Platinum Bio Inc. (Location: Higashihiroshima City, Hiroshima Prefecture, CEO: Keisuke Okuhara, hereinafter 'Our Company') is pleased to announce that the protocol paper "Workflow for Fine-Tuning and Evaluating DNA Language Models for Specific Genomics Issues," which summarizes a protocol for creating AI to address specific genomic challenges using DNA language models, was published online on April 3, 2026, in the life science protocol journal "Bio-protocol" (ISSN: 2331-8325) (peer-reviewed). This paper was authored within the Bio DX Research Lab of the Platinum Bio Joint Research Laboratory established at Hiroshima University, and authors include Kazuya Nakamae, Team Leader of the Bio DX Team in Our Company's Research and Development Department, and Hidemasa Bono, Chief Scientific Officer (CSO).
A DNA language model is a type of artificial intelligence (AI) that learns from a large number of DNA sequences composed of A, T, G, and C, much like AI learns the rules of language. In this protocol, we show step-by-step, in a reproducible manner, the procedure for additional training (fine-tuning) of DNABERT-2, a representative example of a DNA language model, and evaluating its performance for specific research purposes. This protocol enables the construction of AI models that can read meaningful features from DNA sequences, without relying solely on expert intuition or proprietary know-how.
Within the paper, we performed prediction of "unintended RNA effects (RNA off-targets)," which are problematic in genome editing, by fine-tuning DNABERT-2. Furthermore, using a similar fine-tuning method, we also conducted predictions of promoter sequences necessary for creating proteins that carry out life functions from genetic information. We believe that such reproducible protocols can serve as a foundation for life science researchers to predict various genomic challenges and find clues for problem-solving.
In addition, for fine-tuning, we utilized the VM-type GPU cloud service "High-performance VRT" provided by Sakura Internet Inc., a domestic company. The paper reports that while fine-tuning processing can typically be completed in about one day in this environment, a configuration using five other GPUs takes approximately 3-7 days, demonstrating that high-performance GPU clouds significantly influence the speed of research and development. High-speed AI research utilizing domestic cloud infrastructure is also one of this achievement.
Source: Nakamae, K. and Bono, H. (2026). Workflow for Fine-Tuning and Evaluating DNA Language Models for Specific Genomics Issues. Bio-protocol 16(9): e5676. DOI: 10.21769/BioProtoc.5676.
Our Company, with the vision of "Pioneering the Future with Biotechnology" and the mission of "Solving Societal Challenges by Designing Biological Functions," is deploying a platform for genome editing and Bio DX. This publication of literature demonstrates Our Company's efforts to combine AI, genomic data, and computational infrastructure, translating them into practical forms for research sites. Moving forward, Our Company will continue to promote the social implementation of AI-driven genomic analysis and genome editing in fields such as drug discovery/medical care, food & agritech, and environment & energy.
Literature Information Paper Title: Workflow for Fine-Tuning and Evaluating DNA Language Models for Specific Genomics Issues Publication: Bio-protocol Paper Link: https://doi.org/10.21769/BioProtoc.5676
About Platinum Bio Inc. Our Company is a startup established with the aim of implementing Hiroshima University's cutting-edge biotechnology into society. Utilizing our unique Bio DX technology and genome editing technology, we are promoting co-creation projects with various business partners as a platformer that designs biological functions capable of solving global societal challenges. Location: 3-chome-10-23 Kagamiyama, Higashihiroshima City, Hiroshima Prefecture Established: August 30, 2019 Representative: CEO Keisuke Okuhara Business Description: Support for data-driven breeding utilizing Bio DX and genome editing technology.
FACT BOX
- Source: PR TIMES
- Category: Survey