A gene exists in several information worlds at once. Its sequence can be measured, its expression can be counted, its proteins can be placed in networks—and decades of papers describe what researchers have observed about it. Many gene models use the first three and treat the literature as background. LitGene makes the text part of the representation.[1]
The model starts from biomedical language and refines gene representations by aligning unstructured text with Gene Ontology annotations through contrastive learning. In simple terms, it learns that a passage and an ontology concept should be close when they describe related biology, while keeping unrelated descriptions apart.[1]
Two evidence systems constrain one representation
What LitGene combines
| Input | What it contributes | Characteristic limitation |
|---|---|---|
| Scientific literature | Descriptions of experiments, functions, diseases and relationships | Popular genes are described far more often than neglected onesSource 1 |
| Gene Ontology | Structured hierarchy of biological processes, molecular functions and cellular components | Annotations are incomplete and inherit past research choicesSource 1 |
| Contrastive learning | A shared space linking textual and structured meanings | Alignment quality depends on the training pairs and objectiveSource 1 |
| SHAP explanations | Words that most influenced a prediction | Importance inside the model is not a biological mechanismSource 1 |
That hybrid design is intended to reduce a familiar bias. A text-only system can know a great deal about genes that attract funding and very little about genes with sparse literature. Ontology structure can connect an under-studied gene to broader functional categories, although it cannot create experimental evidence that does not exist.
From benchmark tasks to a disease candidate
Reported evaluation and outputs
| Measure | Value | Context |
|---|---|---|
| Protein-property benchmarks | Eight tasks | LitGene achieved competitive predictive accuracy.Source 1 |
| Zero-shot outputs | GO terms, KEGG pathways and gene–disease associations | Predictions were made for relationships not directly trained as labelled task examples.Source 1 |
| IPF candidate | ANGPTL6 | Proposed as a potential gene associated with idiopathic pulmonary fibrosis.Source 1 |
| Independent support | Differential expression in IPF samples | Supports relevance but does not establish causation or therapeutic benefit.Source 1 |
ANGPTL6 is the paper’s demonstration of hypothesis generation. The model identified it as a possible candidate for idiopathic pulmonary fibrosis, and the researchers found differential expression in independent IPF samples. That is a useful narrowing of the search space. A gene can be differentially expressed because it drives disease, responds to damage, marks a cell population or travels with another process.
Interpretability should lead back to evidence
LitGene includes a SHAP-based tool that highlights words influencing a prediction and an automated literature-grounding system that seeks corroborating sources. Those features can help a scientist inspect why the model connected a gene and disease. They should not turn the model’s explanation into circular evidence: a word was influential because it occurred in the literature used to train or ground the system.
The model’s open-source code and public interface also make its claims more inspectable. Researchers can submit genes, diseases, drugs or text and examine the relationships produced. The strongest contribution is not that a transformer has “understood” every gene. It is that literature and ontology can be placed in one interrogable map—making both well-known connections and conspicuous gaps easier to see.
Sources & further reading
- LitGene: an interpretable transformer model integrating text and ontology for gene representation learning Nature Communications · 16 Oct 2026 · Accessed 16 Oct 2026