This study investigates the potential of radiomics combined with deep learning to predict the Programmed Death-Ligand 1 (PD-L1) biomarker in Non-Small Cell Lung Cancer (NSCLC) patients using CT scans. Traditional biomarker testing is invasive and costly, creating a need for non-invasive alternatives. We extracted quantitative features from CT images and applied various machine learning models, including Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Networks (GRU). Initial results showed that the Transformer model achieved better performance, with a Test Mean Squared Error (MSE) of 38.16, compared to 294.59 for ANN and 127.12 for CNN. Further improvements were made with a Complex Transformer architecture, which, after 1000 epochs, reduced the MSE to 17.41. With early stopping at epoch 261, the final model achieved an MSE of 18.25. These findings suggest that radiomic features, combined with advanced deep learning techniques, offer a promising non-invasive alternative for predicting PD-L1 expression, potentially reducing healthcare costs and improving personalized cancer treatment.