Prediction Models in MERGE

MERGE integrates 9 state-of-the-art machine learning models to predict the pathogenicity of genetic variants. Each model brings unique strengths in analyzing different aspects of genomic data.

DANN

DANN (Deleterious Annotation of genetic variants using Neural Networks) uses deep neural networks to predict the functional impact of genetic variants. It combines information from multiple genomic features including conservation scores, functional annotations, and protein-level information.

Type
Deep Neural Network
Primary Feature
Multi-feature Integration

MetaRNN

MetaRNN is a meta-predictor based on recurrent neural networks that integrates predictions from multiple variant effect prediction tools. It leverages sequence context and evolutionary information to provide accurate pathogenicity predictions.

Type
Recurrent Neural Network
Primary Feature
Meta-learning & Ensemble

MVP

MVP (Missense Variant Pathogenicity) prediction model uses deep residual networks to predict the functional impact of missense variants. It integrates sequence conservation, protein structure, and functional annotations.

Type
Deep Residual Network
Primary Feature
Structure & Conservation

gMVP

gMVP (genome-wide Missense Variant Pathogenicity) extends MVP to genome-wide predictions. It incorporates additional genomic context and regulatory information to improve prediction accuracy across the entire genome.

Type
Genome-wide Deep Learning
Primary Feature
Regulatory Context

PrimateAI

PrimateAI leverages deep learning on primate genomic data to predict the pathogenicity of missense variants. It uses evolutionary information from six primate species to identify functionally important positions in the human genome.

Type
Deep Convolutional Network
Primary Feature
Primate Evolution

ESM1b

ESM1b (Evolutionary Scale Modeling) is a large-scale protein language model trained on millions of protein sequences. It captures evolutionary patterns and protein structure information to predict variant effects without explicit structural data.

Type
Protein Language Model
Primary Feature
Sequence Representation

AlphaMissense

AlphaMissense is built on AlphaFold's protein structure prediction capabilities. It combines structural insights with evolutionary information to predict the pathogenicity of missense variants with high accuracy across the entire human proteome.

Type
Structure-based Deep Learning
Primary Feature
3D Protein Structure

MutFormer

MutFormer uses transformer architecture to model the effects of mutations on protein function. It learns contextualized representations of protein sequences and predicts how mutations affect protein stability and function.

Type
Transformer Model
Primary Feature
Attention Mechanism

MisFit

MisFit (Missense Fitness) predicts variant pathogenicity by modeling the fitness effects of mutations. It integrates evolutionary constraints, protein structural information, and functional data to assess the impact of missense variants.

Type
Fitness-based Model
Primary Feature
Functional Fitness