References

The following references were used in the development of this tool:

SIFT: SIFT
PolyPhen-2: PolyPhen-2
AlphaMissense: AlphaMissense
MutationTaster: MutationTaster
MutationAssessor: MutationAssessor
PROVEAN: PROVEAN
REVEL: REVEL
MetaSVM: MetaSVM
MetaLR: MetaLR
M-CAP: M-CAP
CADD: CADD
DANN: DANN
fathmm-XF: fathmm-XF

SIFT (Sorting Intolerant From Tolerant) es un predictor de efecto de variantes no supervisado que utiliza alineamientos múltiples de secuencias para evaluar el impacto funcional de sustituciones de aminoácidos. Parte del principio de que las posiciones conservadas a lo largo de la evolución son funcionalmente importantes, por lo que cambios en estas posiciones tienen mayor probabilidad de ser perjudiciales.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
SIFT - Sorting Intolerant From Tolerant	Empirical	Unsupervised	Multiple sequence alignment	No	< 0.05	Multiple sequence alignment only	View	View

PolyPhen-2 (Polymorphism Phenotyping v2) is a supervised variant effect predictor based on a Naive Bayes classifier trained with structural and evolutionary conservation features. It evaluates whether an amino acid substitution may negatively affect protein structure and function. PolyPhen-2 provides two distinct models: HDIV (optimized for sensitivity) and HVAR (optimized for clinical specificity).

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
PolyPhen-2 - Polymorphism Phenotyping v2	Naive Bayes classifier	Supervised	Structural and evolutionary features	No	> 0.5	HumDiv / HumVar	View	View

AlphaMissense is an advanced predictor that combines supervised and unsupervised approaches using a custom masked language model designed to detect the functional effects of missense variants. Developed by DeepMind, AlphaMissense estimates variant pathogenicity based on protein sequence data and inferred structural knowledge, providing broad and up-to-date coverage of the human proteome.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
AlphaMissense	Custom masked language model	Unsupervised / Supervised	Protein sequence and inferred structure	No	> 0.5	Observed and unobserved variants in humans and primates	View	View

MutationTaster is a supervised variant effect predictor that uses a Random Forest model to classify genetic variants as pathogenic or benign. It incorporates a wide range of biological features, including DNA sequence, splicing effects, conservation data, and functional annotations. It is particularly useful for predicting the effects of non-coding and splicing variants.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
MutationTaster2021	Random forest	Supervised	Sequence, conservation, splicing, annotation	No	Classification output only	gnomAD, HGMD, ClinVar	View	View

MutationAssessor is an unsupervised variant effect predictor based on evolutionary conservation. It estimates the functional impact of amino acid substitutions by analyzing how conserved the mutated position is across multiple sequence alignments. This method is especially useful for identifying functional mutations in cancer-related genes and other disease-associated loci.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
MutationAssessor	Combinatorial entropy optimisation	Unsupervised	Multiple sequence alignment	No	Medium-High: >3.5 Low-Medium: >1.935 None-Low: >0.8	Multiple sequence alignment only	View	View

PROVEAN (Protein Variation Effect Analyzer) is an unsupervised predictor that evaluates the functional impact of amino acid substitutions and small indels using homologous sequence alignments. It compares the similarity scores between reference and variant sequences to assess the evolutionary conservation. PROVEAN is suitable for analyzing both coding and non-coding variants.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
PROVEAN - Protein Variation Effect Analyzer	Empirical	Unsupervised	Homologous sequence alignment	No	< -2.5	Multiple sequence alignment only	View	View

REVEL (Rare Exome Variant Ensemble Learner) is a supervised meta-predictor based on a Random Forest algorithm that integrates outputs from multiple tools such as SIFT, PolyPhen-2, MutationTaster, and MutationAssessor. It is designed to classify rare missense variants more accurately by leveraging the strengths of each individual method. REVEL is widely used due to its robust performance and high precision in variant pathogenicity prediction.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
REVEL - Rare Exome Variant Ensemble Learner	Random Forest	Supervised meta-predictor	Ensemble of multiple tools	Yes	> 0.5	HGMD, ESP, ARIC, 1000 Genomes	View	View

MetaSVM is a supervised meta-predictor that uses a Support Vector Machine (SVM) model to combine outputs from various individual variant effect predictors. It aims to improve the classification accuracy for missense variants by leveraging the strengths of multiple features and tools. MetaSVM is often used alongside MetaLR, both trained on the same input features but using different machine learning algorithms.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
MetaSVM & MetaLR	Support Vector Machine / Logistic Regression	Supervised meta-predictor	SIFT, PolyPhen-2, LRT, MutationAssessor, etc.	Yes	MetaSVM > 0	UniProt	View	View

MetaLR is a supervised meta-predictor that uses logistic regression to combine the outputs of several individual variant effect predictors. It is trained using the same features and data as MetaSVM, offering a complementary approach to variant classification. While MetaSVM uses an SVM model, MetaLR provides a probabilistic score via logistic regression, helping to improve the robustness of missense variant interpretation.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
MetaSVM & MetaLR	Support Vector Machine / Logistic Regression	Supervised meta-predictor	SIFT, PolyPhen-2, LRT, MutationAssessor, etc.	Yes	MetaLR > 0.5	UniProt	View	View

M-CAP (Mendelian Clinically Applicable Pathogenicity Score) is a supervised meta-predictor based on gradient boosted trees, designed to classify Mendelian variants with high sensitivity. It integrates outputs from multiple tools and conservation metrics to predict pathogenicity while minimizing the misclassification of benign variants. M-CAP is especially useful in clinical genetics to reduce the number of variants of uncertain significance (VUS) in exome analyses.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
M-CAP - Mendelian Clinically Applicable Pathogenicity Score	Gradient boosted trees	Supervised meta-predictor	Various predictors + conservation features	Yes	> 0.025	HGMD, ExAC	View	View

CADD (Combined Annotation Dependent Depletion) is a supervised meta-predictor that integrates various genomic annotations into a single score estimating the deleteriousness of genetic variants. It uses logistic regression trained on simulated (neutral) and observed (potentially pathogenic) variants to distinguish between benign and pathogenic changes. CADD provides both raw and scaled (PHRED-like) scores and is widely used in genome-wide variant prioritization.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
CADD - Combined Annotation Dependent Depletion	Logistic regression	Supervised meta-predictor	Multiple genomic annotations	Yes	PHRED >10 or >20	Simulated variants vs. real variants from healthy individuals	View	View

DANN (Deleterious Annotation of genetic variants using Neural Networks) is a deep learning-based meta-predictor that uses a deep neural network to estimate the functional impact of genetic variants across the genome. It shares the same input features and training data as CADD, but replaces logistic regression with a neural network model. DANN captures more complex patterns in genomic data, making it a powerful alternative for predicting variant pathogenicity.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
DANN	Deep Neural Network	Supervised meta-predictor	Same as CADD	Yes	Not defined	Simulated variants vs. real variants from healthy individuals	View	View

fathmm-XF is a variant effect predictor based on evolutionary conservation and functional annotations. It uses a comprehensive set of features including sequence conservation, regulatory motifs, and epigenetic signals to estimate the pathogenicity of both coding and non-coding variants. The tool is designed to achieve high precision while minimizing false positives, making it suitable for clinical variant prioritization.

Predictor	Methodology	Type	Feature Categories	Nucleotide Level Only	Recommended Threshold	Training Data	Notes	Reference
fathmm-XF	Ensemble learning (random forest + SVM)	Conservation / structure-based	Conservation, regulatory elements, epigenomic features	Yes	> 0.5 (for pathogenicity)	Pathogenic and benign variants from ClinVar and ENCODE annotations	View	View

What is a Variant Effect Predictor?

A Variant Effect Predictor (VEP) is a bioinformatics tool designed to analyze and predict the functional impact of genetic variants in the human genome. These variants may include single nucleotide substitutions, insertions, deletions, or other genomic changes. By identifying how such mutations affect genes, transcripts, and proteins, VEPs are essential for understanding the potential clinical implications of variants found in genetic studies.

Types of Predictors and Included Tools

Type	Included Predictors
Supervised (Classical ML)	PolyPhen-2 (HDIV, HVAR), MetaSVM, MetaLR, M-CAP, MutationTaster
Unsupervised / Empirical	SIFT, PROVEAN
Conservation / Structure-based	MutationAssessor, fathmm-XF
Deep Learning	DANN, AlphaMissense
Metapredictors	REVEL, CADD

Supervised methods use machine learning algorithms trained on labeled datasets of variants classified as benign or pathogenic. Once trained, these models can generalize to new variants and predict their likely impact. Examples include PolyPhen-2, MetaSVM, MetaLR, M-CAP, and MutationTaster.

Unsupervised methods do not rely on prior labels but instead use evolutionary conservation or physicochemical properties to infer impact. SIFT and PROVEAN follow this approach.

Structure and conservation-based predictors analyze how conserved a region is across species and whether a variant disrupts structural stability or key regions of the protein. MutationAssessor and fathmm-XF are part of this category.

Deep learning predictors like AlphaMissense and DANN use neural networks to detect complex patterns in biological data, improving prediction accuracy.

Metapredictors combine the output of multiple tools to provide a more reliable prediction. REVEL and CADD are widely used examples that aggregate scores from many underlying methods.