An international research team co-authored by Dr. Sushmita Basu, Assistant Professor in the Department of Biological Sciences at Birla Institute of Technology and Science (BITS) Pilani, has developed an artificial intelligence (AI) tool to identify protein regions that interact with DNA, potentially improving research into gene regulation and genetic processes.
Named DNAreader, the tool predicts DNA-binding residues in proteins across both structured and intrinsically disordered regions. The study was published in the journal Nucleic Acids Research.
Proteins interact with DNA to regulate gene activity, replicate and repair genetic material, and maintain genome organisation. Identifying the specific amino acid residues responsible for these interactions remains challenging, particularly in intrinsically disordered regions, where proteins lack a stable three-dimensional structure.
Existing prediction tools can also mistakenly classify protein regions that interact with other molecules as DNA-binding sites, reducing prediction reliability.
AI model combines contrastive learning with balanced training
The research team developed DNAreader using two techniques to improve prediction accuracy and reduce false positives.
The first is a batch-training strategy that balances examples of DNA-binding residues, residues interacting with other ligands and non-binding residues across structured and disordered protein regions. The second uses contrastive learning to help the model distinguish between these different residue types.
The approach enables DNAreader to assess structured and disordered protein regions within a unified prediction framework.
“Accurately identifying DNA-binding residues is particularly challenging when proteins contain intrinsically disordered regions. DNAreader addresses this gap by bringing structured and disordered regions within a unified prediction framework,” said Dr. Sushmita Basu, Assistant Professor in the Department of Biological Sciences at BITS Pilani.
The international collaboration involved Dr. Jian Zhang of Xinyang Normal University, China; Dr. Sushmita Basu of Virginia Commonwealth University and BITS Pilani; Dr. Jingjing Qian of Hebei University of Technology, China; and Dr. Lukasz Kurgan of Virginia Commonwealth University, US.
Benchmark results show improved prediction accuracy
DNAreader was evaluated against 14 existing prediction tools using independent test datasets designed to minimise sequence similarity between training and test proteins.
According to the study, DNAreader achieved an area under the receiver operating characteristic curve (AUROC) of 0.86 to 0.89, compared with 0.75 to 0.79 for the leading existing tools.
AUROC measures a model’s ability to distinguish between positive and negative cases, with higher values indicating better discrimination.
The researchers also demonstrated the tool’s performance using the human estrogen-related receptor ERRβ, where it identified DNA-binding residues across structured and disordered regions. In a separate control enzyme, it avoided incorrectly flagging residues as DNA-binding sites.
These capabilities could help researchers investigate protein-DNA interactions and improve computational approaches to studying biological systems.
Free online access for researchers
DNAreader is available through a free web server that accepts protein sequences in FASTA format, allowing researchers to submit sequences for analysis without requiring specialised computational infrastructure.
The tool can be accessed at the DNAreader web server. Its source code and datasets are also available on GitHub.
The research highlights the growing role of AI in computational biology, where machine-learning methods are being used to analyse complex biological data and identify molecular interactions that can be difficult to establish experimentally.