About MERGE Project

Project Overview

The MERGE (Model Ensemble for Ranking Genomic Effects) framework is a pipeline for research on variant effects. It combines genomic and protein model features in classifiers trained on labelled benchmark data. Chromosome-held-out evaluation reduces overlap between training and evaluation data, but does not establish independence from all upstream training data. MERGE routes each variant to a dedicated panel rather than scoring everything with one general-purpose model.

There are four panels — coding, splice, non-coding and non-SNV (indel / MNV / delins) — each with its own feature set and its own classifier, trained separately. The panel is chosen automatically: a variant whose REF or ALT is not a single base goes to the non-SNV panel; otherwise ANNOVAR annotation decides between coding, splice and non-coding.

The panels use selected features from genomic and protein models: Enformer, Evo2 (7B, 7B-base and 1B-base), Carbon-3B, GENERATOR and GENERATOR-v2, Nucleotide Transformer, HyenaDNA, AlphaMissense, ESM-1b and GPN-MSA. Any feature that cannot be computed for a given variant is median-imputed by the ensemble, and every prediction reports which features were computed and which were imputed. Enformer supplies seven summaries across all 5,313 human tracks; feature counts refer to scalar inputs, not model families. NT-v2 scores can be displayed separately but are not selected by the current panels. AlphaGenome is excluded from ensemble training.

For splicing and non-coding variants only, an optional MERGE + AlphaGenome average combines the MERGE score and a label-free AlphaGenome percentile with equal weights. The percentile is a lookup against benchmark scores, not a fitted classifier or a calibrated pathogenicity probability. Without an AlphaGenome value, this average is unavailable. Labels such as Pathogenic and Benign describe score bands; they are not ACMG/AMP classifications or estimates of an individual's disease risk.

Benchmark performance

Four fixed datasets were evaluated with three chromosome-held-out folds. The table reports unweighted fold-mean AUROC / AUPRC; AUPRC denotes average precision. These are benchmark estimates, rather than performance measured on the models' full-data training fits.

The AlphaGenome component is the splicing score for splicing variants and the raw maximum for noncoding variants. The noncoding quantile maximum is a separate score (AUROC 0.9774; AUPRC 0.6855). The fixed average has the highest mean noncoding AUROC, while the AlphaGenome raw maximum has the highest mean noncoding AUPRC.

All primary splicing comparisons use the same 35,015 variants, including 571 missing AlphaGenome scores imputed from each training fold before percentile conversion. Online predictions offer the separate average only when an AlphaGenome value is available. Online feature availability and scoring windows can differ from the benchmark; see the model descriptions for current inference settings.

Independent BRCA1 functional validation

The frozen MERGE-Coding model was evaluated on 1,421 jointly scored missense variants, including 263 loss-of-function variants, from published BRCA1 saturation genome editing data. It had the highest observed AUROC among 43 predictors: 0.9105 (95% CI 0.8899–0.9286), with AUPRC 0.7158 (95% CI 0.6632–0.7675). The paired difference from VEST4 was not statistically resolved. These functional labels were not used to train MERGE. BRCA1 SGE data source.

Command-line Tool (merge-cli)

merge-cli provides scriptable access to MERGE and supports the same four-panel design as this website. Model bundles may differ between releases; record the CLI and model versions when comparing results. It is published on PyPI. Its software is licensed under the MIT licence — the source carries no field-of-use restriction. The four current ensemble model files are also MIT-licensed; the AlphaGenome reference percentile table is CC BY-NC 4.0. See the terms above.

It covers all four variant panels — coding, splice, non-coding and non-SNV (indel / MNV / delins) — and outputs from several genomic and protein models (AlphaGenome, Enformer, Evo2-7B / 7B-base / 1B-base, Carbon-3B, GENERATOR, GENERATOR-v2, Nucleotide Transformer, NT-v2, HyenaDNA, AlphaMissense, ESM-1b, GPN-MSA).

1. Install

Requires Python 3.11 or newer.

2. Remote mode (default) — nothing to configure

Scores are computed by this server; the MERGE ensemble runs locally on your CPU. The variant panel (coding / splice / non-coding / non-SNV) is chosen automatically.

3. Local mode (optional, GPU)

Local mode can run supported models on your own hardware. Each model is deployed into its own conda environment and served over HTTP; deploy the ones you need. AlphaGenome uses an external API and requires your own key. Where data are processed depends on your configured services.

Use scoring windows and conventions that match the ensemble training procedure. Hardware, software versions and missing features can change results. Missing ensemble inputs are median-imputed, and the CLI reports which features were computed and which were imputed.

4. Use it from an AI assistant (MCP)

merge-cli ships an MCP server, so assistants such as Claude Desktop, Claude Code and Codex can call MERGE directly. Both steps are needed: the first registers the tools, the second makes the assistant reach for them.

The MCP server always runs in remote mode, so no GPU is required. Full documentation ships with the package (TUTORIAL.md) and on the PyPI project page.

The Research Team