A bold new research initiative from Kingston University London aims to revolutionise genetic science by harnessing artificial intelligence, supported by funding from Amazon Science. The project will focus on deciphering the vast, non‑coding regions of human DNA, often referred to as the dark genome, which hold critical insights into disease and personalised medicine.

The research team, led by senior lecturer in computer science Dr Farzana Rahman, Knowledge Exchange and Research Institute Director for Cyber, Engineering and Digital Technologies
Professor Jean-Christophe Nebel and PhD candidate Megha Hegde, has been awarded the prestigious AWS AI Research Award. Their goal is to develop and train highly efficient Language Learning Models (LLMs) to decode non-coding(dark) DNA.

For decades non-coding DNA, which makes up most of the human genome, was dismissed as junk. However, studies now confirm it exerts powerful control over gene expression, with over 90 per cent of disease-associated variants residing in these non-coding regions. Understanding these is key to personalised medicine but has remained a significant computational hurdle.

Existing genomic LLMs treat DNA sequences like language, but their architecture struggle to process the tens of thousands of base pairs needed to capture the distant regulatory signals.

A previous study by the Kingston research team, published in MDPI Genes, discovered not all layers in a genomic LLM are essential, showing that pruning non-critical layers, could almost halve fine-tuning time while retaining accuracy pointing to the need for leaner LLM architectures.

Building on this learning, this latest study will focus on novel architectures that promise significantly more efficient handling of very long DNA sequences. The team will use Amazon’s Trainium hardware and AWS Neuron software stack to train Mamba, Hyena and RKWV, models uniquely suited to model the sparse, long-range interactions found in regulatory DNA.

The researchers aim to demonstrate that state-of-the-art variant interpretation can be achieved at a lower cost and with a smaller carbon footprint than current methods reliant on comparable graphics processing unit clusters.

If successful, the project could transform personalised medicine by improving predictions of splicing changes, transcription factor binding, chromatin interactions and tissue‑specific effects. It could also highlight why certain mutations are pathogenic.

Professor Nebel said LLMs emphasised the transformative potential of the initiative noting that LLMs are now being used to decode the language of life: our genetic code. “We’re exploring how these models are advancing our ability to predict the effects of human genetic mutations. This is a highly complex challenge that requires significant computing power and the Amazon Research Award will enable us to develop faster, more powerful solutions. In the face of human suffering, every processor cycle counts.”

Dr Rahman added the project will address how resource-intensive the LLMs are. “We’re investigating pruning and novel architectures that reduce model size and energy consumption while maintaining accuracy. Our project embodies this mission by enabling the development of smaller, more sustainable genomic language models and by balancing scientific ambition with environmental stewardship, we hope to bring precision medicine closer to reality.”

PhD researcher Meg said deep learning and LLMs have revolutionised the way technology can enable and accelerate scientific discovery. “It is our responsibility as computer scientists to ensure that our positive scientific contributions are not outweighed by their negative impacts on the environment, such as the huge amounts of water and electricity consumed. Promisingly, recent research has shown even small changes to model architectures are able to significantly reduce resource usage while preserving accuracy. Our paper is a small step towards solving this crucial puzzle.”