Decoding Life with AI: DNA Language Models, Memory Transformers, and the Roadmap to AI-powered Genomics

In this talk, AI researcher Mikhail Burtsev outlines a roadmap for decoding biology with AI. He explains how genome-scale language models can learn the language of life, from DNA sequences to gene regulation, and how memory-augmented transformers can maintain long-range context to improve predictions. The presentation covers practical applications such as promoter prediction in Drosophila, disease risk forecasting from ECG data, and protein structure insights inspired by AlphaFold, while emphasizing transfer learning and interpretability. Burtsev argues that these advances can speed up scientific discovery by linking molecular, cellular, and organ-level processes, and he sketches a future where AI accelerates biology at scale.

Overview

This talk by Mikhail Burtsev explores how artificial intelligence, and in particular large language models, can be repurposed to decode the language of life. He begins with a personal trajectory from neural networks to natural language processing and then shifts to biology, arguing that genome-scale data offers a universal medium for learning the rules that govern living systems. The central idea is to treat the genome as a programming language of biology, where sequences of nucleotides encode instructions that manifest as cellular behavior, organ function, and disease outcomes. Burtsev outlines a roadmap that combines foundational language models with biology-oriented fine tuning to predict, interpret, and reason about biological phenomena.

AI and the Biology Puzzle

The talk situates biology as a highly complex system with multiple organizational levels, from genes to organs. Unlike plain natural language, genomic information exhibits multi-scale interactions, including short-range and long-range regulatory relationships. Burtsev argues that scale and memory are crucial: current large language models excel at local, short-range context, but long-range genomic interactions require mechanisms that can maintain and utilize memory across vast sequences. This recognition motivates the development of memory-augmented architectures for biology.

DNA Language Models and GAM

Drawing on transfer learning paradigms, Burtsev describes two-stage training for DNA language models. First, a large model is pre-trained on natural language or genomic data to learn general representations. Then it is fine-tuned on domain-specific tasks such as promoter prediction, gene regulation inference, or disease association. He introduces the concept of GAM, a DNA language model built to understand genetic sequences and predict biological function, and discusses initial results on promoter prediction in Drosophila, including cross-species generalization where the model trained on one species shows predictive power on another. The goal is to derive a universal genomic understanding that transfers across species.

Memory and Long-Range Genomics

A key focus is extending model memory to handle long genomic contexts. Burtsev explains memory-augmented transformers and an associative recurrent memory approach that allows models to read and retain information over vast sequences, enabling more accurate predictions of gene regulation, binding sites, and other regulatory elements. He highlights breakthroughs in processing tens of millions of tokens and the potential to model the human genome in a way that mirrors how biology operates across scales.

Applications and Practical Outcomes

The discussion moves from theory to concrete applications. Case studies include promoter identification in model organisms, disease risk forecasting from electrocardiogram data, and protein structure predictions inspired by AlphaFold. Burtsev emphasizes that genome-scale models can accelerate hypothesis generation, guide experiments, and provide interpretable insights via attribution methods that connect sequence motifs to predicted outcomes. The potential to integrate genomic models with patient data for precision medicine is highlighted, alongside the importance of robust evaluation and data availability.

Interpretability and Biological Insight

Interpretability is presented as a bridge between black box prediction and scientific understanding. Techniques like feature attribution show which DNA segments most influence predictions, aligning model findings with known transcription factor motifs and regulatory mechanisms. This approach can reveal connections between cellular level processes and organ-level function, offering a common ground for linking diverse biological scales and generating new hypotheses about aging, development, and disease.

Roadmap, Challenges, and the Future

Finally Burtsev sketches a roadmap for the next 5 to 10 years. The path starts with learning the genetic code through genome-scale language models, followed by scaling and memory enhancements to capture long-range interactions. As more genomes are sequenced, the models will grow in capability not merely to predict outcomes but to interpret and connect processes across cellular, organ, and systems levels. He envisions a future where AI-generated hypotheses, simulations, and designs accelerate biology and healthcare, while also addressing ethical and practical considerations in data use, reproducibility, and governance. The talk closes with an aspirational glimpse of the near future where AI collaborates with researchers to engineer new insights and potentially novel biological constructs, all grounded in a commitment to credible, science-based discovery.

Conclusion

Burtsev concludes by presenting a concrete research agenda at the London Institute of Mathematical Sciences and emphasizes the collaboration between AI and biology as a path to understanding life’s language and improving human health, with a roadmap built on translation from genome sequence to function, context-aware learning, and cross-scale integration.