Indian researchers have developed an AI tool that predicts interactions of intrinsically disordered proteins.
These are crucial for understanding diseases.
This open-source model outperforms existing tools and can aid disease research, drug design and understanding complex biological systems.
In a breakthrough that could sharpen how scientists understand disease at the molecular level, researchers at the National Centre for Biological Sciences (NCBS), part of the Tata Institute of Fundamental Research, have developed an artificial intelligence tool that predicts how some of the cell’s vital proteins interact.
The deep-learning, open-source model, called Disobind, focuses on intrinsically disordered proteins (IDP) or regions (IDR). These are flexible, shapeshifting protein segments that do not adopt a single fixed structure. These proteins play critical roles in health and disease, from cell signalling and gene regulation to cancer progression and neurodegeneration.
Traditional structural biology and computational tools have struggled to capture how such proteins bind to others, largely because IDRs change shape depending on context. Disobind addresses this gap by analysing protein sequences using protein language models — AI systems trained on millions of known protein sequences — to predict which parts of a disordered protein will bind to a specific partner.
Crucially, the model considers both proteins involved in an interaction. “Context influences interaction in the case of these floppy proteins,” the researchers noted, making partner-aware prediction essential for biological relevance.
IDRs often act like molecular glue, enabling transient yet precise interactions inside cells. They help assemble dynamic cellular hubs known as condensates, regulate which genes are switched on or off, assist protein folding and quality control, and allow proteins to move efficiently through crowded cellular environments. Their centrality to life also means that when these interactions go wrong, disease can follow.
Kartik Majila, lead author of the study, and his colleagues compared Disobind with leading predictors, including AlphaFold-based multimer models. Disobind showed consistently higher accuracy, particularly when tested on protein pairs it had not encountered during training. Combining Disobind with existing structure-prediction tools further boosted performance.
“Applications span from disease biology to drug design,” said Shruthi Viswanath, who leads the Integrative Structural Biology Lab at NCBS. “With Disobind, we can begin to reveal new interaction motifs linked to disease, suggest intervention points for regulating IDR-mediated interactions across the proteome, and better position disordered segments within large molecular assemblies.”
The tool has use in diverse biological systems, including immune signalling pathways and proteins involved in cancer and neurodegenerative disorders — areas where IDRs are increasingly recognised as key but poorly understood players.
As health research shifts towards systems biology and precision medicine, tools like Disobind could help bridge a long-standing blind spot in molecular understanding. The software is open-source and freely available, allowing researchers worldwide to explore the hidden logic of the cell’s most flexible components.
The study was supported by India’s Department of Atomic Energy, Department of Science and Technology (SERB), and Department of Biotechnology.