This repository presents an academic research project that offers a comprehensive assessment of EVO 2, a state-of-the-art AI model developed for genomic analysis. The study investigates EVO 2 by thoroughly examining its data pipeline architecture, revealing how its structure, training processes, and data handling strategies give rise to broader concerns about risk, ethics, and governance in AI-driven genomics.
EVO 2 was developed in collaboration between the Arc Institute, NVIDIA, Stanford, UC Berkeley, and UCSF. Capable of processing genomic sequences up to 1 million base pairs, EVO 2 utilizes a novel architecture (StripedHyena 2) and was trained on a massive OpenGenome2 dataset containing over 9.3 trillion nucleotides.
This project does not aim to build or extend EVO 2, but instead to critically assess the implications of its data pipeline, including:
- Data quality controls and privacy protections
- Risks of genetic misuse or unintended consequences in genome editing
- Bias inheritance from organismal diversity in the training data
- Informed consent in genomic data collection and use
- Transparency in excluded pathogen data and its impact on research access
- Fair use and equitable treatment in predictions and applications
- International regulation needs and ethical frameworks
- Public accountability and legal alignment (e.g., GDPR, EU Bioethics)
- Real-time oversight of AI systems in biomedical research
- EVO2_Assessment_Report.docx – Full research document including references
- Data-Pipeline-chart-of-Evo-2.png - A visual of the EVO 2 data pipeline designed by me
- README.md
This project is for academic, educational, and policy discourse purposes. No proprietary data or private genomic information is used or accessed. All conclusions are based on publicly available documentation and open-source model behavior.
Open-access under CC BY-NC 4.0. Refer to the LICENSE file for reuse conditions.