Benchmarking Machine Learning Models for HIV-1 Protease Inhibitor Resistance Prediction: Impact of Data Set Construction and Feature Representation | JoVE Visualize

Area of Science:

* Computational biology and bioinformatics
* Machine learning in drug discovery
* Viral infectious diseases and resistance mechanisms

Background:

* Drug resistance in viral infections, particularly HIV-1, poses a major global health threat.
* Machine learning (ML) models are increasingly used to predict antiviral drug resistance from genomic data.
* HIV-1 protease inhibitors (PIs) are a critical treatment, making resistance prediction vital.

Purpose of the Study:

* To systematically evaluate existing ML models for predicting HIV-1 PI resistance.
* To assess the impact of different data preprocessing strategies on model performance.
* To propose and validate a novel, stringent approach for assessing model generalizability.

Main Methods:

* Compared various ML models (neural networks, Random Forest, KNN, logistic regression) on three distinct HIV-1 protease datasets.
* Investigated different preprocessing techniques, including ambiguous sequence handling and data expansion.
* Utilized zScales physicochemical descriptors and Rosetta energy terms as model features.
* Implemented a clustering-based validation approach for robust generalizability assessment.

Main Results:

* Data expansion preprocessing artificially inflates performance metrics by introducing redundancy.
* Clustering-based validation provides a more stringent and reliable assessment of model generalizability.
* Physicochemically informed logistic regression models (zScales LR, Rosetta LR) achieved performance comparable to complex neural networks.
* zScales LR demonstrated superior computational efficiency and interpretability over Rosetta LR.
* Mutual information analysis identified distinct resistance mechanisms: zScales highlighted specific hotspots, while Rosetta revealed interconnected energetic networks.

Conclusions:

* Data set construction and preprocessing choices critically influence apparent ML model performance in resistance prediction.
* Well-chosen physicochemical features can yield accurate and interpretable HIV-1 PI resistance models, rivaling complex neural networks.
* The proposed clustering-based validation and physicochemical feature representation offer a robust framework for developing clinically relevant resistance prediction tools.