Structural-Protein-Sequences

My aim is to classify maximum protein types in the Structural Protein Sequences dataset using multiple machine learning and deep learning classification algorithms. The dataset is from Kaggle.

Data Preprocessing

• Format the classification names
• Remove the outliers for phValue and Molecular Weight. However, there are some high values in Molecular Weight which I think "Molecular weight is calculated using sequence using the weight of each amino acid. Thus, if a sequence is long, then it can have a large weight"
• Remove crystallizationMethod, crystallizationTempK, publicationYear and pdbxDetails. crystallizationMethod and crystallizationTempK have 22% missing values
• Drop some rows based on : densityMatthews, densityPercentSol, resolution, and sequence
• Replace phValue nan value with the average value of that particular class and remove all classes if all phValues are null
• Didn't consider classes that have less than 100 values per class
• Added class variable as a target value. Assign a unique class number to each protein class

Data Classification

I compared a few models with cross-validation and hyperparameter tuning with Grid Search and Randomized Search. The following models were compared:

• Random Forest
• Decision Tree
• Gradient Boosting
• Gaussian Naive Bayes
• Multinomial Naive Bayes
• Support Vector Machine (SVM)
• K Nearest Neighbors (KNN)
• Logistic Regression with OneVsRest

Random Forest is the best among all above models. KNN also works well whereas all other models give very poor performance.

I also implemented Character level 1d CNN and LSTM using keras and TensorFlow in which CNN performs very well.