMissense represents an important step but not a completed simulation technology.

AlphaMissense predicts variant pathogenicity.

FuncVEP learns from experimentally measured functional effects.

AlphaGenome predicts molecular functions such as gene expression and splicing from DNA sequence.

These technologies address different biological layers.

Integrating them may eventually enable more systematic analysis of how genetic variants alter molecular functions, change cellular states, and contribute to disease.

However, classifying variant pathogenicity, predicting molecular functional changes, and simulating biological system dynamics are three distinct capabilities.

Biological simulation must also account for temporal dynamics, interactions among cells, environmental changes, therapeutic interventions, and uncertainty.

Current genomic AI research is advancing some of the component technologies needed for such integrated models.

For readers wishing to understand the broader research history, the following papers provide a useful reading sequence.

14.1 EVE (2021)

Disease variant prediction with deep generative models of evolutionary data [6]

This paper uses deep generative models of evolutionary sequence data to predict variant effects.

It helps clarify which ideas AlphaMissense inherited and which aspects it improved.

14.2 AlphaMissense (2023)

Accurate proteome-wide missense variant effect prediction with AlphaMissense [1]

This is the central paper discussed in this article.

Particularly important are Figure 1 on training design, Figure 2 on clinical classification, Figure 3 on MAVE and GCK analyses, and Figure 4 on gene essentiality.

14.3 Making Sense of Missense (2024)

Making sense of missense [21]

This critical review examines the limitations of pathogenicity predictors, including AlphaMissense, in relation to disease mechanisms and functional experiments.

Examples involving IRF6 and CPA1 illustrate why functional effects and clinical pathogenicity must be distinguished.

14.4 Independent Evaluation in a Rare Disease Cohort (2025)

Discordance between a deep learning model and clinical-grade variant pathogenicity classification in a rare disease cohort [22]

This study evaluates AlphaMissense using real rare disease patient data.

It helps explain why strong performance on curated benchmarks does not necessarily translate into equally strong performance in patient-level causal variant identification.

14.5 popEVE (2025)

Proteome-wide model for human disease genetics [11]

This paper explores how variant-level predictions can be used to prioritize causal variants in patients.

14.6 VESM (2026)

Compressing the collective knowledge of ESM into a single protein language model [12]

This paper examines how information learned from protein sequences can be integrated and compressed to achieve high-performance variant effect prediction.

It also raises important questions about AlphaMissense's dependence on population allele frequency.

14.7 FuncVEP (2026)

Prediction of human missense variant effects from functional evidence [13]

This research illustrates how variant prediction is developing toward greater use of experimentally measured functional evidence.

14.8 AlphaGenome (2026)

Advancing regulatory variant effect prediction with AlphaGenome [15]

This paper demonstrates how AI-based genomic prediction is expanding beyond protein amino acid substitutions to regulatory DNA, gene expression, and splicing.

15. Overall Assessment: What Has AlphaMissense Achieved, and What Has It Not?

AlphaMissense's place in the history of genomic AI can be summarized as follows.

AlphaMissense demonstrated that biological knowledge accumulated through protein evolution and structural learning can be transferred to large-scale, high-performance prediction of human missense variant pathogenicity.

This was a significant achievement.

Only a small fraction of the enormous number of human genetic variants have established clinical classifications.

AlphaMissense provided large-scale predictions for variants without existing classifications.

It also demonstrated that a foundation model originally developed for protein structure prediction could be adapted to a practical problem in clinical genomics.

However, AlphaMissense has not yet explained how individual variants alter specific molecular functions, how those changes affect particular cells and tissues, or why they cause disease in particular patients.

This distinction can be summarized as a progression of research capabilities:

  1. Earlier research: Infer whether variants may be harmful from evolutionary conservation and sequence characteristics.
  2. AlphaMissense: Integrate evolutionary and structural knowledge for large-scale pathogenicity prediction across the human proteome.
  3. Research in 2024–2026: Combine functional experiments, patient information, and multiple models to improve prediction and interpretation.
  4. Future research challenges: Explain causal relationships from genetic variants to molecular functions, cellular states, and disease, including responses to interventions.

These developments should not be interpreted as a single linear progression.

Variant pathogenicity classification, protein functional prediction, patient-level causal variant identification, and regulatory DNA analysis are different research problems.

Models such as popEVE, VESM, FuncVEP, and AlphaGenome are therefore better understood as related but distinct efforts rather than simple replacements for AlphaMissense.

15.1 The September 2026 Correction to the Original Paper

The original AlphaMissense paper includes a correction dated September 23, 2026 [1].

The supplementary description of classification thresholds was corrected to clarify that the thresholds were calculated using the full, class-imbalanced ClinVar dataset rather than a class-balanced dataset.

This distinction matters because the evaluation conditions for ClinVar AUROC and the conditions used to determine classification thresholds are not identical.

Researchers attempting to reproduce the results, recalibrate the scores, or compare AlphaMissense rigorously with other predictors should take this correction into account.

15.2 Key Questions for Evaluating Future Research

When evaluating future research related to AlphaMissense, it is important to ask more than whether a model achieves a higher AUROC.

Is the model predicting evolutionary constraints?

Is it predicting changes in protein function?

Is it predicting pathogenicity for a particular disease?

Or is it explaining the causal mechanisms that produce disease in a patient?

These capabilities are related, but they are not equivalent.

In particular, comparing AlphaMissense, popEVE, VESM, and FuncVEP on shared independent variant datasets could provide important insights.

Stratifying evaluations by functional effect direction, disease mechanism, population frequency, gene, and protein structure would make it possible to examine which biological capabilities have genuinely improved.

Such studies could also help clarify which forms of learned biological knowledge are responsible for higher prediction performance.

15.3 What AlphaMissense Reveals About the Current State of AI in Biology

Understanding AlphaMissense means more than understanding a single variant prediction model.

It reveals that machine learning has reached a stage where statistical regularities learned from enormous biological datasets can be used to evaluate previously unseen variants.

However, successful statistical prediction and understanding biological causation remain different achievements.

AlphaMissense is an especially important case study for examining this gap.

The central question for future research is this: When a variant is predicted to be pathogenic, which molecular function does it change, in what direction, and in which cellular environment? Through what biological processes does that change lead to symptoms in a patient, and can those processes be modified by treatment?

AlphaMissense's lasting significance is not that it has already answered these questions.

Rather, it demonstrated the feasibility of large-scale variant effect prediction from evolutionary and structural knowledge, while helping clarify the next scientific challenges: understanding functional mechanisms, explaining disease causation, and eventually developing more comprehensive biological simulations.

References

[1] J. Cheng et al., "Accurate proteome-wide missense variant effect prediction with AlphaMissense," Science, vol. 381, eadg7492, 2023. https://doi.org/10.1126/science.adg7492. (Includes the correction dated September 23, 2026.)

[2] P. C. Ng and S. Henikoff, "SIFT: Predicting amino acid changes that affect protein function," Nucleic Acids Research, vol. 31, pp. 3812–3814, 2003. https://doi.org/10.1093/nar/gkg509.

[3] I. Adzhubei, D. M. Jordan, and S. R. Sunyaev, "Predicting functional effect of human missense mutations using PolyPhen-2," Current Protocols in Human Genetics, vol. 76, pp. 7.20.1–7.20.41, 2013. https://doi.org/10.1002/0471142905.hg0720s76.

[4] N. M. Ioannidis et al., "REVEL: An ensemble method for predicting the pathogenicity of rare missense variants," The American Journal of Human Genetics, vol. 99, pp. 877–885, 2016. https://doi.org/10.1016/j.ajhg.2016.08.016.

[5] T. A. Hopf et al., "Mutation effects predicted from sequence co-variation," Nature Biotechnology, vol. 35, pp. 128–135, 2017. https://doi.org/10.1038/nbt.3769.

[6] J. Frazer et al., "Disease variant prediction with deep generative models of evolutionary data," Nature, vol. 599, pp. 91–95, 2021. https://doi.org/10.1038/s41586-021-04043-8.

[7] L. Sundaram et al., "Predicting the clinical impact of human mutation with deep neural networks," Nature Genetics, vol. 50, pp. 1161–1170, 2018. https://doi.org/10.1038/s41588-018-0167-z.

[8] J. Jumper et al., "Highly accurate protein structure prediction with AlphaFold," Nature, vol. 596, pp. 583–589, 2021. https://doi.org/10.1038/s41586-021-03819-2.

[9] A. Rives et al., "Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences," Proceedings of the National Academy of Sciences, vol. 118, e2016239118, 2021. https://doi.org/10.1073/pnas.2016239118.

[10] PrimateAI-3D-related research, "PrimateAI-3D," Science, 2023. (Study integrating protein structure and primate genetic diversity for variant effect prediction.)

[11] "Proteome-wide model for human disease genetics," Nature Genetics, 2025. https://www.nature.com/articles/s41588-025-02400-1.

[12] "Compressing the collective knowledge of ESM into a single protein language model," Nature Methods, 2026. https://www.nature.com/articles/s41592-026-03050-9.

[13] B. Kayaalp et al., "Prediction of human missense variant effects from functional evidence," Nature Genetics, vol. 58, pp. 2350–2360, 2026. https://www.nature.com/articles/s41588-026-02727-3.

[14] "PATHOS: Predicting variant pathogenicity by combining protein language models and biological features," Artificial Intelligence in the Life Sciences, 2026. https://www.sciencedirect.com/science/article/pii/S2667318526000139.

[15] "Advancing regulatory variant effect prediction with AlphaGenome," Nature, 2026. https://www.nature.com/articles/s41586-025-10014-0.

[16] "Calibration and evaluation of machine-learning algorithms for missense variant classification under ACMG/ClinGen recommendations," Genetics in Medicine, 2026. https://doi.org/10.1016/j.gim.2026.102606.

[17] L. P. Mirabueno, "New pathogenicity predictions available for human variants," Ensembl Blog, June 25, 2026. https://www.ensembl.info/2026/06/25/new-pathogenicity-predictions-available-for-human-variants/.

[18] AlphaMissenseR, R/Bioconductor package for AlphaMissense prediction analysis and visualization, 2025. https://bioconductor.org/.

[19] S. Richards et al., "Standards and guidelines for the interpretation of sequence variants: A joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology," Genetics in Medicine, vol. 17, pp. 405–424, 2015. https://doi.org/10.1038/gim.2015.30.

[20] MissenseDB, integrated missense variant prediction resource including AlphaMissense, VESM, and VESPA, 2026.

[21] "Making sense of missense," 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC11683568/.

[22] S. W. Kong et al., "Discordance between a deep learning model and clinical-grade variant pathogenicity classification in a rare disease cohort," npj Genomic Medicine, vol. 10, article 17, 2025. https://www.nature.com/articles/s41525-025-00480-w.

[23] ClinGen Sequence Variant Interpretation Working Group, recommendations for applying computational evidence to variant interpretation. https://clinicalgenome.org/working-groups/sequence-variant-interpretation/.

[24] "Calibration of additional computational tools," 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12208618/.