Target Audience

  • Researchers and professionals interested in biological simulation and AI models that predict biological phenomena.
  • Readers interested in personalized medicine, genomic analysis, and identifying disease-causing genetic variants.
  • Those seeking to understand AI-based protein function prediction and its potential applications in drug discovery and development.

Summary

  • AlphaMissense applies AI representations learned from protein structure and evolutionary information to predict the pathogenicity of human genetic variants.
  • Published in 2023, it represented an important milestone in genomic AI by providing predictions for approximately 71 million possible human missense variants.
  • By 2026, models such as popEVE, VESM, FuncVEP, and AlphaGenome have extended this research toward patient-level variant prioritization, protein function prediction, and regulatory DNA analysis.
  • The field is progressing from predicting whether a variant is harmful toward understanding how variants alter molecular functions and contribute to disease.
  • These developments have important implications for personalized medicine, drug discovery, and the longer-term advancement of biological simulation.

1. What Did AlphaMissense Change?

AlphaMissense is a research achievement that applies AI models trained on protein structure and evolutionary information to predict the pathogenicity of human genetic variants across the entire human proteome.

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

Its historical significance goes beyond achieving higher prediction accuracy than previous variant effect prediction models.

More importantly, AlphaMissense brought together three research areas that had previously developed along largely separate paths.

  1. Evolutionary biology and population genetics: These fields estimate how harmful a genetic variant might be based on evolutionarily conserved amino acids and variants observed in populations.
  2. Deep learning for proteins: This field learns complex dependencies between amino acids and structural information from large collections of protein sequences.
  3. Clinical genomics: This field aims to identify which variants, among the many detected in a patient's genome, are most likely to contribute to disease.

By integrating these areas, AlphaMissense provided pathogenicity scores for approximately 71 million possible missense variants in the human proteome [1].

By 2026, research had advanced further in three particularly important directions.

  1. Connecting pathogenicity prediction to the identification of disease-causing variants in individual patients: Examples include popEVE (2025).
  2. More accurately distinguishing and integrating changes in protein function and clinical pathogenicity: Examples include research published in 2026 that incorporates experimental functional data.
  3. Expanding predictions beyond proteins to include gene regulation, splicing, and other DNA-level mechanisms: Examples include AlphaGenome (2026).

AlphaMissense therefore remains an important foundational technology, but it is not the final solution to predicting the pathogenicity of every type of genomic variant.

Understanding its position requires first distinguishing the biological phenomena that different models attempt to predict.

2. What Exactly Do These Models Predict?

A genetic variant can affect human disease through several biological stages.

Broadly speaking, the process begins with a change in the DNA sequence.

For example, a nucleotide may change from A to G.

Such a change may produce different molecular consequences.

  • A change in the protein sequence: A missense variant may alter an amino acid in the encoded protein.
  • A change in how a gene functions: A DNA variant may alter gene expression, RNA splicing, or other regulatory processes.

These molecular changes may subsequently affect cellular or physiological functions and, in some cases, contribute to disease.

AlphaMissense and AlphaGenome focus on different parts of this process.

  • AlphaMissense: Primarily predicts the pathogenicity of amino acid substitutions using protein structural and evolutionary information.
  • AlphaGenome: Primarily predicts the effects of DNA sequences on molecular functions such as gene regulation and RNA splicing.

However, AlphaMissense does not explicitly calculate every stage between a DNA variant and the development of a disease.

Changes in Protein Function Are Not the Same as Pathogenicity

An important distinction is that a change in protein function is not necessarily equivalent to causing disease.

For example, a variant may reduce the activity of an enzyme without causing a clinically meaningful disorder.

Conversely, some diseases arise when a variant causes an abnormal increase in enzyme activity.

AlphaMissense does not directly measure or predict every possible functional consequence.

Instead, it primarily estimates variant pathogenicity using evolutionary constraints and structural features [1].

This distinction becomes particularly important when comparing AlphaMissense with more recent approaches.

From the perspective of biological simulation, predicting whether a variant is harmful and predicting how that variant changes molecular or cellular states are fundamentally different tasks.

3. The Evolution of Machine Learning in Genomic Variant Research

One useful way to understand the evolution of this field is to examine what kinds of biological data have been used to train predictive models.

The major developments can be summarized as follows.

Period Representative Models and Research Main Technological Characteristics
2003–2016 SIFT, PolyPhen-2, CADD, REVEL Use evolutionary conservation and manually designed biological features
2017–2021 DeepSequence, PrimateAI, EVE Learn variant tolerance from evolutionary sequences and population genetic data using deep learning
2021–2023 AlphaFold2, ESM-1b, ESM-2, ESM-1v Develop large-scale representations of protein structure and function
2023 AlphaMissense Integrate AlphaFold-derived structural and evolutionary representations with weak supervision from population genetics
2024–2025 Large-scale MAVE studies, popEVE Improve generalization to less-studied genes and connect variant predictions to patient-level disease analysis
2026 VESM, FuncVEP, AlphaGenome Advance knowledge distillation, experimental functional data integration, and noncoding variant prediction

3.1 First Generation: Predictions Based on Handcrafted Features

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

Their basic principle can be illustrated by the following reasoning.

If an amino acid at a particular position has remained almost unchanged throughout evolution, that position is probably important for protein function. A mutation at that position may therefore be harmful.

These approaches were already capable of producing useful predictions.

However, they faced limitations in capturing complex biological relationships.

For example, two amino acids that are far apart in the linear protein sequence may be physically adjacent in the three-dimensional structure.

Similarly, the effect of a mutation at one position may depend on the identities of amino acids at other positions.

Representing these relationships with manually designed features was difficult.

Later, ensemble methods such as REVEL improved prediction accuracy by combining multiple existing predictors [4].

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

Beginning around 2017, researchers increasingly attempted to learn the effects of mutations directly from evolutionarily related protein sequences.

Representative approaches included DeepSequence and EVE [5][6].

Consider a mutation that changes an amino acid from A to B at a particular position.

These models estimate how unusual that substitution would be relative to patterns observed across naturally evolved protein sequences.

Unlike simple conservation scores, such models can learn constraints that depend on combinations of multiple amino acids.

Meanwhile, PrimateAI (2018) used variants observed in humans and other primates as training signals [7].

AlphaMissense strongly inherited this evolutionary modeling approach.

3.3 Third Generation: The Emergence of Protein Foundation Models

The development of AlphaFold2 in 2021 marked a major advance in protein three-dimensional structure prediction [8].

At approximately the same time, protein language models such as ESM were progressing rapidly [9].

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

Two major technological directions began to converge.

  • AlphaFold-based approaches: Particularly effective at representing three-dimensional structures and spatial relationships between amino acids.
  • ESM-based approaches: Learn general-purpose representations from large protein sequence datasets and apply them to various downstream biological tasks.

AlphaMissense is a representative example of adapting AlphaFold-derived representations to pathogenicity prediction.

It is also important to recognize that AlphaMissense was not the only research effort moving in this direction.

PrimateAI-3D, introduced in 2023, also combined protein structural information with genetic variation observed in primates [10].

Several research groups were therefore developing related ideas around the same time.

4. The Technical Innovations of AlphaMissense

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

The model uses a network derived from AlphaFold, combining protein structure prediction with learning objectives designed to capture evolutionary sequence constraints.

4.1 Fine-Tuning AlphaFold for Pathogenicity Prediction

AlphaMissense training consists of two major stages.

Stage 1: Protein Structure Prediction and Masked Amino Acid Prediction

The model uses an AlphaFold-like network to learn structural information and constraints embedded in evolutionarily related protein sequences.

In addition to predicting protein structure, the model is trained to predict amino acids that have been masked from the input sequence.

Stage 2: Fine-Tuning Using Human and Primate Population Variants

Variants frequently observed in populations are primarily used as examples of benign variation.

Variants that have not been observed are treated as candidates for the pathogenic class.

The model ultimately produces a pathogenicity score for each missense variant.

Calibrated scores are used to assign variants to three categories.

  • Likely benign: The variant is predicted to have a low likelihood of pathogenicity.
  • Ambiguous: The prediction is uncertain.
  • Likely pathogenic: The variant is predicted to have a high likelihood of pathogenicity.

One particularly important aspect is that AlphaMissense does not use manually curated pathogenicity labels from databases such as ClinVar as its primary fine-tuning targets.

Training directly on clinical annotations can cause a model to inherit biases associated with previously studied disease genes and historical clinical interpretations.

AlphaMissense attempts to reduce this problem by using the observed distribution of genetic variants in populations.

However, it would be inaccurate to describe AlphaMissense as a model that never uses clinical labels.

ClinVar data are used for model selection, validation, and score calibration.

Furthermore, an unobserved variant is not necessarily pathogenic.

This means that the training labels contain noise.

The paper also introduces self-distillation to reduce the impact of these noisy labels [1].

4.2 Using Structural Knowledge Rather Than Predicting Mutant Structures

Another important distinction is that AlphaMissense does not directly predict the three-dimensional structure of a protein after a mutation.

For example, it does not directly calculate how many ångströms a structure changes following a mutation or how much the binding energy changes.

Instead, it uses the structural and evolutionary representations learned by AlphaFold to estimate pathogenicity.

This is a representative example of a broader approach in modern AI:

Knowledge learned by a large foundation model can be transferred to a different biological task.

4.3 How Significant Were the Experimental Results?

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

Evaluation Metric AlphaMissense Result Comparison or Context
ClinVar classification AUROC 0.940 EVE achieved 0.911
Mean Spearman correlation with MAVE data 0.514 ProteinGym benchmark covering 72 proteins
Number of possible missense variants predicted Approximately 71 million Approximately 216 million possible amino acid substitutions
Proportion of variants assigned a confident category 89% Thresholds selected to target 90% precision on ClinVar

These values come from the original 2023 study and should not be interpreted as direct comparisons against models published in 2026.

One of the most important aspects of the evaluation is that AlphaMissense was tested not only against ClinVar annotations but also against experimental measurements collected through Multiplexed Assays of Variant Effect (MAVEs).

These assays measure the functional effects of large numbers of protein variants.

Agreement with MAVE measurements supports the conclusion that AlphaMissense captures biologically relevant information rather than merely reproducing existing clinical labels.

However, a mean Spearman correlation of 0.514 also indicates that AlphaMissense does not perfectly reproduce the magnitudes of functional effects.

4.4 What the GCK Example Reveals About the Limitations of AlphaMissense

Figure 3 of the original paper includes an analysis of variants in GCK, the gene encoding glucokinase [1].

Glucokinase is an enzyme involved in glucose metabolism.

AlphaMissense performs relatively well in identifying pathogenic variants associated with reduced GCK activity.

However, some variants that cause disease through abnormally increased enzyme activity are classified as benign or ambiguous.

This example illustrates a fundamental limitation.

Predicting whether a mutation is evolutionarily harmful is not equivalent to predicting the direction of its functional effect or its disease mechanism.

The original paper also investigates gene-level predictions in Figure 4.

By calculating average AlphaMissense scores across variants within a gene, the researchers show that the model can help identify functionally important short genes that are difficult to evaluate using conventional population-based statistical approaches.

This finding provides an important conceptual link to subsequent research on disease gene discovery and patient-level variant prioritization.

5. Comparison with Major Research Developments from 2024 to 2026

Research following AlphaMissense has expanded beyond improving binary pathogenicity classification.

Recent studies explore new prediction targets, training strategies, and connections to clinical decision-making.

Six particularly relevant models are summarized below.

Model Year Main Difference from AlphaMissense
PrimateAI-3D 2023 Combines primate genetic diversity with structural information in a closely related approach
popEVE 2025 Makes variant severity more comparable across genes and emphasizes identification of disease-causing variants in patients
VESM 2026 Achieves strong variant effect prediction using protein sequence alone at inference time
FuncVEP 2026 Incorporates experimentally measured functional effects into model training
PATHOS 2026 Combines embeddings from multiple protein language models with biological features
AlphaGenome 2026 Extends variant effect prediction toward regulatory DNA, gene expression, and RNA splicing

5.1 popEVE (2025): From Pathogenicity Prediction to Disease Variant Discovery

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

popEVE is particularly important when examining the development of the field after AlphaMissense.

AlphaMissense primarily addresses the question:

"Is this variant likely to be pathogenic?"

popEVE places greater emphasis on a question that is closer to clinical decision-making:

"Among many candidate variants, which is most likely to be relevant to a patient's disease?"

A central feature of popEVE is its attempt to make variant severity scores comparable across the proteome while accounting for evolutionary constraints that differ between genes.

The paper reports candidate variants across 442 genes in a cohort involving severe developmental disorders, including 123 novel candidate genes.

It also demonstrates that candidate variants can be prioritized even when only the child's exome data are available, without requiring genomic information from both parents.

Position relative to AlphaMissense

popEVE can be understood as extending the large-scale variant prediction paradigm toward patient-level identification of disease-associated variants.

However, candidate variant prioritization and binary variant pathogenicity classification are different evaluation tasks.

It would therefore be inappropriate to determine which model is superior using a single AUROC value.

5.2 VESM (2026): Can High Accuracy Be Achieved Without Structural Inputs?

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

VESM was published in Nature Methods in March 2026.

It provides an interesting comparison with AlphaMissense from the perspective of model architecture.

AlphaMissense achieves high performance by combining protein structural information, multiple sequence alignments (MSAs), and population genetic information.

However, models relying on such inputs can incur substantial computational costs and data preparation requirements.

VESM uses co-distillation to consolidate knowledge from multiple ESM models.

Its objective is to achieve strong variant effect prediction using only protein sequences during inference.

The study reports performance comparable to or exceeding that of complex hybrid models across multiple benchmarks.

It also evaluates effects associated with continuous clinical phenotypes in biobank datasets.

Position relative to AlphaMissense

In the AlphaMissense approach, explicitly using structural information and MSAs provides important predictive advantages.

VESM demonstrates the possibility of recovering comparable predictive capabilities through improved protein language model training, without explicitly providing all such information at inference time.

This can be interpreted as a shift:

From improving accuracy by integrating more information sources toward compressing learned biological knowledge into efficient predictive models.

5.3 FuncVEP (2026): Learning Directly from Experimental Functional Measurements

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

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

This study addresses limitations of AlphaMissense by using a different source of training information.

AlphaMissense relies heavily on evolutionary observations and population frequencies as training signals.

In contrast, FuncVEP emphasizes learning the effects of variants on protein function from diverse experimental functional datasets.

The paper reports comparisons against 48 existing predictors.

It reports improvements from 78.8% to 84.6% on functional benchmarks and from 90.1% to 92.4% on clinical benchmarks.

The study also reports 210 new gene–phenotype associations related to inborn errors of immunity.

These performance values are specific to the evaluation conditions used in the FuncVEP study.

They cannot be compared directly with the AUROC of 0.940 reported in the original AlphaMissense paper.

Position relative to AlphaMissense

The research direction is clear.

Rather than relying only on evolutionary evidence to identify variants that are unlikely to be tolerated, FuncVEP attempts to learn more directly from experimentally observed functional changes.

This is an important development toward connecting pathogenicity prediction with biological mechanisms.

However, experimental functional measurements also depend on factors such as the cell system used and the biological activity being measured.

Consequently, experimental functional effects do not always correspond directly to clinical pathogenicity.

5.4 PATHOS (2026): Combining Multiple Foundation Models

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

PATHOS combines embeddings obtained from two protein language models, ESM-C 600M and Ankh2 Large.

It also incorporates phylogenetic information, allele frequencies, and protein annotations.

The paper compares PATHOS with 65 existing predictors.

It reports a Matthews correlation coefficient (MCC) of 0.591 on a separately curated clinical dataset and 0.826 on a ClinVar dataset.

Position relative to AlphaMissense

AlphaMissense is centered on an integrated network derived from AlphaFold.

PATHOS instead uses representations from several existing foundation models as features for pathogenicity prediction.

This approach illustrates how high-performing predictive models can be developed by integrating knowledge from existing foundation models rather than building a new large-scale model entirely from scratch.

5.5 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.

However, it is highly relevant to understanding the broader direction of machine learning in genomic research.

Whereas AlphaMissense primarily focuses on amino acid substitutions in proteins, AlphaGenome takes DNA sequences of up to approximately one million bases as input.

It predicts molecular properties including:

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

The paper reports performance matching or exceeding the strongest external baseline models in 25 of 26 variant effect prediction evaluations.

Another noteworthy connection is that Jun Cheng and Žiga Avsec, who contributed to AlphaMissense, also participated in the AlphaGenome research.

Although the prediction targets differ, the research reflects an expansion of the underlying idea:

Learning biological functions directly from molecular sequences.

In this case, the focus expands from protein-level consequences to DNA regulatory mechanisms.

Position relative to AlphaMissense

AlphaMissense and AlphaGenome are better understood as complementary approaches than as direct competitors.

Category AlphaMissense AlphaGenome
Main inputs Protein sequences, MSAs, missense variants DNA sequences
Main predictions Pathogenicity of missense variants Effects of DNA sequences on molecular functions and regulation
Primary strengths Amino acid substitutions in proteins Noncoding regions, gene regulation, RNA splicing
Main limitations Does not directly explain the direction of functional changes or disease specificity Is not a specialized model for directly classifying the pathogenicity of protein amino acid substitutions
Potential applications Prioritization of candidate variants in rare diseases Analysis of gene regulatory and splicing abnormalities

An important future research challenge is to integrate predictions across these different molecular levels into a unified patient- or disease-oriented framework.

However, this should be understood as a possible research direction, not as evidence that such a complete and fully validated model already exists.

6. Which Models Are Actually Better in 2026?

An important consideration is that newer models are not necessarily better than AlphaMissense for every application.

A study published in May 2026 provides an example of this distinction.

"Calibration and evaluation of machine-learning algorithms for missense variant classification under ACMG/ClinGen recommendations," published in Genetics in Medicine, compared AlphaMissense, PrimateAI-3D, and REVEL using clinical data [16].

The study reported several differences.

  • REVEL performed particularly well overall for autosomal recessive disorders and certain autosomal dominant disorders involving genes relatively tolerant of loss-of-function variation.
  • AlphaMissense showed stronger performance for autosomal dominant variants in genes where haploinsufficiency is an important disease mechanism.
  • Different models exhibited strengths for different disease mechanisms, suggesting that they can provide complementary information.

AlphaMissense therefore remains a useful predictive tool in 2026, but it is not necessarily the optimal standalone model for every variant.

6.1 Choosing Models According to the Research Objective

Primary Objective Relevant Models or Methods Rationale
Broad evaluation of human missense variants AlphaMissense Provides precomputed predictions across the human proteome
Identification of candidate variants in rare disease patients popEVE + clinical information Emphasizes patient-level variant prioritization
Efficient variant effect prediction from sequences VESM, VespaG Reduces inference requirements and computational complexity
Evaluation of changes in protein function FuncVEP, MAVE data Emphasizes experimental functional evidence
Clinical classification of missense variants REVEL, AlphaMissense, PrimateAI-3D Requires calibration and evaluation appropriate to the disease mechanism
Evaluation of noncoding variants AlphaGenome and related models Predicts effects on gene regulation, RNA splicing, and related processes

It is also important to recognize that placing scores from multiple models side by side does not automatically create independent evidence.

Different models may rely on overlapping evolutionary information or training datasets.

Their correlations and potential data circularity must therefore be considered when evaluating combined predictions.

7. How AlphaMissense Connects to Technologies Used Today

Three developments are particularly important in understanding how AlphaMissense has moved from research toward practical applications.

7.1 Integration into Common Genomic Analysis Tools

In June 2026, Ensembl release 116 introduced AlphaMissense and ESM1b prediction scores into its variant annotation resources, replacing MutationAssessor and MetaLR [17].

This reflects a shift in AlphaMissense's role.

Rather than remaining only a research model demonstrating strong performance, its predictions are becoming part of the commonly available information used to interpret genomic variants.

In 2025, AlphaMissenseR was also released to support analysis, visualization, and validation of AlphaMissense predictions in the R/Bioconductor environment [18].

7.2 Clinical Use Requires Converting Model Scores into Evidence Strength

Suppose AlphaMissense produces a pathogenicity score of 0.95.

This should not be interpreted as meaning that a patient has a 95% probability of developing a disease.

The score is an estimate of pathogenicity calibrated using ClinVar data.

It is not a direct estimate of disease penetrance in the general population or of the probability that a specific variant explains an individual patient's condition.

Clinical applications require careful calibration within frameworks such as those established by ClinGen and the American College of Medical Genetics and Genomics (ACMG).

In particular, computational predictions may contribute to PP3/BP4 evidence under ACMG/AMP variant interpretation guidelines [19].

Research published in 2025 also investigated calibration of AlphaMissense, ESM1b, and VARITY scores within this framework.

Thus, progress in the field involves not only improving model accuracy but also developing methods for translating AI outputs into evidence that can be used appropriately in clinical interpretation.

7.3 Development of Integrated Multi-Model Databases

The 2026 version of MissenseDB integrates predictions from AlphaMissense, VESM, and VESPA.

It provides access to variant scores and their positions in protein structures for approximately 19,400 human proteins [20].

This illustrates a broader direction in practical genomic analysis.

Future systems may increasingly rely on comparing outputs from multiple specialized models rather than making decisions based on a single large model.

8. Important Problems That AlphaMissense Still Cannot Solve

Four challenges are particularly important when evaluating future developments.

Challenge Why It Is Difficult Emerging Research Direction
Direction of functional effects Both loss-of-function and gain-of-function variants can cause disease Integration of functional assay data and mechanism-specific models
Disease specificity The same variant may have different consequences depending on tissue or disease context Integration of phenotype, cell-type, and tissue information
Generalization to unknown genes Strong performance on known disease genes does not guarantee equivalent performance on less-studied genes Proteome-wide calibration and approaches such as popEVE
Circularity in evaluation datasets Indirect overlap between training and evaluation data can inflate estimated performance Independent clinical datasets and new MAVE experiments

Among these challenges, the gap between predicting pathogenicity and explaining disease mechanisms is particularly important.

A high AlphaMissense score does not necessarily explain which molecular function is altered, in what direction the function changes, or which cells are affected.

Addressing these questions requires integrating information across multiple biological levels.

Relevant data may include genomic sequences, protein structures, functional experiments, cellular states, and patient phenotypes.

These challenges are also important for biological simulation, personalized medicine, and drug development.

A prediction that a variant may be pathogenic can help identify candidates for further investigation.

However, understanding how that variant changes molecular and cellular behavior may be necessary to investigate disease mechanisms and potential therapeutic targets.

Achieving this requires more than pathogenicity prediction alone.

It requires integrating information from different levels of biological organization.

The following papers provide a useful reading sequence for understanding the historical development of this research area.

9.1 EVE (2021)

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

This study uses deep generative models trained on evolutionary sequence data to estimate the effects of genetic variants.

It provides a foundation for understanding which ideas AlphaMissense inherited and how its approach differs from earlier methods.

9.2 AlphaMissense (2023)

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

This is the central paper discussed in this article.

Figure 1 is particularly useful for understanding the training architecture.

Figures 2 and 3 provide important evidence regarding classification performance and agreement with experimental functional measurements.

9.3 popEVE (2025)

Proteome-wide model for human disease genetics [11]

This study provides a useful perspective on how variant-level predictions can be connected to identifying disease-causing variants in individual patients.

9.4 VESM (2026)

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

This paper helps explain the differences between integrating multiple biological information sources and developing high-performing sequence-only models.

9.5 FuncVEP (2026)

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

This study illustrates the growing importance of incorporating experimentally measured functional effects into predictive models.

9.6 AlphaGenome (2026)

Advancing regulatory variant effect prediction with AlphaGenome [15]

This paper demonstrates how genomic AI is expanding beyond protein missense variants toward gene regulation, RNA splicing, and other molecular functions.

10. The Overall Position of AlphaMissense

The historical role of AlphaMissense can be summarized as follows:

It connected foundation models developed for protein structure prediction to the practical and clinically important problem of interpreting genetic variants in humans.

By 2026, research had progressed toward a broader set of objectives.

  1. Earlier approaches: Predict whether genetic variants are harmful.
  2. AlphaMissense: Integrate evolutionary and structural knowledge to predict variant pathogenicity at proteome-wide scale.
  3. Current research: Combine functional experiments, patient information, and multiple predictive models.
  4. Future direction: Explain how genetic variants lead to changes in molecular functions, cellular processes, and disease.

The long-term significance of AlphaMissense may lie not only in its improvements in pathogenicity classification accuracy at the time of publication.

It also demonstrated that combining large-scale biological pretraining with task-specific weakly supervised learning can be effective for practical genomic variant interpretation.

From this perspective, models such as popEVE, VESM, and FuncVEP should not be viewed simply as replacements for AlphaMissense.

They represent different attempts to address limitations and opportunities revealed by this line of research.

10.1 Implications for Biological Simulation, Personalized Medicine, and Drug Discovery

From the perspectives of biological simulation, personalized medicine, and drug discovery, these developments share an important direction.

That direction is connecting the prediction of molecular functions from biological sequences to predictions relevant to patients and diseases.

AlphaMissense represents an important step by estimating the pathogenicity of genetic variants.

popEVE moves closer to identifying disease-associated variants in patients.

FuncVEP incorporates experimental functional evidence.

AlphaGenome expands prediction toward DNA regulatory mechanisms.

These studies operate on different types of information and different biological levels.

If such predictions can eventually be integrated, they may contribute to a more systematic understanding of how genetic variants affect molecular functions, cellular states, and disease processes.

However, this remains a research direction.

It does not mean that current models can already simulate every biological process within the human body accurately.

In personalized medicine, an important objective is to evaluate whether an individual's genetic variants are associated with disease.

In drug discovery and development, understanding how those variants alter molecular functions and contribute to disease mechanisms is also important.

Research following AlphaMissense is expanding the knowledge needed to connect these two objectives.

10.2 The 2026 Correction to the AlphaMissense Paper

The original AlphaMissense paper includes a correction issued in September 2026 [1].

The supplementary materials were updated to clarify that the classification thresholds were based on the full, non-balanced ClinVar dataset rather than the class-balanced ClinVar dataset.

This correction should be considered when reproducing the model's evaluation procedures or calibrating classification thresholds.

10.3 How Should Future Research Be Evaluated?

One particularly valuable next step would be to compare AlphaMissense, popEVE, VESM, and FuncVEP using the same independent variant dataset.

Such a comparison would help distinguish improvements in different capabilities rather than relying solely on performance claims reported in separate papers.

Understanding AlphaMissense is therefore about more than understanding one genomic variant prediction model.

It provides a way to understand how machine learning is progressing from learning the structures and evolutionary constraints of biological molecules toward explaining disease mechanisms and supporting personalized medicine.

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 September 2026 correction.)

[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. (Related to the development of DeepSequence.)

[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. (Research combining structural information with genetic variation across primates.)

[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] “Prediction of human missense variant effects from functional evidence,” Nature Genetics, 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.

[17] Ensembl, “Ensembl Release 116,” 2026. https://www.ensembl.org/.

[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. (See also the 2025 research on calibration of AlphaMissense, ESM1b, and VARITY scores.)

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