Jindu Life Science published GeneLLM, a multi-omics foundation model trained directly on raw sequencing data, in Nature Communications and Advanced Science this month. The release shifts biological modeling away from human-curated annotations toward systems that interpret unprocessed transcriptome signals, cutting precision medicine sequencing costs by 83 percent while maintaining high predictive accuracy.
How the model processes biological language
GeneLLM operates on the same principle as large language models: predicting the next token. Instead of letters or words, the system predicts adenine, uracil, guanine, and cytosine in RNA sequences. Researchers divided RNA fragments into seven-base windows and fed tens of trillions of reads through hundreds of NVIDIA A100 GPUs during pre-training. This approach bypasses traditional gene annotation pipelines, allowing the architecture to surface hidden disease patterns that manual labeling often misses.
Connecting artificial intelligence to wet laboratories
Model accuracy alone does not accelerate discovery. Jindu paired GeneLLM with the BioFord system, a physical automation layer designed to run experiments inside existing laboratory environments. The system translates research protocols into machine-executable commands, manages device scheduling across heterogeneous instruments, and logs every operational parameter. Five specialized AI agents now handle literature review, experimental design, protocol optimization, instrument coordination, and result analysis. Failed runs feed directly back into the training loop, turning negative outcomes into actionable data rather than discarded notes.
Building infrastructure over foundational parameters
The startup completed four funding rounds in 2025, securing nearly RMB 100 million in a Series A led by Gaotai Investment alongside earlier backing from Sequoia Capital China Seed Fund, Chuangdongfang Investment, and Nanshan Strategic Emerging Industries Investment. CEO Kim Young-sung said, "During our R&D process, we gradually realized that AI for Bioscience is not simply about stacking models and data. For those conducting experiments, the ultimate challenge remains solving the dilemma of getting results that can't be executed or that are incorrect even after execution." Rather than competing in foundational model training or building capital-intensive robotic facilities, the company targets the integration gap between software and physical bench work.
Why this matters for science and research professionals
Laboratory workflows currently stall when digital predictions meet physical constraints. Teams that adopt automated execution layers can compress experiment design cycles from months to weeks while maintaining auditable data trails. Researchers should evaluate whether their current instrumentation supports programmable control and standardized logging before investing in proprietary end-to-end platforms. For teams integrating these tools, following the AI Learning Path for Research Scientists provides practical frameworks for bridging computational outputs with wet-lab validation. The field will reward groups that treat failed trials as structured data rather than lost time.
Your membership also unlocks:
