• Citer

Protein melting temperature prediction: A protein language processing predicting the melting temperature from protein sequences

par Pirard, Delanoe
Président du jury Flot, Jean-François
Promoteur Rooman, Marianne
Co-Promoteur Pucci, Fabrizio
Publication Non publié, 2023-01-27

Mémoire

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

The field of the study of proteins is a vast field of the study of biology. In fields such as medicine where certain diseases are linked to proteins such as neurodegenerative diseases, or even in food products or biotechnological products, the study of the thermal stability of proteins is essential.Recently, a new dataset including about 48,000 proteins with their associated melting temperature was published. With this considerable increase in the amount of data available on this subject, it becomes possible to make much more efficient Machine Learning models, as is the case with the recent ProTstab2 model.In this work, we tried to create a predictive model of melting temperature using several Deep Learning concepts. First of all, the model is based on a new attention algorithm, the Sparse Attention Mechanism, which should allow fairly good performance while reducing the consumption of computer resources. Then, the concept of pretraining which is supposed to bring us better predictions by already training the model on a large amount of data before specializing it. We also use SentencePiece a tokenization software which does not need hand crafted rules to understand every language.Unfortunately, the results were not what we expected. All the models converge towards the mean of the melting temperatures composing our datasets. In discussion, we will explain what could be the reasons why the models do not work well and what are the areas for improvement.