Deep Learning-Based RNA Tertiary Structure Prediction and Generation

Project duration: 2024 - 2027

Funding: Croatian Science Foundation   

RNA has long been neglected as a potential target for drug development due to a lack of detailed structural understanding. However, this perception is changing as its significance is increasingly acknowledged, particularly considering that non-coding RNAs make up a substantial portion of the genome, ranging from 70% to 75%. Similar to proteins, RNAs adopt complex tertiary structures that exert control over various stages of their life cycle, including transcription, translation, degradation, and transportation. The advent of AI-driven tools, notably in protein structure determination such as AlphaFold2, coupled with growing insights into the intricate folding patterns and druggability of RNA, has reignited interest across academia and industry. Nevertheless, significant hurdles remain in adapting protein-like tools for accurately predicting RNA tertiary structures. This project proposal aims to address these challenges and gain a profound understanding of the limitations of current deep-learning approaches in predicting RNA tertiary structures as well as developing new methods. Our approach involves our recently developed RNA large language model RiNALMo as an input to structure prediction. RiNALMo showed surprising generalization when we fine-tuned it to the secondary structure prediction. Moreover, on top of this, we will explore the development of deep generative methods for modeling the conformational space of RNA molecules and designing RNA tertiary structures. We believe the developed methods will have an important impact on structural biology, especially for drugging RNA.

Project team

Project members:

  • Prof. dr. sc. Mile Šikić - head
  • Dr. sc. Tin Vlašić (A*STAR GIS, Singapore)
  • Rafael Josip Penić - phd student
  • Dr. sc. Roland G. Huber (A*STAR BII, Singapore)