FusOn-pLM: Advancing Precision Therapy for Fusion Oncoproteins through Enhanced Protein Language Modeling

Protein language models (pLMs) are a type of machine learning model that has recently emerged as a powerful tool for analyzing and understanding proteins. These models are trained on large datasets of protein sequences and learn to capture the structure, function, and evolutionary information of proteins from the sequence data alone5. By treating amino acid sequences as "sentences" and individual amino acids as "words," pLMs can identify patterns and relationships between different parts of a protein sequence, much like how natural language processing models understand human languages.

pLMs contribute to understanding protein functions in several ways:

1. Predicting protein structures: pLMs have been used to predict protein structures with high accuracy, even outperforming traditional experimental methods in some cases4.

2. Identifying functional sites: By analyzing the patterns and relationships within a protein sequence, pLMs can help identify functional sites, such as active sites or binding sites, which are crucial for the protein's function.

3. Understanding protein evolution: pLMs can reveal evolutionary relationships between proteins, helping researchers trace the evolutionary history of a protein family and understanding how functions have evolved over time.

4. Designing new proteins: pLMs can be used to design new proteins with desired properties, such as improved stability or specific functions, by generating novel sequences that conform to the learned patterns and relationships.

5. Informing drug discovery: By providing detailed insights into protein function and structure, pLMs can aid in the design of new drugs that target specific proteins5.

In summary, protein language models have revolutionized the field of protein research by providing a powerful tool to analyze, understand, and design proteins, contributing to advancements in drug discovery, protein engineering, and our understanding of protein function and evolution.

FusOn-pLM, a novel protein language model developed by researchers at Duke University, has shown superior performance in predicting the behavior and properties of fusion oncoproteins compared to other models and embeddings. The key ways in which FusOn-pLM outperforms other models are:

1. Improved representation of fusion oncoproteins: FusOn-pLM employs an innovative masked language modeling approach, focusing on key residues likely involved in protein interactions. This method improves the representation of fusion oncoproteins, outperforming the base ESM-2 model and other embeddings in various benchmarks.

2. Enhanced embeddings for therapeutic targeting: The embeddings generated by FusOn-pLM are specifically designed to aid in the therapeutic targeting of challenging fusion oncoproteins. By fine-tuning the advanced ESM-2 pLM on a large dataset of fusion oncoprotein sequences, FusOn-pLM achieves better performance in tasks related to fusion oncoproteins.

3. Superior performance in benchmark tasks: FusOn-pLM was evaluated on several benchmark tasks, including predicting phase separation of fusion oncoproteins, their localization in the cell, and their associations with specific cancers. The model demonstrated superior performance in these tasks, outperforming other models and embeddings.

4. Effective identification of intrinsically disordered regions: FusOn-pLM excels in identifying intrinsically disordered regions and their physicochemical properties, which are crucial for understanding the behavior of fusion oncoproteins. It outperforms other embedding methods in this regard.

5. Distinct separation of fusion oncoproteins: Visualization techniques showed that FusOn-pLM embeddings distinctly separate fusion oncoproteins from their components, reflecting their unique characteristics and biological relevance. This ability to differentiate fusion oncoproteins from their components is a significant advantage over other models.

Overall, FusOn-pLM's enhanced embeddings and focused training strategy enable it to outperform other models and embeddings in predicting the behavior and properties of fusion oncoproteins.