MERGE uses selected genomic and protein model features in four separately trained classifiers (coding, splicing, non-coding and non-SNV). Enformer contributes seven summaries across 5,313 human tracks. A model does not contribute to every panel; NT-v2 is displayed separately and is not selected by the current panels. Missing ensemble inputs are median-imputed and listed with each result. AlphaGenome is excluded from ensemble training and contributes only to the optional separate average for splicing and non-coding variants.
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 ↗
AlphaGenome predicts functional genomic signals from DNA sequence, including gene expression, chromatin accessibility and RNA splicing. MERGE displays summaries of its predicted variant effects. These molecular predictions do not directly establish clinical outcomes.
Enformer predicts gene expression and other functional genomic tracks from DNA sequence using convolution and transformer attention. The model uses an effective sequence context of 196,608 bp; the original interface accepts a longer input and crops it internally. MERGE summarizes predicted REF-to-ALT effects across all 5,313 human output tracks.
Evo 2 is an autoregressive genomic model trained across all domains of life. MERGE evaluates the 7B, 7B-base and 1B-base checkpoints as separate features. Each feature is the mean per-token log-likelihood of the reference sequence minus that of the alternate sequence, using complete haplotypes for non-SNVs.
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. MERGE uses REF-to-ALT sequence likelihood differences to score supported SNVs and non-SNVs; its deployment window is shorter than the native model context.
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.
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.
HyenaDNA is a genomic language model using the Hyena operator for long sequence contexts at single-nucleotide resolution. Context capacity depends on the checkpoint; the family includes models supporting up to one million bases. This capacity should not be read as the sequence window used by every MERGE request.
AlphaMissense predicts effects of missense substitutions using protein sequence and structural context. MERGE retrieves precomputed AlphaMissense scores through dbNSFP and uses them in the coding panel; variants without a matched score have this input median-imputed.
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.
GPN-MSA (Genomic Pre-trained Network with Multiple Sequence Alignments) uses multispecies whole-genome alignments to estimate nucleotide-level evolutionary constraint. MERGE matches official precomputed scores to coding, splicing and noncoding SNVs by chromosome, position, reference and alternate allele. GPN-MSA is not an input to the non-SNV panel.