Fine-tune and deploy the ProtBERT model for protein classification using Amazon SageMaker
Proteins, the key fundamental macromolecules governing in biological bodies, are composed of amino acids. These 20 essential amino acids, each represented by a capital letter, combine to form a protein sequence, which can be used to predict the subcellular localization (the location of protein in a cell) and structure of proteins.
The study of protein localization is important to comprehend the function of protein, which is essentially to structure, function, and regulate the body’s tissues and organs. Protein localization has great importance for drug design and other applications. For example, we can investigate methods to disrupt the binding of the spiky S1 protein of the SARS-Cov-2 virus. The binding of the S1 protein to the human receptor ACE2 is the mechanism which led to the COVID-19 pandemic [1]. It also plays an important role in characterizing the cellular function of hypothetical and newly discovered proteins [2].
Protein sequences are constrained to adopting particular 3D shapes (referred to as protein 3D structure) optimized for accomplishing particular functions. These constraints mirror the rules of grammar and meaning in natural language, thereby allowing us to map algorithms from natural language processing (NLP) directly onto protein sequences. During training, the language model learns to extract those constraints from millions of examples and store the derived knowledge in its weights. [1] Although existing solutions in protein bioinformatics [11, 12, 13, 14, 15,16] usually have to search for evolutionary-related proteins in exponentially growing databases, language models offer a potential alternative to this increasingly time-consuming database search because they extract features directly from single protein sequences. Additionally, the performance of existing solutions deteriorates if a sufficient number of related sequences can’t be found; for example, the quality of predicted protein structures correlates strongly with the number of effective sequences found in today’s databases [17].
