Indrasis Dasgupta, Abhay Nath, Sanskruti Patel, Shovanlal Gayen
Matrix metalloproteinase-12, also known as macrophage metalloelastase, plays a critical role in extracellular matrix remodelling and is strongly implicated in various pathological conditions, including cancer, inflammatory disorders, cardiovascular diseases, and pulmonary dysfunction. Consequently, the identification of potent and selective MMP-12 inhibitors has become an important objective in drug discovery. The current study presents a computational framework that integrates chemical space analysis, machine learning, and deep learning to predict MMP-12 inhibitory activity using a curated data set of 481 compounds. A novel deep learning architecture, termed the Multi-Scale Residual Attentive Regression Network, is proposed, and this architecture integrates multiscale convolutional feature extraction, dilated gated convolutions, and attention mechanisms to detect complex nonlinear structure-activity relationships. The proposed model demonstrated superior predictive ability with a test R2 of 0.924, significantly outperforming other predictive models. The proposed model performance was also evaluated using scaffold-based splitting to assess both predictive ability and generalizability. Model interpretability was investigated using SHAP analysis, which identified key fingerprint bits contributing to MMP-12 inhibitory potency. Furthermore, the MMP-12 inhibitor prediction tool (MMP-12i predictor) was developed based on the proposed model. In addition, selectivity models were also developed to discriminate selective MMP-12 inhibitors from nonselective ones, revealing structural features associated with isoform selectivity. Overall, this framework provides a robust and interpretable strategy for predicting MMP-12 inhibitory activity and identifying molecular determinants of selectivity. The current study is limited by a moderate data set size and the lack of experimental validation, both of which should be addressed in future work.