Helen Power1, Dr. Marltan Wilson1, Dr. Candice Jones1, Dr. Andrew Warden1
1Advanced Engineering Biology Future Science Platform, Environment, Commonwealth Scientific and Industrial Research Organisation, Black Mountain, Australia
Biography:
Presenter 1: Helen Power is a researcher working in the Data-Driven Molecular Design (DDMD) Theme in the Advanced Engineering Biology Future Science Platform at CSIRO. She is focused on enhancing the throughput and accuracy of computational workflows for solving a wide-range of problems involving protein design, optimisation and property prediction. She is passionate about leveraging AI-based tools to accelerate scientific discoveries and making these tools more accessible to the scientific community.
Presenter 2: Andrew Warden leads the Data-Driven Molecular Design (DDMD) Theme in the CSIRO Advanced Engineering Biology Future Science Platform. He has a background in synthetic chemistry, protein engineering, and molecular modelling which he applies broadly in the environmental, agricultural and health domains. The DDMD Theme is focused on rapidly implementing and assessing AI-based tools for protein and small molecule design, and developing accessible and high-throughput workflows for both specialist and non-specialist scientists.
Abstract:
Situation: Developing novel biosensors has significant potential to solve various medical, environmental, agricultural and biotechnological challenges. One effective sensing mechanism involves engineering binding proteins, which emit a signal upon detecting their target. Traditional design approaches require extensive experimentation, labour intensive protein design and have high failure rates. Recently, the explosion of deep-learning tools has unlocked the potential for rapid design of protein-based biosensor recognition elements against almost any target molecule, completely from scratch. However, robust workflows rely on substantial computational resources and experimental validation of such approaches is still a bottleneck.
Task: The purpose of this study is to construct and validate a high-throughput computational workflow for the rapid development of protein-based biosensors.
Action: The workflow employs three recently developed deep-learning tools: RFdiffusion for generating biosensor backbone structures, LigandMPNN for decoding amino acid sequences of each structure and AlphaFold for predicting the structure of protein-binder complexes. Due to the high memory and compute requirements of Alphafold, the Google Vertex AI platform was used to considerably increase prediction throughput. In addition to deep-learning methods, physics-based tools including molecular dynamics simulations (Amber22) and RosettaScripts, were used to assess and optimise designs.
Result: Our computational workflow has generated thousands of biosensor designs against various targets including organisms, proteins and small molecules. Experimental testing is underway to validate the workflow. The computational approach enables significantly accelerated biosensor development.