XGBMUT: Predicting the Functional Impact of Missense Mutations Using an Extreme Gradient Boost Classifier

Abstract

Millions of new mutations have been discovered largely due to advancements in genome projects, but characterizing their effects through traditional wet-lab experiments remains labor-intensive and time-consuming. Functional prediction algorithms offer a solution by enabling the efficient screening of mutations, thereby saving time and resources. The objective of this study was to develop a competitive algorithm for predicting the functional impact of missense mutations. A unified database and substitution matrices containing predictor variables specifically for missense mutations were initially constructed. Subsequently, values for the predictor variables were collected from the training and test sets derived from the ClinVar and HumsaVar databases. A series of supervised machine learning classifiers were then trained, and their performance was evaluated using the test set. The best-performing model was additionally compared against ten currently available functional prediction algorithms. The proposed algorithm, XGBMut, demonstrates exceptional accuracy in classifying missense mutations while also exhibiting a competitive performance. Additionally, a user-friendly graphical interface was developed to enhance accessibility for professionals in various fields. Unlike most existing methods, XGBMut eliminates the need for a web server dependency and the installation of third-party software, making it a more versatile tool for users.

Figure 1

Workflow for data processing, model training, validation, and functioning of the XGBMut classifier. (A) Workflow for XGBMut construction, validation, and benchmarking. The ClinVar and HumsaVar databases were initially processed to remove duplicates and select valid missense mutations for constructing the training and test sets, respectively. These cleaned data sets were then submitted to a function designed to automatically extract local and global predictor variables for the corresponding mutations, based on information from the previously prepared unified database and substitution matrices. The training set was used to develop the XGBMut model, while the test set, containing unseen mutations, was independently evaluated to assess the model’s performance. Finally, the performance of well-established methods for predicting the functional effects of missense mutations was compared to that of XGBMut using the test set, ensuring the model’s viability. (B) Workflow for XGBMut functioning. For unlabeled mutations, an initial viability check ensures that only valid mutations proceed to predict the variable extraction. Invalid mutations are filtered out during this step and saved in a separate file for user reference. Valid mutations were then submitted to a function designed to automatically extract local and global predictor variables for the corresponding mutations based on information from the previously prepared unified database and substitution matrices. Once predictor variables are fully extracted from the submitted data, they are input into the XGBMut model, which predicts the deleteriousness probabilities for each mutation. These predictions are then saved to an output file. Arrows represent the data flow, while icons illustrate the computational processing and model implementation steps.

Figure 2

Performance of XGBMut on the test set and its comparison with currently available functional prediction algorithms. (A) Confusion matrix calculated for the final model (XGBMut). TN: true negatives; TP: true positives; FN: false negatives; FP: false positives. (B) ROC curve calculated for XGBMut and ten functional prediction algorithms within the test set. The output of VariPred is not shown, as it is provided as a class label rather than a probability score, making the AUC-ROC calculation infeasible. (C) Blue bars indicate the accuracy of each algorithm, while the dashed black lines represent their corresponding coverage within the test set, i.e., the percentage of total observations (mutations) effectively analyzed by each algorithm.

Figure 3

Graphical user interface and command-line interface of the XGBMut software. (A) Graphical user interface. The main menu of the program is on the left, while the analysis loading menu is on the right. (B) Command-line interface. The image corresponds to the Windows version of the interfaces.