👤 Assif Assad

The viability of cells and the integrity of the genome depend on the detection and repair of damaged DNA through intricate mechanisms. Cancer treatment employs chemotherapy or radiation therapy to eliminate neoplastic cells by causing substantial damage to their DNA. In many cases, improved DNA repair mechanisms lead to resistance to these medicines; therefore, it is essential to expand efforts to develop drugs that can sensitise cells to these treatments by inhibiting the DNA repair process. Multiple studies have demonstrated a correlation between the overexpression of Apurinic/Apyrimidinic Endonuclease (APE1), the primary mammalian enzyme responsible for excising apurinic or apyrimidinic sites in DNA, and the resistance of cells to cancer therapies; in contrast, APE1 downregulation increases cellular susceptibility to DNA-damaging agents. Thus, the effectiveness of existing therapies can be improved by promoting the targeted sensitization of cancer cells while protecting healthy cells. The current study aims to employ explainable artificial intelligence (XAI) to enhance the accuracy and reliability of machine learning models for the prediction of APE1 inhibitors. Various ML-based regression models are employed to predict the pIC50 value of different medicines. Bayesian optimization and the Permutation Feature Importance (PFI) approach are employed to determine the best hyperparameters of machine learning models and to discover the most significant features for recognizing drug candidates that target APE1 enzymes, respectively. To acquire comprehensive elucidations for the predictive models in our research, two XAI methodologies, namely SHAP and LIME, are used. The SHAP analysis reveals that the features 'C1SP2' and 'ASP-2' are essential in influencing the model's predictions. The SHAP values demonstrate variability for features such as 'maxHBint2' and 'GATS1s,' signifying that their impact is dependent on specific instances within the dataset. The LIME study corroborates these findings, demonstrating that 'C1SP2' and 'ASP-2' are the most significant positive contributors, whereas features like 'SHCHnX,' 'nHdCH2,' and 'GATS1s' result in a decrease in the predicted values. Due to the limited sample size of the APE1 dataset, direct training on this dataset posed challenges in model generalization and reliability. To overcome this limitation, the BACE-1 dataset is leveraged for model training, enabling the ML models to learn from a more extensive and diverse chemical space. Among the tested algorithms, XGBoost demonstrated superior predictive performance, achieving R Show less

Single-stranded DNA (ssDNA) containing four guanine repeats can form G-quadruplex (G4) structures. While cellular proteins and small molecules can bind G4s, it has been difficult to broadly assess their DNA-binding specificity. Here, we use custom DNA microarrays to examine the binding specificities of proteins, small molecules, and antibodies across ∼15,000 potential G4 structures. Molecules used include fluorescently labeled pyridostatin (Cy5-PDS, a small molecule), BG4 (Cy5-BG4, a G4-specific antibody), and eight proteins (GST-tagged nucleolin, IGF2, CNBP, FANCJ, PIF1, BLM, DHX36, and WRN). Cy5-PDS and Cy5-BG4 selectively bind sequences known to form G4s, confirming their formation on the microarrays. Cy5-PDS binding decreased when G4 formation was inhibited using lithium or when ssDNA features on the microarray were made double-stranded. Similar conditions inhibited the binding of all other molecules except for CNBP and PIF1. We report that proteins have different G4-binding preferences suggesting unique cellular functions. Finally, competition experiments are used to assess the binding specificity of an unlabeled small molecule, revealing the structural features in the G4 required to achieve selectivity. These data demonstrate that the microarray platform can be used to assess the binding preferences of molecules to G4s on a broad scale, helping to understand the properties that govern molecular recognition. Show less