Prediction Models in MERGE

MERGE combines 14 model features, drawn from the 10 models below, into one ensemble probability. Each of the four variant panels (coding, splice, non-coding and non-SNV) uses its own subset of these features, so a given model does not contribute to every prediction. Any feature that cannot be computed for a variant is median-imputed, and every prediction reports which features were computed and which were imputed.

Not for clinical use. MERGE is a research prototype, not a medical device, and must not be used for diagnosis or any other clinical purpose. The MERGE software itself is MIT-licensed, but the models below are third-party works, each under its own licence — AlphaGenome, Nucleotide Transformer v1/v2, dbNSFP and ANNOVAR restrict commercial use — and obtaining them and complying with their terms is the user's responsibility. Full terms of use ↗

Genomic foundation models (DNA sequence)

AlphaGenome

Read Paper ↗

AlphaGenome is an advanced deep learning framework designed to model whole-genome contexts and predict the functional impact of genetic variants across the entire human genome. It leverages deep evolutionary constraints to map sequence variations to potential clinical outcomes.

AlphaGenome runs on Google DeepMind's servers rather than ours, and is updated there over time. It is the one model in MERGE whose scores are not reproducible indefinitely; every other model reproduces bit-for-bit.
Type
Genomic Foundation Model
Primary Feature
Regulatory Variant Effects
Ensemble Features
1

Enformer

Read Paper ↗

Enformer predicts gene expression and chromatin state directly from sequence, using transformer attention to carry information across a 393,216 bp receptive field. This lets it capture the effect of distal enhancers and other long-range regulatory elements that convolutional models truncate.

Type
Long-Range Regulatory Model
Primary Feature
Expression and Chromatin Effects
Ensemble Features
1

Evo 2 is a foundation model that directly links DNA sequence to function across vast evolutionary timescales. Trained across domains of life, it uncovers deep, cross-species genomic principles to robustly predict variant effects. MERGE scores variants with the difference in mean per-token log-likelihood between the reference and alternate sequences.

Three Evo 2 variants are separate ensemble features: evo2_7b, evo2_7b_base and evo2_1b_base. On this deployment evo2_1b_base is not computed — it requires a GPU with FP8 support (compute capability 8.9 or higher) — so its feature is median-imputed from the training distribution.
Type
Cross-species Foundation Model
Primary Feature
Deep Evolutionary Principles
Ensemble Features
3

Carbon-3B

Model Card ↗

Carbon-3B is a 3-billion-parameter genomic foundation model for eukaryotic DNA and RNA, built on a hybrid tokenizer that encodes DNA in fixed 6-nucleotide blocks. Its very large native context makes it well suited to scoring insertions, deletions and other structural edits, where the altered sequence must be read as a whole.

Carbon uses the same summed log-likelihood-delta scoring as during training, but over an 8,192 bp context window instead of the 24,576 bp used to fit the ensemble — the larger window exceeds the memory of the deployment GPU. Scores are therefore on a slightly different scale from the published feature tables.
Type
Long-Context Foundation Model
Primary Feature
Sequence Likelihood Shift
Ensemble Features
1

GENERator & GENERator-v2

Read Paper ↗

GENERator is a long-context generative genomic foundation model trained on 386 billion nucleotides of eukaryotic DNA. MERGE uses the 1.2B eukaryote models and scores a variant by the change in negative log-likelihood the model assigns to the sequence when the reference allele is replaced by the alternate.

Type
Generative Genomic Model
Primary Feature
Sequence Likelihood Shift
Ensemble Features
2

Nucleotide Transformer (NT & NT-v2)

Read Paper ↗

The Nucleotide Transformer is pre-trained on a vast, diverse collection of DNA sequences and learns the fundamental language of genomes. MERGE uses two of its models — the 500M human-reference model and the 500M multi-species v2 model — as masked language models, scoring a variant by the Euclidean distance between the reference and alternate sequence embeddings.

Type
Masked DNA Language Model
Primary Feature
Embedding Shift
Ensemble Features
2

HyenaDNA

Read Paper ↗

HyenaDNA is a long-range genomic foundation model capable of processing contexts of up to 1 million base pairs at single-nucleotide resolution. It uses the Hyena operator to capture distant regulatory interactions and variant effects that traditional models miss.

Type
Long-Context Foundation Model
Primary Feature
Long-Range Interactions
Ensemble Features
1

Protein-level and conservation models

AlphaMissense

Read Paper ↗

AlphaMissense is built on AlphaFold's protein structure prediction capabilities. It combines deep 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
Ensemble Features
1

ESM-1b (Evolutionary Scale Modeling) is a large-scale protein language model trained on millions of protein sequences. It captures evolutionary patterns to predict how mutations affect protein function without requiring explicit structural data.

Type
Protein Language Model
Primary Feature
Protein Representation
Ensemble Features
1

GPN-MSA

Read Paper ↗

GPN-MSA (Genomic Pre-trained Network with Multiple Sequence Alignments) integrates deep learning representations with whole-genome alignments. It effectively maps evolutionary constraints to predict pathogenic variations in both coding and regulatory regions.

Type
Genomic Pre-trained Network
Primary Feature
Multiple Sequence Alignments
Ensemble Features
1