EnTao-GPM

Separating pathogenic mutations from benign polymorphisms is the rate-limiting step in clinical genome interpretation. Rule-based scores such as SIFT and PolyPhen-2 lean on amino acid properties and conservation, which leaves them weak on non-coding and rare variants, while ensemble predictors like REVEL depend on hand-engineered features. DNA foundation models changed the framing by learning pathogenicity signal end-to-end from raw sequence, with Evo 2 demonstrating zero-shot variant effect prediction across coding and non-coding regions and DNABERT-2 showing that contextual sequence embeddings improve single-nucleotide variant classification.

EnTao-GPM (Enlightened Tao Germline Pathogenic Mutations) is a July 2025 arXiv preprint from Fudan University and BioMap Research that specializes this machinery for inherited disease. It follows a three-stage recipe: start from the TrinityDNA backbone, continue self-supervised pretraining on the genomes of 27 laboratory mammals to produce an intermediate model the authors call TrinityDNA-LabFauna, then fine-tune that model on clinically annotated variants from ClinVar and the Human Gene Mutation Database (HGMD) to emit a calibrated probability that a given mutation is disease-causing.

The design bet is narrowing rather than broadening. Where Evo and Evo 2 scale pretraining across all domains of life, EnTao-GPM restricts its corpus to the mammals actually used as human disease models, on the argument that their conservation patterns map more directly onto human pathogenicity — especially in non-coding sequence, where most of the interpretive burden falls.

#Key Features

• Mammal-targeted cross-species pretraining: Continued pretraining spans 27 mammalian genomes — human, mouse, rat, macaque, pig, dog and other laboratory species — assembled from NCBI with matching Ensembl cDNA, totaling roughly 80 billion nucleotides.
• Bio-inspired backbone: The TrinityDNA architecture contributes a Groove Fusion module (parallel convolutions of kernel size 3, 5 and 7 encoding major and minor groove structure), sliding multi-window attention that gives each head a different window size, and a gated reverse-complement mechanism that processes both strands through a shared encoder.
• Two deployment variants: EnTao-GPM Fast scores SNVs from the reference sequence alone, supporting whole-genome mutation scanning; EnTao-GPM Pro takes both reference and mutant sequences and handles insertions, deletions and multi-point mutations.
• Clinically curated supervision: Fine-tuning uses 338,916 high-confidence ClinVar variants (April 2025 release) plus 266,044 non-overlapping HGMD disease-causing mutations, restricted to the DM class that shows 97.2% concordance with ClinVar pathogenic labels.
• Parameter-efficient specialization: Low-Rank Adaptation plus a feedforward classification head adapt the pretrained backbone without full-weight retraining.

#Technical Details

The backbone is a roughly 1-billion-parameter bidirectional transformer trained with a masked language modeling objective at a 100 kb context window, using Megatron and DeepSpeed with FlashAttention 2 and 4D parallelism on A100 GPUs in BF16. Zero-shot on a held-out ClinVar SNV test set, TrinityDNA-LabFauna reaches 0.943 AUROC and 0.961 AUPRC, above the same-size Evo 2 1B model (0.927 / 0.946) and close to Evo 2 40B (0.950 / 0.955) at a fraction of the parameters; it clears AIDO.DNA-7B and MambaDNA by a wide margin on the same split. After fine-tuning, EnTao-GPM Fast reaches 0.963 AUROC and 0.974 AUPRC on the 5,406-variant ClinVar SNV test set, and EnTao-GPM Pro reaches 0.933 AUROC and 0.966 AUPRC on the harder 10,987-variant set combining SNVs, non-SNVs and held-out HGMD mutations — a split where the zero-shot backbone drops to 0.704 AUROC. The paper describes an interpretability layer that pairs sequence embeddings with LLM-generated statistical explanations of a variant call; the reported benchmarks cover the pathogenicity classifier itself.

#Applications

The intended users are clinical geneticists and diagnostic laboratories triaging variants from exome and genome sequencing. Inference accepts VCF or CSV input and returns per-variant pathogenicity probabilities through either a Python library or a hosted server API, running on a GPU with at least 40 GB of memory. The Fast variant suits large-scale batch screening — the authors report a whole-genome mutation scan across GRCh38 — while Pro covers the indel and multi-point cases that reference-only scanning cannot express. Research uses include prioritizing candidate variants in rare-disease cohorts and scoring non-coding regions where missense predictors are silent.

#Impact

EnTao-GPM demonstrates that a domain-narrowed pretraining corpus can substitute for raw scale: a 1B-parameter mammalian model matches a 40B all-of-life model on ClinVar zero-shot, and supervised specialization on clinical labels adds several points on top. Several caveats bound the result. The work remains an arXiv preprint without peer review, and evaluation is confined to held-out ClinVar and HGMD splits rather than prospective clinical cohorts, so the reported margins over Evo 2 reflect a single curated test distribution. HGMD is a licensed commercial resource, which limits independent retraining. Model weights have not been publicly released; access runs through direct contact with the authors. A first-author repository published under an MIT license in June 2026 provides evaluation utilities — benchmark construction, metric computation with bootstrap confidence intervals, and Evo 2 baseline scoring — but not the model itself.