Hallucination Co-Design of Protein–RNA Complexes

Exploring hallucination and model-inversion strategies for protein–RNA binder design.

Overview

This project explores generative co-design of protein–RNA complexes through a hallucination / model-inversion framework. The central idea is to use pretrained structure predictors as structure-aware scoring functions to guide interface generation and optimization.

This work is being carried out at the Chinese University of Hong Kong under the supervision of Prof. Yu Li.

What I worked on

I have been prototyping a workflow for protein–RNA binder design inspired by hallucination-based and BindCraft-style optimization strategies. The project focuses on how multimodal structural information can be incorporated into iterative design.

Methods / Pipeline

The design framework explores confidence- and geometry-based objectives, including:

  • structure-confidence signals analogous to pLDDT / PAE
  • interface contacts
  • steric clashes
  • multimodal interface quality constraints

The broader goal is to use these objectives to iteratively optimize candidate protein–RNA interfaces.

Current status

This project is currently at the prototyping stage. It has helped me think more deeply about how pretrained complex predictors can be repurposed as differentiable or quasi-differentiable scoring modules for molecular design.

Research significance

This direction is closely aligned with my long-term interests in AI-driven biomolecular design, especially beyond protein-only systems.