Shiru, a functional ingredients discovery company, has presented a paper entitled “Improving Protein Subcellular Localization Prediction with Structural Prediction & Graph Neural Networks” at NeurIPS, the leading conference of the AI research community.

The simple and effective method combines Language Models (LM) and Graph Neural Networks (GNN) to predict protein subcellular localization — with implications for function prediction.

Predicting where a protein localizes or resides in a cell is a huge challenge in biotechnology. Where a protein resides is an important indicator of its function and manufacturability. Many processes, such as disease mechanisms, drug performance, regulation of metabolic processes, and signaling cascades, depend on a protein’s localization.

Previous work in the field has shown how Language Models (LM) and Graph Neural Networks (GNN) can independently provide efficient localization predictions when trained on protein DNA sequences and 3D structures, respectively. Shiru AI researchers created a method for combining the two different kinds of protein representations used by LM and GNN, then used real-world data to show how ensembling them outperforms the reigning state-of-the-art method.

“This paper validates our approach and offers a generalized method that opens new possibilities toward creating a sequence-to-function map of the protein universe,” says Geoffroy Dubourg-Felonneau, Shiru’s Machine Learning Lead and the paper’s lead author. “We are continually developing new and improved methods for protein representation, and we show in this case that the combination of protein structure information and language modeling yields a significant improvement on the task of subcellular localization prediction.”

“Shiru is pioneering the use of advanced machine learning techniques to reveal proteins with similar function but dissimilar sequence and structure relative to a target protein,” says Lawrence Lee, Shiru’s Chief Technology Officer. “This new approach enhances our ability to uncover the hidden food functionality of proteins.”

The new AI method is the latest from Shiru’s science and technology team. In September 2022, Shiru was awarded a key patent for its protein discovery platform, which covers machine learning combined with lab analysis for developing naturally-occurring proteins as functional food ingredients. Shiru’s Flourish™ technology is being proven through partnerships with global food leaders Puratos and CP Kelco. Shiru also recently announced that Impossible Foods veteran Ranjani Varadan joined as Chief Scientific Officer.

Learn more or use the green "Request Information" button below to contact the supplier for more details.