Design

Prediction and Validation of Potential Adhesive Proteins

In the third part of the project, we aim to expand CoPlat's application range and transform it into a versatile platform. To accomplish this, our team aims to find “potential adhesive proteins” with machine learning techniques.

1. Collection of protein database
First, we searched for words such as "adhesive," "cohesive," "sticky," etc., in UniProt's function and keyword searching field to collect protein data and expand our database.
Then, we analyzed the data and organized it into positive data for our project database. Then, we collected negative data according to the length of the positive data to test the stability of our model.

2. Establishment and Testing of Machine Learning Prediction Model
To attain our objective, we must build a classifier to identify potential adhesive proteins to enhance the flexibility and abundance of CoPlat in use. We employed multiple protein sequence coding methods, including AAPC, AA index, and ESM, and transformed them into vector representations for machine learning model testing with limited data.
Then, we adopted three machine learning models: Gradient Boost, Random Forest, and SVM, each coupled with three encoding methods. These models can effectively distinguish between adhesive and non-adhesive proteins while providing a probability score.
To ensure the accuracy and fairness of the models, we chose SVM as the model and used balanced training and Jackknife methods to prevent bias and assess accuracy. Stability tests and overfitting tests further validate the reliability of the model.
Finally, we applied the tested and analyzed models to datasets with different characteristics, such as specific GO domains, sequence lengths, specific motifs, etc., to achieve accurate prediction of adhesive proteins.

3.Software Interface
We have developed a tool, "FoxyProt," for identifying proteins with adhesive properties. To increase accessibility, we developed a user interface with a graphical user interface. Key steps of the software include:

A. Model export We used the joblib module to enable efficient export of machine learning models to ensure that future teams can easily load and use the models without retraining.

B. Graphical user interface We used the Tkinter library to create a graphical user interface that allows users to input data, make predictions, and visualize results easily.

C. Current functionality Users can select a fasta file, and then by embedding features, our SVM model categorizes the proteins and outputs the probability of being an adhesive protein.

This part involved using machine learning techniques and modeling to predict and validate potential adhesive proteins. We expanded the database with multiple protein sequence coding methods, built high-performance machine learning models, and provided future teams with easy-to-use tools and interfaces to easily identify adhesive proteins.