Target Audience

  • Readers interested in AlphaMissense, how it works, its place in the history of genomic AI research, and the strengths and limitations of its predictions.
  • Researchers and practitioners interested in precision medicine, identifying disease-causing variants in rare diseases, AI-based protein function prediction, and drug discovery.
  • Readers interested in biological simulation and AI technologies for predicting biological phenomena, including their potential and the challenges that remain unresolved.

Summary

  • AlphaMissense is an AI model that uses protein structural and evolutionary information to predict the pathogenicity of human missense variants at scale. Its original 2023 paper demonstrated strong performance against known clinical annotations and experimental measurements of variant effects.
  • A fundamental assumption underlying AlphaMissense is that variants that are poorly tolerated by evolution are more likely to be associated with human disease. However, evolutionary fitness, changes in protein function, and clinical pathogenicity are not equivalent.
  • Subsequent studies have identified limitations, including missed gain-of-function pathogenic variants, false-positive pathogenicity predictions, reduced performance in intrinsically disordered regions, and potential indirect circularity arising from the use of population allele frequencies.
  • AlphaMissense pathogenicity scores do not directly represent an individual patient's probability of developing a disease, the probability that a variant causes a specific condition, the direction of protein functional changes, or treatment outcomes. Clinical applications require validation and calibration appropriate to the disease mechanism.
  • Future research needs to connect pathogenicity predictions derived from evolutionary and structural information with functional experiments, patient phenotypes, cellular and tissue contexts, and the effects of therapeutic interventions. These challenges are also important for advancing precision medicine, drug discovery, and biological simulation.

1. The Implicit Assumptions Underlying AlphaMissense

AlphaMissense is a model for predicting the pathogenicity of missense variants, developed by Jun Cheng, Žiga Avsec, and colleagues at Google DeepMind and published in Science in 2023 [1].

AlphaMissense combines protein sequence, structural, and evolutionary information with population variant data from humans and other primates. It provides pathogenicity scores for approximately 71 million possible missense variants in the human proteome.

Its technological significance lies in adapting representations learned by an AlphaFold-derived system to the practical challenge of interpreting human genomic variants.

However, understanding the biological assumptions that make this approach successful, and the extent to which those assumptions hold, requires further examination.

The central question is as follows.

AlphaMissense is based on the idea that learning evolutionary constraints makes it possible to infer the pathogenicity of human genetic variants to a considerable extent. However, being evolutionarily disadvantageous, altering protein function, and causing disease in a patient are three distinct phenomena.

AlphaMissense exploits statistical relationships among these phenomena.

Although the model has demonstrated that these relationships are informative for many variants, it has not established that they hold across all disease mechanisms.

1.1 What AlphaMissense Infers—and What It Does Not Explain Causally

The distinction between AlphaMissense predictions and the underlying biological causal processes can be understood through four stages.

  1. Evolutionary and population-level patterns of genetic variation
    These patterns provide information about which amino acid substitutions have been tolerated during evolution.

  2. Variant pathogenicity scores
    AlphaMissense estimates the pathogenicity of individual variants using structural, evolutionary, and population-frequency information.

  3. Changes in protein function
    In biological systems, variants may affect enzymatic activity, protein stability, molecular interactions, cellular localization, or cause gains or losses of function.

  4. Disease and clinical manifestations in patients
    Changes in protein function may influence disease onset, severity, age of onset, and treatment response, depending on cellular and tissue conditions, genetic background, and other factors.

Importantly, AlphaMissense does not simulate these four stages sequentially.

Instead, it primarily uses evolutionary, structural, and population-frequency information to estimate pathogenicity scores directly.

Consequently, even when a variant receives a high pathogenicity score, the model does not directly explain which protein function is altered, how that alteration affects particular cells, or how it ultimately contributes to disease in a patient.

1.2 Seven Implicit Assumptions of AlphaMissense

An analysis of AlphaMissense's training methodology and output design suggests seven major implicit assumptions.

Implicit Assumption Why It May Be Problematic
1. Evolutionary constraints reflect human pathogenicity Evolutionary fitness and the development of a specific human disease are not equivalent.
2. Common variants tend to be benign, while unobserved variants tend to be deleterious The relationship between rarity and pathogenicity is probabilistic and has exceptions.
3. Pathogenicity can be inferred from protein sequence and structure Gene expression, cellular states, environmental factors, and other contextual information are not adequately represented.
4. Disruption of protein function leads to pathogenicity Loss of function can be harmless, while gain of function can cause disease.
5. The effects of individual variants can be evaluated independently Interactions among multiple variants and the broader genetic background are not directly evaluated.
6. Pathogenicity can be represented by a single score Disease specificity, tissue context, inheritance patterns, and disease risk are compressed into one value.
7. Strong performance on known clinical datasets generalizes to previously unseen variants Dataset biases, indirect circularity, and unknown disease mechanisms remain potential problems.

These are not necessarily assumptions explicitly stated by the authors of the original paper.

Rather, they are assumptions inferred from the model's training methodology, input information, and output design.

A critical scientific commentary published in 2024 also identified several limitations of AlphaMissense, including insufficient interpretability, an inability to explain functional changes, and a lack of disease specificity [2].

2. Fundamental Assumptions Revealed by the History of Variant Effect Prediction

To understand the assumptions underlying AlphaMissense, it is useful to examine the history of variant effect prediction not only in terms of improvements in predictive accuracy, but also in terms of the biological proxies used to infer pathogenicity.

Research Direction Representative Methods Core Idea Underlying Assumption
Evolutionary conservation SIFT, PolyPhen-2 Variants at conserved positions may be deleterious. Evolutionary conservation corresponds to functional importance.
Learning sequence coevolution DeepSequence, EVE Learn constraints across protein sequences. The distribution of naturally evolved sequences contains information about protein function.
Weak supervision using population genetics PrimateAI Use variants observed in populations as training signals. Population allele frequencies reflect variant tolerance.
Integration of structure and sequence AlphaFold2, AlphaMissense Combine three-dimensional structure and evolutionary information. Structural constraints help predict variant effects.
Disease- and function-specific prediction Research from 2024–2026 Incorporate functional experiments and patient information. The meaning of pathogenicity depends on disease mechanisms and evaluation contexts.

A key point is that AlphaMissense did not fundamentally abandon the earlier research paradigm of using evolutionary constraints as a proxy for pathogenicity.

Rather, it can be understood as a model that improved the accuracy and scalability of this paradigm through deep learning representations derived from AlphaFold and training based on population allele frequencies.

2.1 The Assumption That Evolutionary Fitness Approximately Corresponds to Pathogenicity

This is arguably the most fundamental assumption.

The underlying idea is that amino acid substitutions that have been poorly tolerated during evolution are likely to be disadvantageous to proteins or organisms, and that this information can help predict human disease risk.

However, three concepts need to be distinguished.

  • Evolutionary fitness: The extent to which a genetic variant is likely to persist in a population through survival and reproduction.
  • Protein function: How a variant affects enzymatic activity, structure, molecular interactions, or other functional properties.
  • Clinical pathogenicity: Whether a variant causes disease under a particular inheritance pattern or disease context.

For example, a variant that causes disease after reproductive age may experience weaker natural selection than a variant that causes severe effects before reproduction.

Conversely, a reduction in a particular protein function does not necessarily harm human health. In some cases, it may even protect against disease.

Therefore, evolutionary constraints provide a powerful source of information, but they are not equivalent to the causal mechanisms of disease.

2.2 The Assumption That Unobserved Variants Are More Likely to Be Pathogenic

Figure 1B of the original AlphaMissense paper illustrates how training labels are constructed [1].

Rather than using clinically curated pathogenicity labels as its primary training data, AlphaMissense uses population variant data from humans and other primates.

In particular, frequently observed variants are treated as weak signals of benignity, while variants absent from population datasets are treated as weak signals of pathogenicity.

Rare observed variants receive reduced weights during training.

However, a variant may be unobserved for several different reasons:

  • It may genuinely be subject to strong negative selection.
  • The available sample size may be insufficient.
  • The variant may have arisen recently.
  • Its frequency may differ across populations.
  • The underlying mutation may have a low probability of occurring.

In other words, not all unobserved variants are pathogenic.

[ P(\text{Pathogenic}\mid\text{Unobserved}) < 1 ]

The original paper acknowledges that the training dataset contains numerous incorrect labels and uses self-distillation to reduce the resulting noise [1].

The important distinction is that self-distillation can reduce label noise, but it does not resolve the underlying biological problem that being unobserved is not equivalent to being pathogenic.

2.3 The Assumption That Knowledge of Protein Structure Is Sufficient to Understand Pathogenicity

AlphaMissense uses structural and evolutionary representations learned by an AlphaFold-derived neural network.

It does not directly simulate the three-dimensional structure of a mutated protein or the energy changes caused by a mutation.

Consequently, its output does not directly reveal:

  • How many ångströms the protein structure changes after a mutation.
  • How much protein stability changes in kcal/mol.
  • How much the enzymatic reaction rate changes.
  • Whether binding affinity to a molecular partner increases or decreases.
  • Whether the protein changes its subcellular localization.

This is not simply a matter of insufficient structural resolution.

Even if the three-dimensional structure of a mutated protein could be predicted accurately, that would not necessarily determine its function or clinical pathogenicity.

Indeed, the original AlphaMissense paper explicitly notes that AlphaFold itself is relatively insensitive to individual amino acid substitutions and cannot accurately predict structural changes caused by point mutations [1].

3. What Did the Original AlphaMissense Experiments Actually Validate?

The original AlphaMissense paper evaluated the model using several benchmarks involving clinical pathogenicity annotations and experimental measurements of variant effects [1].

It is important to distinguish what each evaluation metric actually demonstrates.

3.1 Major Results Reported in the Original Paper

Evaluation Result What Was Evaluated
ClinVar classification AUROC 0.940 Ability to distinguish known benign and pathogenic variants
DDD cohort AUROC 0.809 Discrimination between de novo variants found in patients and controls
Mean Spearman correlation with ProteinGym MAVE data 0.514 Rank correlation with experimentally measured variant effects
Fraction of variants assigned a classification 89% Coverage achieved using thresholds calibrated against ClinVar

These results are reported in Figures 2 and 3 and the main text of the original paper.

However, they come from different evaluation settings and should not be interpreted as measures of clinical diagnostic accuracy.

The conclusions supported by each result—and those that remain unsupported—can be summarized as follows.

Evaluation What the Result Supports What It Does Not Establish
ClinVar AUROC of 0.940 The model distinguished known clinical labels with high accuracy in the evaluation dataset. The same accuracy will necessarily be achieved for previously unseen variants in real patients.
DDD AUROC of 0.809 The model demonstrated useful discrimination in a particular developmental disorder dataset. It can identify disease-causing variants in every rare disease.
Mean MAVE correlation of 0.514 Predictions are associated with experimental measurements of functional effects across diverse proteins. The model accurately reproduces the magnitude or mechanism of molecular functional changes.
Classification coverage of 89% The model can assign classifications to a large proportion of variants using specified thresholds. The true pathogenicity of 89% of variants has been experimentally or clinically established.

3.2 What Does It Mean That 89% of Variants Were Classified?

The interpretation of the 89% figure deserves particular attention.

It means that, using thresholds established from known clinical data, AlphaMissense assigned either a "likely benign" or "likely pathogenic" classification to approximately 89% of all possible human missense variants.

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

In the original paper, classification thresholds were chosen to achieve 90% precision for both classes in the ClinVar dataset [1].

Approximately 32% of variants were classified as likely pathogenic, 57% as likely benign, and the remainder as ambiguous.

This was an important achievement because it provided predictions for an enormous number of variants whose clinical significance had not previously been established.

However, providing a prediction must be distinguished from demonstrating the true pathogenicity of a variant.

3.3 The September 2026 Correction

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

The correction clarified that the ClinVar dataset used to determine classification thresholds was the full, non-class-balanced dataset rather than a class-balanced dataset.

This distinction matters when reproducing the methodology or recalibrating the model's scores.

It is also important to remember that AlphaMissense pathogenicity scores cannot be directly interpreted as an individual patient's probability of developing a disease or as the probability that a variant causes a particular condition.

4. Specific Limitations Identified by Subsequent Research

Independent studies published between 2024 and 2026 have shown that several limitations of AlphaMissense are not merely theoretical concerns.

They can also be observed in specific genes and real patient datasets.

4.1 Failure to Identify Some Gain-of-Function Pathogenic Variants: GCK

GCK (glucokinase) is an enzyme involved in glucose metabolism.

Loss-of-function variants in GCK can cause maturity-onset diabetes of the young (MODY), while gain-of-function variants can cause hyperinsulinemic hypoglycemia.

Thus, different directions of functional change in the same gene can lead to distinct diseases.

Figures 3F and 3G of the original AlphaMissense paper illustrate this problem [1].

AlphaMissense performs relatively well in predicting loss-of-function variants in GCK.

However, it sometimes classifies pathogenic gain-of-function variants, such as T65I, as ambiguous or benign.

This example demonstrates the difference between identifying variants that appear evolutionarily unusual and understanding the direction of functional changes that cause disease.

To predict pathogenicity accurately, a model must account not only for how much protein function is impaired, but also for whether function is abnormally increased or whether a new function is acquired.

4.2 Functional Changes Do Not Necessarily Imply Pathogenicity: CPA1

A 2024 commentary also discusses the analysis of variants in CPA1 [2].

In CPA1, reduced catalytic activity alone is insufficient to explain the pathogenicity associated with the relevant disease.

Nevertheless, AlphaMissense may assign high pathogenicity scores to variants that reduce catalytic activity.

This means that even when a model correctly identifies a functional impairment, that impairment may not be the cause of a particular disease.

It illustrates an important distinction between evaluating protein function and evaluating clinical pathogenicity.

4.3 Misclassifying Benign Variants as Pathogenic: IRF6

Experimental studies involving IRF6 have reported discrepancies between AlphaMissense predictions and experimentally assessed functional effects [2][3].

According to the 2024 commentary, AlphaMissense classified 15 of 18 variants experimentally assessed as benign as pathogenic.

Within this particular experimental dataset, approximately 83.3% of the variants assessed as benign were therefore assigned pathogenic predictions.

However, this result comes from a limited experimental dataset involving a single gene.

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

Nevertheless, it demonstrates that systematic prediction errors can occur depending on the gene and the relevant biological function.

4.4 Reduced Performance in a Real-World Rare Disease Cohort

A study published in npj Genomic Medicine in 2025 evaluated AlphaMissense using a cohort of 7,454 patients and family members representing more than 45 rare diseases [4].

The study reported the following results.

Metric Result
Precision 32.9%
Recall 57.6%

These values reflect agreement with variants classified as pathogenic by experts under the definitions used in that study.

The evaluation population and metrics differ from those used to calculate the ClinVar AUROC in the original paper.

The study also identified prediction instability in proteins containing intrinsically disordered regions (IDRs).

This finding is consistent with the reduced performance in disordered regions reported in the original AlphaMissense paper [1].

The broader implication is that a model achieving a high AUROC on a dataset of previously characterized variants may perform differently when applied to patients with rare diseases.

However, because the classification criteria and composition of candidate variants differ between the studies, these results should not be interpreted as evidence that the model itself became less accurate between 2023 and 2025.

4.5 Population Allele Frequencies Can Introduce Indirect Circularity into Evaluation

This is one of the most important concerns raised by subsequent research.

The 2026 VESM paper highlights a potential problem in the evaluation of AlphaMissense [5].

Although AlphaMissense does not directly train on pathogenicity labels from ClinVar, it does use human population allele frequencies during training.

Clinical classifications recorded in ClinVar also use population allele frequencies as part of the supporting evidence.

Therefore, even when training labels do not directly overlap, information can still be shared between the training data and the criteria used to construct evaluation labels.

The VESM study used the March 2025 version of ClinVar and reported that AlphaMissense performance depends on minor allele frequency.

It also found that VESM performed better in an evaluation excluding higher-frequency variants used during training.

These observations do not invalidate the achievements of AlphaMissense.

However, they suggest that part of its strong performance may reflect not only generalizable knowledge of protein structure and evolution, but also population-genetic information shared with clinical classification procedures.

This issue is particularly important when considering the independence of training datasets and benchmarks used to evaluate AI models.

4.6 Conventional Models Can Outperform AlphaMissense Under Certain Clinical Conditions

An independent evaluation published in Genetics in Medicine in May 2026 compared AlphaMissense, PrimateAI-3D, and REVEL [6].

The results were not uniform across disease mechanisms.

  • AlphaMissense performed particularly well for variants in genes associated with haploinsufficiency in autosomal dominant diseases.
  • REVEL performed better for autosomal recessive variants and for certain autosomal dominant genes that are relatively tolerant of loss of function.
  • The best-performing predictor varied depending on the underlying disease mechanism.

Thus, AlphaMissense's strong performance in the original 2023 study does not establish that it is the optimal predictor for every clinical variant interpretation task.

In addition, a 2025 calibration study based on ClinGen recommendations reported that thresholds specified by the developers of AlphaMissense cannot simply be translated into levels of clinical evidence without additional calibration [7].

Therefore, in clinical applications, model scores should not be treated as final pathogenicity classifications without considering the target disease, gene, and intended use.

5. What Has AlphaMissense Not Yet Validated?

When discussing the limitations of AlphaMissense, it is important to distinguish among three categories:

  1. Questions that were not sufficiently evaluated in the original paper.
  2. Questions for which subsequent research has not established a general solution.
  3. Questions that fall outside the model's intended prediction targets.

Something that has not been validated is not necessarily something that has been demonstrated to be impossible.

5.1 Unvalidated Questions and Capabilities Outside the Model's Scope

Question Current Assessment Reason
Can the model accurately predict structural changes after mutation? Outside its direct scope It outputs pathogenicity scores rather than mutated protein structures.
Can it distinguish gain of function from loss of function? Known failure cases exist It fails to distinguish the direction of functional change in examples such as GCK.
Can it determine whether a variant causes a specific disease? Generally unresolved Disease and phenotype information are not direct inputs.
Can it predict disease penetrance, severity, or age of onset? Outside its scope It is not an individualized disease-risk model.
Does it achieve equal accuracy across ancestral populations? Generalization is not sufficiently guaranteed Performance may depend on allele frequencies and the representativeness of population datasets.
Can it predict the combined effects of multiple variants? Not directly validated It was primarily designed for single amino acid substitutions.
Can it predict functional effects in specific cell types or tissues? Outside its scope Tissue identity and cellular state are not direct inputs.
Can it predict drug response or treatment outcomes? Not established in the original study Pathogenicity and pharmacological response are different prediction targets.
How much does it improve prospective clinical diagnostic yield? Not demonstrated in the original study Evaluation of the entire clinical workflow is required.
Does it retain the same performance for novel proteins or sequences with limited evolutionary information? Generalization remains uncertain It depends on the availability and richness of multiple sequence alignments and evolutionary information.

Among these questions, one of the most important concerns is what pathogenicity scores actually represent and how they relate to the causal mechanisms of disease.

5.2 A Fundamental Unvalidated Question: The Causal Meaning of Prediction Scores

Let the output of AlphaMissense be denoted by (s(v)).

The model directly estimates a pathogenicity score associated with variant (v).

However, what clinicians and drug developers ultimately want to know may be a quantity such as:

[ P(D\mid v,g,e,a) ]

Here:

  • (D) represents a particular disease.
  • (v) represents a genetic variant.
  • (g) represents genetic background.
  • (e) represents environmental conditions.
  • (a) represents factors such as age.

This quantity represents the probability of developing a particular disease given a variant, genetic background, environmental conditions, and age-related factors.

It is not equivalent to the pathogenicity score produced by AlphaMissense.

Furthermore, drug discovery requires answering causal questions about what changes when a therapeutic intervention is applied.

For example:

[ P(D\mid do(T=t),v,g,e,a) ]

Here, (T) represents a therapeutic intervention.

The notation (do(T=t)) comes from causal inference and represents the effect of actively implementing an intervention, rather than merely observing patients who happen to receive treatment.

AlphaMissense is not designed to estimate such patient-specific disease probabilities or the causal effects of therapeutic interventions.

Therefore, there remains a substantial gap between obtaining a pathogenicity score and explaining disease causation or predicting treatment outcomes.

6. Research Priorities Remaining as of 2026

Considering the directions in which the field is developing, five research challenges are particularly important.

6.1 Disease Mechanism-Specific Prediction

The first challenge is distinguishing among loss of function (LoF), gain of function (GoF), dominant-negative effects, haploinsufficiency, and other pathogenic mechanisms.

AlphaMissense essentially represents pathogenicity using a single score.

However, different types of functional change in the same protein can cause different diseases.

As illustrated by the GCK example, some gain-of-function pathogenic variants are not adequately captured by the model [1].

Therefore, distinguishing the mechanisms through which variants cause disease directly addresses one of AlphaMissense's most fundamental limitations.

6.2 Connecting Predictions to Functional Experiments and Causal Mechanisms

The second challenge is predicting which molecular functions are altered, using experimental data such as:

  • Multiplexed Assays of Variant Effect (MAVEs)
  • Protein stability measurements
  • Enzymatic activity assays
  • Binding affinity measurements
  • Subcellular localization experiments

AlphaMissense shows meaningful correlations with experimentally measured variant effects.

However, it does not directly predict the magnitude or direction of functional changes.

Models such as FuncVEP, published in 2026, advance a research direction that incorporates experimentally measured functional data into variant effect prediction [8].

Nevertheless, functional assay results can depend on the cell system, measurement method, and biological property being evaluated.

Therefore, agreement with a functional assay does not necessarily constitute proof of clinical pathogenicity.

6.3 Incorporating Patient Phenotypes and Disease Context

The third challenge is predicting how variants relate to particular diseases, inheritance patterns, and patient phenotypes.

AlphaMissense assigns pathogenicity scores to variants themselves.

In clinical practice, however, the critical task is often to determine which of the many variants identified in a patient is responsible for the patient's condition.

Research such as popEVE, published in 2025, emphasizes prioritizing candidate disease-causing variants at the patient level [9].

This represents a shift from variant-level pathogenicity classification toward identifying causal variants in individual patients.

However, prioritizing candidate variants and demonstrating that a particular variant causes a patient's disease remain distinct tasks.

6.4 Evaluating Generalization Using Truly Independent Data

The fourth challenge is evaluating predictive performance separately across population allele frequencies, genes, diseases, and types of functional assays, while carefully controlling for overlap with information used during training.

A particularly important question is how much of the high AUROC reported in existing studies reflects genuine generalizable predictive ability.

AlphaMissense avoids direct training on ClinVar pathogenicity labels partly to reduce dependence on previously established clinical classifications.

However, its use of population allele frequencies creates the possibility of indirect information sharing with clinical labels [5].

Consequently, independent evaluations using newly generated clinical datasets and functional experiments are essential.

6.5 Multiscale Prediction from Molecules to Cells and Organisms

The fifth challenge is modeling how changes in protein function propagate through signaling pathways, metabolism, cellular processes, tissues, and ultimately organism-level phenotypes.

This is a long-term and technically demanding research challenge, especially in the context of biological simulation.

AlphaMissense estimates pathogenicity from protein sequences and evolutionary information, but it does not directly simulate cellular states or organism-level physiological phenomena.

In the future, it will be important to connect changes in protein function with intracellular reactions, tissue-level behavior, and disease progression.

In addition, research published in 2026 has explored gene- and domain-specific calibration of pathogenicity predictors, reporting more useful results than those obtained using universal genome-wide thresholds [10].

This suggests that developing increasingly large AI models is not the only important research direction.

Understanding how to interpret existing model predictions in the appropriate biological context is also essential.

7. What Experiments Would Actually Test AlphaMissense's Assumptions?

From a research perspective, it would be particularly valuable not only to compare how different predictors rank on benchmarks, but also to identify the biological conditions under which each model fails.

For example, AlphaMissense, REVEL, PrimateAI-3D, VESM, and FuncVEP could be evaluated using a common set of datasets.

7.1 Hypotheses and Experimental Approaches Worth Testing

Research Question Experimental or Evaluation Method What We Want to Determine
Relationship between evolutionary information and pathogenicity Stratified evaluation by allele frequency and evolutionary conservation How much predictive performance depends on evolutionary and population-frequency information
Differences between loss of function and gain of function Stratified evaluation using experimentally characterized LoF and GoF variants Whether prediction errors depend on the direction of functional change
Contribution of structural information Evaluation by MSA depth, intrinsically disordered regions, and structured regions Under which conditions structural and evolutionary information are most useful
Differences among disease mechanisms Evaluation by disease and inheritance pattern How strongly predictive performance depends on genes and disease mechanisms
Generalization to unseen variants Temporally separated evaluation using new ClinVar and MAVE datasets Whether models demonstrate predictive ability beyond fitting known annotations
Contribution to clinical diagnosis Evaluation of candidate variant ranking and confirmed diagnoses in patient cohorts How much the models improve real diagnostic workflows

7.2 Why AUROC Alone Is Insufficient

The original AlphaMissense paper reported a high ClinVar classification AUROC of 0.940 [1].

However, this metric alone is insufficient for clinical applications.

Several evaluation metrics should be considered together.

  • AUROC: How well the model distinguishes pathogenic from benign variants across decision thresholds.
  • AUPRC: How precision and recall trade off, including in datasets with relatively few positive examples.
  • False-positive rate: The proportion of benign variants incorrectly classified as pathogenic.
  • False-negative rate: The proportion of pathogenic variants that are missed.
  • Calibration error: How closely predicted scores correspond to observed proportions of pathogenic variants.
  • Patient-level causal variant identification rate: How effectively the model helps identify the relevant disease-causing variants in actual patients.

In rare disease diagnosis, it is often necessary to identify a very small number of causal variants among an enormous number of candidates.

Therefore, even a model with strong variant-level discrimination may not necessarily rank the true causal variant at the top for a particular patient.

7.3 The Importance of Independent Evaluation Data

Clinical evaluations must also ensure that the process used to establish ground-truth labels is independent of the predictor being assessed.

Existing clinical classifications may incorporate information from earlier computational predictors or population allele frequencies.

If the AI model being evaluated shares information with the clinical classification process, performance may be overestimated.

In addition, whether a protein functional change measured in a MAVE experiment actually causes disease must be validated separately.

Therefore, future evaluations should distinguish among variant pathogenicity classification, molecular functional changes, and causal relationships with patient disease.

8. Particularly Important References

Several studies are especially useful for understanding AlphaMissense's implicit assumptions, challenges, limitations, and unvalidated aspects.

8.1 The Original AlphaMissense Paper (2023)

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

This is the foundational paper discussed throughout the article.

Particularly important sections include:

  • Figure 1: Weakly supervised learning methodology
  • Figure 2: Evaluation and calibration using clinical data
  • Figure 3: MAVE benchmarks and analysis of GCK variants

The paper is essential for understanding noise in the training labels, the use of evolutionary and structural information, and the distinction between pathogenicity and functional changes.

Paper: https://doi.org/10.1126/science.adg7492

8.2 Making Sense of Missense (2024)

Making sense of missense [2]

This commentary critically examines the limitations of pathogenicity predictors, including AlphaMissense.

It is particularly useful for discussions of IRF6, CPA1, disease specificity, and interpretability.

The paper helps clarify why changes in protein function do not necessarily correspond to clinical pathogenicity.

Paper: https://pmc.ncbi.nlm.nih.gov/articles/PMC11683568/

8.3 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 [4]

This study independently evaluates AlphaMissense using data from patients with rare diseases.

It is important for understanding why benchmark performance reported in the original study may not generalize directly to real clinical settings.

It is also useful for examining prediction instability in proteins containing intrinsically disordered regions.

Paper: https://www.nature.com/articles/s41525-025-00480-w

8.4 VESM (2026)

Compressing the collective knowledge of ESM into a single protein language model [5]

This study is particularly important because it examines AlphaMissense's dependence on population allele frequencies and the possibility of indirect circularity with clinical labels.

It also helps clarify the distinction between models that integrate multiple information sources, such as AlphaMissense, and protein language models that use only sequence information during inference.

Paper: https://www.nature.com/articles/s41592-026-03050-9

8.5 Clinical Calibration and Evaluation Under ACMG/ClinGen Recommendations (2026)

Calibration and evaluation of machine-learning algorithms for missense variant classification under ACMG/ClinGen recommendations [6]

This study compares AlphaMissense, PrimateAI-3D, and REVEL.

It is useful for understanding how the performance of pathogenicity predictors depends on disease mechanisms and inheritance patterns.

It also addresses the important question of how pathogenicity scores should be interpreted as clinical evidence.

Paper: https://doi.org/10.1016/j.gim.2026.102606

8.6 Clinical Calibration of Computational Prediction Scores (2025)

Calibration of additional computational tools [7]

This study examines how outputs from models such as AlphaMissense can be translated into levels of evidence suitable for clinical use.

It is important for understanding why thresholds specified by model developers should not necessarily be used directly in clinical decision-making.

Paper: https://pmc.ncbi.nlm.nih.gov/articles/PMC12208618/

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

AlphaMissense can be positioned within the history of genomic AI research as follows.

What AlphaMissense achieved was the large-scale and highly accurate application of biological information accumulated through protein evolution and structural learning to the prediction of human missense variant pathogenicity.

This is an important accomplishment.

It made it possible to provide predictions across the human proteome for variants that previously had little or no directly available functional or clinical evidence.

However, what AlphaMissense has not yet achieved is explaining which molecular functions a variant changes, how those changes occur, and why they cause disease in particular cells, tissues, or patients.

Understanding this distinction is essential when considering the future of AI-based biological research.

9.1 From Predicting Evolutionary Constraints to Understanding Disease Mechanisms

AlphaMissense uses statistical regularities learned from evolutionary history to infer the potential effects of genetic variants.

In contrast, biological simulation and mechanistic disease research require models capable of predicting changes in biological states and the causal relationships connecting those changes.

For example, even when a variant is predicted to be pathogenic, its AlphaMissense score alone does not reveal whether the variant reduces protein activity, increases it, or alters interactions with other molecules.

Furthermore, how these functional changes affect cellular processes in particular tissues and ultimately lead to patient symptoms must be investigated separately.

Therefore, future research should move beyond predicting whether a variant is likely to be pathogenic toward understanding how its effects propagate across molecular, cellular, and organismal scales.

9.2 Implications for Precision Medicine, Drug Discovery, and Biological Simulation

In precision medicine, identifying disease-associated variants from a patient's genomic information is a major challenge.

AlphaMissense can provide valuable information for narrowing down large numbers of candidate variants to those more likely to be pathogenic.

However, determining whether a variant actually causes a particular patient's disease requires additional evidence, including patient phenotype, inheritance pattern, family history, and functional experiments.

In drug discovery, understanding the molecular functional changes and disease mechanisms caused by variants becomes even more important.

Identifying a therapeutic target requires more than predicting pathogenicity.

Researchers also need to understand which molecular functions should be modified to correct disease-associated abnormalities.

Biological simulation presents an even broader challenge.

It requires predicting how molecular functional changes caused by genetic variants affect cellular behavior, tissue states, and organism-level physiology.

Therefore, models such as AlphaMissense provide important foundational capabilities, but they do not by themselves simulate biological processes or predict the effects of therapeutic interventions.

9.3 A Fundamental Question for Evaluating Future AI Research

Future research should move beyond asking only how accurately a model can classify a variant as pathogenic.

A more fundamental question is:

When a variant is predicted to be pathogenic, which molecular function does it change, in what direction, and under what cellular conditions? Through which biological processes do these changes lead to a patient's symptoms? And can those processes be modified through therapeutic intervention?

The historical significance of AlphaMissense is not that it has answered these questions.

Rather, its achievement is that it has made large-scale prediction of variant effects possible, providing an important foundation for investigating them.

When evaluating future research, the most important distinction is not simply how accurate a model's predictions are, but how much the model has actually learned and validated about evolutionary constraints, molecular function, and the causal mechanisms of disease.

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] “Making sense of missense,” 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC11683568/.

[3] Experimental functional assessment study of IRF6 variants, 2024. (Research on IRF6 functional assessment discussed in [2].)

[4] Kong et al., “Discordance between a deep learning model and clinical-grade variant pathogenicity classification in a rare disease cohort,” npj Genomic Medicine, 2025. https://www.nature.com/articles/s41525-025-00480-w.

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

[6] Zhuo et al., “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.

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

[8] “Prediction of human missense variant effects from functional evidence,” Nature Genetics, 2026. https://www.nature.com/articles/s41588-026-02727-3.

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

[10] Research on gene- and domain-specific calibration of pathogenicity prediction scores, 2026. (Studies comparing universal genome-wide thresholds with gene- and domain-specific calibration.)