EnTao-GPM: DNA Foundation Model

• The paper introduces EnTao-GPM, a DNA foundation model that improves pathogenic germline mutation classification via cross-species pre-training and fine-tuning on ClinVar/HGMD.
• It leverages advanced transformer architecture with modules like GFM, SMWA, and GRC to capture conserved coding and non-coding genetic motifs.
• An interpretable clinical framework integrates DNA embeddings and LLM-generated ACMG-style evidence, facilitating actionable variant assessment.

EnTao-GPM, introduced in "EnTao-GPM: DNA Foundation Model for Predicting the Germline Pathogenic Mutations" (Lin et al., 29 Jul 2025), is a DNA foundation model for distinguishing pathogenic germline mutations from benign polymorphisms. Developed by Fudan University and BioMap, it combines three linked components: cross-species targeted pre-training on disease-relevant mammalian genomes; germline mutation specialization through fine-tuning on ClinVar and HGMD; and an interpretable clinical framework that integrates DNA sequence embeddings with LLM-based statistical explanations. The model is positioned for both SNVs and non-SNVs, with particular emphasis on conserved pathogenic motifs, including motifs in non-coding regions.

1. Conceptual scope and problem setting

EnTao-GPM is framed around a central precision-medicine problem: pathogenicity classification of germline variants. Its stated purpose is to improve discrimination between pathogenic mutations and benign polymorphisms in a setting where clinically relevant signal resides in both coding and non-coding sequence contexts (Lin et al., 29 Jul 2025).

The system is organized around three innovations given in the source description. First, it performs cross-species targeted pre-training on disease-relevant mammalian genomes, specifically including human, pig, and mouse, with the rationale that pathogenic motifs tend to be conserved across mammals, particularly in non-coding regulatory regions. Second, it is specialized for germline mutation prediction through fine-tuning on ClinVar and HGMD, with explicit support for both SNVs and non-SNVs. Third, it adds an interpretable clinical layer in which DNA-sequence-derived representations are combined with LLM-based statistical explanations intended to produce ACMG-style rationales.

This design places EnTao-GPM within the class of DNA foundation models that use large-scale sequence pre-training and downstream adaptation for biomedical inference. A plausible implication is that its novelty lies less in any single modeling primitive than in the coupling of mammalian comparative pre-training, germline-specific supervision, and an explicit explanation pipeline aimed at clinical interpretability.

2. Backbone architecture and cross-species pre-training

The backbone is adapted from TrinityDNA. In its "LabFauna" stage, it is post-trained on 27 high-quality mammalian genomes, including Homo sapiens, Sus scrofa domesticus (pig), Mus musculus (mouse), and 24 other laboratory mammals (Lin et al., 29 Jul 2025). The pre-training objective is a standard masked language-model objective applied over long contexts of 100 kb windows, with masking and reconstruction of bases. The stated effect is that the model "autonomously captures mammal-specific regularities, such as conserved sequence blocks and the distinction between coding and noncoding regions."

Input DNA is tokenized into overlapping $k$-mers, for example $k=6$, or into subword units. Each token $\text{token}_n$ is mapped to a learned $d$-dimensional embedding,

$$\mathbf{e}_n \;=\; E(\text{token}_n)\quad\in\mathbb{R}^d.$$

The embedding sequence $\{\mathbf{e}_1,\dots,\mathbf{e}_L\}$ is then processed by stacked transformer blocks. The overview also specifies standard self-attention with $Q,K,V\in\mathbb{R}^{\ell\times d_k}$.

Three TrinityDNA-derived modules are highlighted. The Groove Fusion Module (GFM) captures major/minor groove signals via multi-kernel convolutions. Sliding Multi-Window Attention (SMWA) assigns each head a distinct local/global window to alleviate oversmoothing and locality bias. Gated Reverse-Complement (GRC) processes forward and reverse-complement strands in parallel with a gating fusion. Taken together, these components define a sequence model that is explicitly adapted to genomic structure rather than treated as a generic token transformer.

The cross-species design is central rather than incidental. The source explanation ties the choice of mammalian genomes directly to pathogenic motif conservation, especially in non-coding regulatory regions. This suggests that the pre-training corpus is intended to encode evolutionary constraints that are later exploited during clinical pathogenicity prediction.

3. Germline specialization, datasets, and optimization

Fine-tuning is performed for germline mutation classification using curated data from ClinVar and HGMD (Lin et al., 29 Jul 2025). ClinVar uses the April 9, 2025 release; after curation to "pathogenic/likely pathogenic" versus "benign/likely benign," it contains approximately 300,000 SNVs and approximately 35,000 non-SNVs, for a total of 338,916 variants. HGMD Professional 2025.2 contributes 266,044 confirmed pathogenic variants from the DM category, with non-overlap relative to ClinVar. After curation, class balance is described as roughly $50\!:\!50$ benign versus pathogenic within each variant type.

The final training splits are given explicitly. Dataset 1, which is SNV only, has 297,634 training examples and 5,406 test examples. Dataset 2, which combines SNVs, non-SNVs, and HGMD, has 593,973 training examples and 10,987 test examples, including 5,406 ClinVar SNVs, 3,581 ClinVar non-SNVs, and 2,000 HGMD variants.

The optimization objective is standard binary cross-entropy over $N$ labeled variants:

$$\mathcal{L} \;=\; -\frac{1}{N} \sum_{i=1}^N \Bigl[y_i\,\log\hat y_i \;+\;(1-y_i)\,\log(1-\hat y_i)\Bigr].$$

Fine-tuning is carried out via Low-Rank Adaptation (LoRA) on the pre-trained weights for efficiency. The settings reported in the supplement are: AdamW with weight decay $k=6$0; learning rate $k=6$1 with linear warmup over the first $k=6$2 of steps followed by cosine decay; batch size $k=6$3 sequences per GPU; $k=6$4--$k=6$5 epochs with early stopping on validation AUC; and BF16 for forward passes with FP32 gradients.

Two downstream model variants are distinguished in the results summary. EnTao-GPMFast is SNV-only fine-tuned and is associated with whole-genome DMS. EnTao-GPMPro uses mixed fine-tuning and is intended for integrated SNV+indel handling. The distinction is functional: one variant emphasizes large-scale SNV scoring, while the other broadens variant-type coverage.

4. Interpretable clinical framework and ACMG-style evidence aggregation

After a variant is scored with pathogenic probability $k=6$6, EnTao-GPM passes three classes of information to a biomedical LLM such as GPT-4: sequence context embedding features, model confidence $k=6$7, and standard annotations such as conservation and allele frequency (Lin et al., 29 Jul 2025). The interface is prompt templated. The example prompt given is: "Given the following variant: [chr, position, ref→alt], model pathogenicity p=0.92, PhyloP=4.5, ClinVar count=3. Please provide an ACMG-style classification rationale, citing Codes (PVS1, PM2, etc.) and aggregate evidence."

The output is a structured explanation containing evidence codes and a final call. Evidence scores are computed by mapping model probability and feature heuristics to ACMG criteria weights, with examples including PVS1 as very strong and PM2 as moderate. The framework then forms a final "evidence sum" $k=6$8, and the description gives a representative thresholding rule: $k=6$9.

This explanation layer is clinically oriented rather than purely mechanistic. The framework is designed to produce ACMG-style statistical explanations that are actionable for variant assessment and counseling workflows. A plausible implication is that interpretability here is post hoc with respect to the sequence model, because the LLM receives model outputs and annotations after variant scoring rather than constituting the predictive core itself.

The recommended probability thresholds are also explicit. Pathogenic corresponds to $\text{token}_n$0; likely pathogenic to $\text{token}_n$1; variant of uncertain significance (VUS) to $\text{token}_n$2; likely benign to $\text{token}_n$3; and benign to $\text{token}_n$4. These cutoffs situate the model in a workflow where continuous scores are translated into clinically legible categorical calls.

5. Empirical performance and ablation structure

The reported evaluation centers on ClinVar SNV testing and a mixed-variant test set (Lin et al., 29 Jul 2025). On the ClinVar SNV test from Dataset 1 with $\text{token}_n$5, the zero-shot TrinityDNA-LabFauna baseline attains AUC $\text{token}_n$6 and AUPRC $\text{token}_n$7. EnTao-GPMFast reaches AUC $\text{token}_n$8 and AUPRC $\text{token}_n$9. EnTao-GPMPro reaches AUC $d$0 and AUPRC $d$1. At a threshold of $d$2, the derived metrics are approximately Accuracy $d$3, Precision $d$4, and Recall $d$5.

On the mixed test from Dataset 2 with $d$6, zero-shot performance is substantially lower, with AUC $d$7 and AUPRC $d$8, whereas EnTao-GPMPro reaches AUC $d$9 and AUPRC $$\mathbf{e}_n \;=\; E(\text{token}_n)\quad\in\mathbb{R}^d.$$0. This establishes that the mixed fine-tuning stage is not merely incremental for broader variant-type coverage; it materially changes performance in the mixed setting.

Evaluation setting	AUC	AUPRC
Zero-shot on ClinVar SNV test (Dataset 1)	0.943	0.961
EnTao-GPMFast on ClinVar SNV test (Dataset 1)	0.963	0.974
EnTao-GPMPro on ClinVar SNV test (Dataset 1)	0.960	0.970
Zero-shot on mixed test (Dataset 2)	0.704	0.885
EnTao-GPMPro on mixed test (Dataset 2)	0.933	0.966

The ablation studies isolate two design choices. Removing the cross-species LabFauna stage and replacing it with human-only pre-training reduces zero-shot ClinVar SNV AUC by approximately $$\mathbf{e}_n \;=\; E(\text{token}_n)\quad\in\mathbb{R}^d.$$1, from $$\mathbf{e}_n \;=\; E(\text{token}_n)\quad\in\mathbb{R}^d.$$2 to approximately $$\mathbf{e}_n \;=\; E(\text{token}_n)\quad\in\mathbb{R}^d.$$3. Fine-tuning without LoRA, using full-weight updates, yields identical AUC but requires $$\mathbf{e}_n \;=\; E(\text{token}_n)\quad\in\mathbb{R}^d.$$4 higher GPU memory usage. The first result suggests that mammalian comparative pre-training contributes measurable signal beyond human-only pre-training; the second indicates that LoRA is an efficiency choice rather than an accuracy tradeoff within the reported setup.

The summary description also states that EnTao-GPM "outperforms existing SNV-only models in accuracy (AUC↑2–3%), handles indels and complex mutations, and delivers actionable, standards-compliant explanations." Because the detailed comparator list is not reproduced in the supplied data, that statement is best interpreted as a high-level positioning claim rather than a fully enumerated leaderboard.

6. Clinical deployment targets, use cases, and interpretive cautions

Three use cases are identified. First, variant assessment: EnTao-GPMFast is described as enabling rapid scoring of all possible SNVs genome-wide through whole-genome DMS (Lin et al., 29 Jul 2025). Second, rare-variant risk identification: EnTao-GPMPro provides integrated SNV+indel handling for familial studies. Third, personalized treatment planning: the LLM explanation layer maps outputs to actionable ACMG codes intended to guide genetic counselors.

The practical significance of these use cases derives from the model’s joint handling of predictive accuracy and explanation formatting. In the source framing, this combination supports clinical diagnostics through variant assessment, risk identification, and personalized treatment. The same architecture is also presented as useful for research settings in which genome-scale prioritization and non-coding interpretation are important.

A recurring misconception in this area is to equate explanation output with direct mechanistic proof. The supplied description supports a narrower reading. EnTao-GPM produces a pathogenicity probability from a DNA foundation model, then synthesizes ACMG-style evidence using an LLM and heuristic mappings. Accordingly, the explanation framework should be understood as a structured statistical rationale rather than as direct experimental validation. This does not diminish its workflow utility, but it does clarify what kind of interpretability is being claimed.

A second interpretive caution concerns generalization. The reported gains are tied to a specific pipeline: cross-species mammalian pre-training, germline-focused fine-tuning on ClinVar and HGMD, and post hoc ACMG-style explanation. This suggests that the model’s clinical value depends on the interaction of these components rather than on pre-training alone. Within that framing, EnTao-GPM is best understood as a specialized germline pathogenicity system built on a DNA foundation-model backbone, with explicit accommodation of non-SNVs and clinically formatted evidence synthesis.