Gengram AI for Decoding DNA

Описание: AI Breakthrough in Genomics: "Plug-in Brain" Gengram Deciphers DNA's "Book of Life"

A Chinese research team, Genos, has made a significant leap in applying artificial intelligence to life sciences. Inspired by DeepSeek's recent "Engram" model—an external memory mechanism—the team developed "Gengram," a novel plug-in for genome analysis models.

Gengram addresses a key bottleneck: traditional genomic AI models process DNA sequences as individual letters (A,T,C,G), which is computationally inefficient and lacks biological context. Gengram introduces an external "genetic dictionary" that stores common, functional short DNA sequences (motifs). This allows the model to quickly reference prior biological knowledge instead of deducing everything from scratch.

The system features a dynamic gating mechanism, enabling the AI to intelligently decide when to consult this memory bank, focusing on crucial functional regions while ignoring non-coding areas.

Remarkably, during testing, the optimal performance was achieved with a 21-base-pair (bp) window. This aligns perfectly with DNA's double-helix structure (a full turn every ~10.5 bp), suggesting Gengram helps the AI implicitly understand the three-dimensional spatial context of DNA from one-dimensional sequence data.

Despite its small size (~20M parameters), integrating Gengram into large genomic models has led to substantial performance gains. It has set new state-of-the-art records on several tasks, such as boosting the AUC for splice site identification by 16.1%, and enables models to achieve comparable results with far less training data.

The Genos team behind Gengram represents a powerful collaboration in Hangzhou, combining the omics data expertise of BGI Research with the computational prowess of Zhejiang Lab. This development marks a major step toward truly deciphering the complex "book of life."
https://github.com/BGI-HangzhouAI/Gen...
https://github.com/BGI-HangzhouAI/Gen...
https://huggingface.co/BGI-HangzhouAI...