xLSTM-based deep learning model for predicting anticancer Activity of pyrano[3,2-a]phenazine hybrid derivatives:...

Finding novel treatments is a long and complex process involving extensive drug discovery, experimental testing, and preclinical research. Computational quantitative structure-activity relationship (QSAR) modeling can support this process by enabling rapid prediction of biological activity and identification of molecular features associated with activity. This study presents a proof-of-concept deep-learning-based QSAR framework for predicting the biological activity of pyrano[3,2-a]phenazine hybrid derivatives. The proposed framework combines an Extended Long Short-Term Memory (xLSTM) architecture with a hybrid APOADAM optimization strategy, in which Arctic Puffin Optimization (APO) is used for molecular descriptor subset selection and the Adam optimizer is used for xLSTM parameter optimization. Molecular descriptors were generated using the PaDEL-Descriptor tool and subjected to preprocessing, mutual-information-based ranking, and optimizer-based feature selection. The framework was evaluated for four cancer cell lines: HCT116 (colon), MCF7 (breast), HepG2 (liver), and A549 (lung). Model development was performed using 25 compounds with five-fold cross-validation, while 12 compounds were retained as a completely independent test set. On the independent test sets, the APOADAM-xLSTM model achieved RMSE values of 0.2514, 0.1712, 0.2451, and 0.3314 for HCT116, MCF7, HepG2, and A549, respectively. The corresponding Pearson correlation coefficients were 0.8252, 0.9577, 0.9314, and 0.8249. APOADAM also identified cell-line-dependent molecular descriptor subsets, indicating that different molecular features may contribute to biological activity across the investigated cell lines. Additional validation using out-of-fold predictions included Y-scrambling and applicability-domain analysis. The results provided evidence against chance correlations for three cell lines, while Williams-plot analysis showed that most OOF compounds were within the defined applicability domain. Overall, the results demonstrate the potential of the xLSTM-APOADAM framework for QSAR modeling under a constrained-data setting, while the findings remain preliminary and specific to the investigated chemical series. Further validation using larger and more chemically diverse datasets, broader benchmarking, and prospective experimental evaluation is warranted.