SRI graduate fellows explore the evolution of genomic language models
Can AI unlock the hidden rules of our DNA and revolutionize medicine? SRI Graduate Fellows Micaela Elisa Consens and Ben Li explore this question in a new commentary examining the potential of genomic language models to transform biomedical research. From left to right: Micaela Elisa Consens and Ben Li (supplied images).
Can artificial intelligence (AI) decode the fundamental rules of life hidden within our DNA? Could language models predict the risks of genetic diseases or even design novel treatments tailored to individual patients? These questions are at the heart of a new commentary, published this week in npj Digital Medicine by Schwartz Reisman Institute Graduate Fellows Micaela Elisa Consens and Ben Li, on the rapid evolution of genomic language models and their potential to transform biomedical research and clinical medicine.
Genomic language models (gLMs) have emerged as transformative tools for interpreting the complex regulatory functions of DNA. Unlike protein language models that focus on the 2% of human DNA encoding proteins, gLMs analyze the entire genome—including the 98% of non-coding DNA that plays a crucial role in regulating gene expression. Evo2, the latest advancement in the field, represents a major leap forward, scaling up genomic AI to match the computational power of leading text-based large language models.
Consens and Li’s commentary emphasizes the potential of gLMs to transform biomedical research, disease prediction, and personalized medicine. At the same time, they highlight many open questions within the field, including challenges regarding model interpretability and ethical concerns and risks related to gLMs’ potential use.
AI innovation powers new approaches
The rapid advancement of AI in medicine is transforming how we understand and treat disease, with research on interpretability and real-world applications serving a critical role in unlocking the value of AI innovation. Through their 2024–25 SRI fellowships, Consens and Li are exploring how AI technologies are powering new types of biological and medical research and practice.
Consens is a PhD candidate in the University of Toronto’s Department of Computer Science who specializes in interpretability and genomic language models. Affiliated with the Wang Lab and Moses Lab, as well as the Vector Institute, Consens’ research focuses on developing novel computational methods to uncover fundamental insights into human molecular biology, including how AI-driven genomic models can be interpreted and applied to answer real-world biological questions. Her SRI fellowship project aims to enhance the interpretability of transformer-based genomic models, decoding how these models make decisions and translating them into biologically meaningful insights.
“One of the most pressing issues in genomic AI is understanding how these models make their predictions,” Consens explains. “If we can interpret their decision-making processes, we can potentially learn novel biology from them.”
Li is a vascular surgery resident and PhD candidate at U of T’s Institute of Medical Science, where he applies machine learning techniques to improve patient outcomes following major vascular surgery. With over 80 peer-reviewed publications, 60 conference presentations, and 30 research grants and awards, Li is making an impact when it comes to the successful use of AI predictions in surgery. His SRI fellowship project focuses on developing automated surgical risk prediction tools that could be deployed at local, provincial, and international levels to improve risk-mitigation strategies and reduce healthcare costs.
"AI has the potential to improve surgical outcomes by predicting complications before they happen," says Li. "The key is ensuring these models are trained and evaluated on high-quality data so they can be trusted in clinical decision-making."
Language models and the future of genomics
As genomic language models like Evo2 emerge as promising tools for studying regulatory genomics, questions about their interpretability, ethical deployment, and real-world clinical utility become more urgent. The researchers point to improvements in model pre-training, fine-tuning, and context size as crucial to the development of more sophisticated tools. Trained on over 128,000 genomes, Evo2 is a major advancement from the previous largest gLM, which trained on only 850 genomes.
The researchers note that more sophisticated pre-training will offer immense potential through “zero-shot” performance—meaning a model’s ability to succeed in tasks it wasn’t explicitly trained to do—in other domains. “Strong zero-shot performance indicates the model has learned fundamental principles about genomic structure that generalize to new scenarios,” they write. “Potentially, this means gLMs with strong self-supervised zero-shot performance have uncovered new regulatory grammar within the genome—grammar that we can learn from. Uncovering novel genomic grammar would advance our understanding of human disease and transform personalized care across all aspects of medicine.”
Given that almost all leading causes of death or disability in the world have important genetic components, the authors contend that “it is likely that in the future, gLMs could help clinicians estimate the risks of whether a patient will develop these diseases, years before their onset, and implement appropriate personalized preventive strategies.”
However, they also point to challenges in the clinical adoption of gLMs, including researchers’ ability to assess whether the models are genuinely learning contextual relationships or merely memorizing patterns from training. Challenges in designing biologically meaningful benchmarks can result in an approach that “often fails to test models on the complex regulatory patterns they are ultimately intended to discover.”
Ethical considerations regarding the responsible development of gLMs—including questions of privacy and consent, dual-use risks, and access—will also become increasingly prevalent as such techniques approach practical application. While Evo2 was trained on open-source genomic data, the future clinical use of gLMs to analyze individual patient DNA will require privacy safeguards and clear consent mechanisms. Additionally, the ability of these models to generate entire genomes raises concerns about potential misuse, including biosecurity threats. The high computational costs of running gLMs could also create disparities in healthcare access, limiting advanced genomic diagnostics to wealthier populations. To mitigate these risks, the authors highlight that regulatory frameworks must prioritize AI safety, transparency, and equitable implementation to ensure that gLMs serve the public good without exacerbating existing inequalities.
The need for multidisciplinary expertise
Consens and Li’s contributions underscore the multidisciplinary expertise within the Schwartz Reisman Institute’s research community, where scholars from computer science, medicine, and other disciplines collaborate to push the boundaries of technological and biomedical innovation. Their commentary provides a lens on what researchers can look forward to in a future where genomic AI is harnessed to unlock biological discovery and clinical advancements, while also addressing the pressing need for interpretability, ethical oversight, and equitable deployment of these technologies.
"SRI’s graduate fellowship has been an incredible opportunity for me to work with and learn from brilliant scholars across many disciplines,” says Li. “As a clinician with an interest in AI, collaborating with technical experts like Mica is invaluable. This project is just one of many examples of the remarkable work that SRI graduate fellows are engaged in.”
Consens’ research on interpretability has the potential to uncover how AI-driven genomic language models are learning to make decisions, while Li’s work on AI-powered surgical risk prediction addresses the need for data-driven, personalized solutions that can improve patient outcomes. Their research highlights the immense promise and the challenges of integrating AI into biological research and clinical medicine at a time when the field is rapidly evolving.
Want to learn more?
● Read “Genomic language models could transform medicine but not yet” in Nature Digital Medicine.
● Read Micaela Elisa Consens’ review on genome language models in Nature Machine Intelligence.
● Learn more about how SRI researchers are addressing bias in clinical AI models.
