FusOn-pLM

FusOn-pLM fine-tunes the ESM-2 protein language model on fusion oncoprotein sequences using a focused probabilistic masking strategy to generate information-rich embeddings for therapeutic design.

Key Features:

• Fusion oncoprotein-specific fine-tuning: Fine-tunes ESM-2 on fusion oncoprotein sequences to produce optimized representations for this protein class.
• Focused probabilistic masking: Employs a binding-site probability predictor to focus masked language modeling on key amino acid residues, generating more fusion oncoprotein-aware embeddings.
• Improved benchmark performance: Outperforms baseline ESM-2 representations and manually-constructed biophysical embeddings on fusion oncoprotein-specific benchmarks and disorder prediction tasks.

Scientific Applications:

• Therapeutic design for fusion oncoproteins: Enables downstream design applications targeting fusion oncoproteins, which are challenging to address with small molecule-based methods or structure-based binder design approaches due to their intrinsically disordered nature, large size, and lack of well-defined druggable pockets.
• Disorder prediction: Improves performance on protein disorder prediction tasks compared to baseline representations.

Methodology:

FusOn-pLM fine-tunes ESM-2 using a novel masked language modeling strategy with a binding-site probability predictor that focuses masking on key amino acid residues to generate fusion oncoprotein-aware embeddings.