Ⅰ. Data Collection

Positive Antimicrobial Peptides:

Dataset source： APD、DADP、DRAMP、DBAASP、YADAMP

Filter Criteria: Specific to Streptococcus mutans

Steps:

• select the samples that are specific to streptococcus mutans
• Remove 7 duplicate sample sequences
• Remove 5 samples with MIC greater than 500
• Remove 1 sample sequence with imprecise MIC values

Results:

206 positive Antimicrobial Peptides

Negative Antimicrobial Peptides:

Dataset source： AntiBP2

Filter Criteria: Not possessing antibacterial properties

Step:

• select the samples without antibacterial properties
• Remove 0 duplicate sample sequences
• Remove 5 samples that do not contain common amino acid sequences

Results：

994 negative Antimicrobial Peptides

Construction:

Anti-S. Mutans Peptides（ASP） Dataset

Ⅱ. Data Processing

Objective：

• Eliminating the phenomenon of homologous bias

• Calculating physicochemical descriptors

• Balancing the ASP Dataset

Calculating physicochemical descriptors:

​ This part can be divided into three steps：

• Basic physicochemical descriptors calculation
• Amino acid composition calculation (single peptide and dipeptide)
• CTD sequence derivation calculation

Feature calculation table(part)

Features： 578

Significant Feature ： MIC（Minimum Inhibitory Concentration） (./data/ASP.csv)

Ⅲ. Mann-Whitney U test

Objective:

Filtering negative and positive samples without significant differences in specific features

Approach:

our experiment dataset has these characteristics as blow:

• non-normal distribution
• small sample size

The best way to test the two sample/group is Mann-Whitney U test

Mann-Whitney U test result are as follows:

Results:

ASP dataset after filtering statistically irrelevant features (./result/ASP_filtering.csv)

Ⅳ. Feature Extraction

Objective:

gather features that play crucial role in the MIC value of antimicrobial peptides

Method_1： Random Forest

steps:

• Select the columns "LABEL" from the ASP dataset as y variable
• Drop the columns "SEQUENCE"，“MIC” and "LABEL" from the ASP dataset as X variable
• Separate the variable X and y as X_train,y_train,X_test,y_test
• Balance X_train and y_train by applying Borderline-SMOTE
• Use Random Forest model to select the feature and calculate the importance of features
• Extract features with an importance greater than N times the average, and use the RF model for feature extraction
• Based on the accuracy of local optima, optimize the parameter N
• Save extracted features and corresponding importance values

Evaluating indicator：

	Accuracy	Precision	Recall	AUC
Test set	0.97	0.95	0.94	0.99
Train set	1.00	1.00	1.00	1.00

Final Selected Features：

Method_2： XGboost

steps:

• Select the columns "LABEL" from the ASP dataset as y variable
• Drop the columns "SEQUENCE"，“MIC” and "LABEL" from the ASP dataset as X variable
• Separate the variable X and y as X_train,y_train,X_test,y_test
• Balance X_train and y_train by applying Borderline-SMOTE
• Use XGBoost model to select the feature and calculate the importance of features
• Using Bayesian optimization for training set parameter tuning
• Extract features with an importance greater than N times the average, and use the RF model for feature extraction
• Based on the accuracy of local optima, optimize the parameter N
• Save extracted features and corresponding importance values

Evaluating indicator：

	Accuracy	Precision	Recall	AUC
Test set	0.98	0.96	0.95	0.99
Train set	1.00	1.00	1.00	1.00

Method_3： Neural Network

steps:

• Select the columns "LABEL" from the ASP dataset as y variable
• Drop the columns "SEQUENCE"，“MIC” and "LABEL" from the ASP dataset as X variable
• Separate the variable X and y as X_train,y_train,X_test,y_test
• Balance X_train and y_train by applying Borderline-SMOTE
• create a deep learning model of neural network with linear, RELU,Linear,RELU,Linear,Sigmod structure
• train the neural network model and make a feature classification prediction
• use AHSP library to evaluate the feature importance
• Save extracted features and corresponding importance values

Evaluating indicator：

	Accuracy	Precision	Recall	AUC
Test set	0.96	0.83	0.92	0.99
Train set	1.00	1.00	1.00	1.00

Ⅴ. Regression prediction

Objective:

Using the features extracted in the previous step, perform regression fitting using MIC as the indicator to obtain the relationship between MIC and the indicator

Method_1： Regression of Random Forest

steps:

• obtain the selected features and Corresponding original ASP dataset as X variable
• standardized the X variable data by applying Standard_Scaler
• Select the columns "MIC" from the ASP dataset as y variable
• Perform logarithmic transformation on the MIC indicator represented by variable y to achieve linear fitting
• Separate the variable X and y as X_train,y_train,X_test,y_test
• Use Linear Regression model to fitting the MIC indicator
• Map the results of MIC to the 0-1 interval using an improved softmax function and provide a box plot
• save the regression evaluating metrics and figures

Evaluating indicator：

	R squared error	Squared Error	Mean Absolute Error
Random Forest	0.58979	2.14157	1.04971

Method_2：Regression of XGBoost

steps:

• obtain the selected features and Corresponding original ASP dataset as X variable
• standardized the X variable data by applying Standard_Scaler
• Select the columns "MIC" from the ASP dataset as y variable
• Perform logarithmic transformation on the MIC indicator represented by variable y to achieve linear fitting
• Separate the variable X and y as X_train,y_train,X_test,y_test
• Use Linear Regression model to fitting the MIC indicator
• Map the results of MIC to the 0-1 interval using an improved softmax function and provide a box plot
• save the regression evaluating metrics and figures

Evaluating indicator：

	R squared error	Squared Error	Mean Absolute Error
Random Forest	0.5807	2.1890	1.0927

Method_3： Nerual Network

steps:

• obtain the selected features and Corresponding original ASP dataset as X variable
• standardized the X variable data by applying Standard_Scaler
• Select the columns "MIC" from the ASP dataset as y variable
• Perform logarithmic transformation on the MIC indicator represented by variable y to achieve linear fitting
• Separate the variable X and y as X_train,y_train,X_test,y_test
• Use Linear Regression model to fitting the MIC indicator
• Map the results of MIC to the 0-1 interval using an improved softmax function and provide a box plot
• save the regression evaluating metrics and figures

Evaluating indicator：

	R squared error	Squared Error	Mean Absolute Error
Random Forest	0.5807	2.1890	1.0927