Enhancing protein function prediction: integrating neural networks with ARGOT 2.5 web server

The prediction of protein function through in silico methods is becoming more prevalent and important. An increasing amount of gene products is continuously sequenced thanks to the accessibility of New Generation Sequencing (NGS) techniques. Unfortunately, this high output of information is not met by an appropriate amount of research regarding the functions of those proteins. Due to the accurate but resource-intensive nature of experimental evidence, the field of molecular biology cannot keep up with the number of gene products that are discovered regularly. The category of Automated Function Prediction (AFP) techniques was born to address this problem and generate the required annotations necessary to describe the functions of the proteins that cannot be adequately studied due to time or resource constraints. AFPs aim to be a fast and cost-effective substitute for wet lab research while keeping a similar level of accuracy. Argot 2.5 is an AFP created by the group of Computational Medicine of the University of Padua, designed to predict protein function starting from its amino-acidic sequence. The efforts of this project revolve around the implementation of a new evaluation system that estimates the accuracy of each prediction through the use of a neural network. This neural network will weigh Argot’s multiple outputs to provide to the end user a curated selection of the most reliable functions predicted by Argot. In addition, the research team noticed that Argot predicted some functions that were wrong at the time of the prediction but were recognized as right years later after the publication of further experimental evidence. This means that Argot may have predictive capabilities that can be used to lead the direction of the experimental research. This project also comprehends the development of a second neural network focused on finding which function predictions, scored as false positives at the time they were generated, have a high probability of being corroborated in the future, which may be worthy of further focused research. Both neural networks developed in this thesis project provided an increase in the performance of the function predictions generated by the Argot 2.5 server and will be implemented in the novel version of the tool.