Generative AI Models for the Protein Scaffold Filling Problem.

Area of Science:

• Proteomics and Bioinformatics
• Artificial Intelligence in Life Sciences

Background:

• De novo protein sequencing is vital for understanding protein functions, drug discovery, and evolutionary studies.
• Mass spectrometry techniques like top-down and bottom-up tandem MS are common but often yield incomplete protein sequences with gaps (scaffolds).
• The protein scaffold filling problem aims to infer complete protein sequences by filling these gaps.

Purpose of the Study:

• To address the protein scaffold filling problem using advanced generative AI techniques.
• To evaluate and compare the performance of various AI models, including convolutional denoising autoencoders, transformers, and GPT models.
• To assess model efficacy on both real and generated datasets.

Main Methods:

• Application of generative AI models: convolutional denoising autoencoder, transformer, and generative pretrained transformer (GPT).
• Comparison with a convolutional long short-term memory (CLSTM)-based sequence model.
• Performance evaluation using real and generated protein scaffold datasets.

Main Results:

• All proposed generative AI models demonstrated outstanding prediction accuracy in protein scaffold filling.
• The GPT-2 model achieved 100% accuracy in both gap-filling and full sequence determination for the MabCampth protein scaffold.
• GPT-2 significantly outperformed other evaluated models on the MabCampth dataset.

Conclusions:

• Generative AI, particularly GPT-2, presents a highly effective solution for the protein scaffold filling problem.
• These AI-driven methods can accurately reconstruct complete protein sequences from incomplete mass spectrometry data.
• The findings highlight the potential of AI in advancing proteomics research and applications.