Neuropsychiatry is rapidly evolving, seeking to bridge radiographic imaging and molecular pathology through machine-learning classification models. This study investigates the feasibility of leveraging class-specific autoencoders to extract latent spatial features associated with O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status in patients with glioblastoma multiforme (GBM). The framework consists of a 2D ResNet backbone that extracts spatial features from each MRI slice, followed by a Long Short-Term Memory (LSTM) layer that models inter-slice temporal dependencies. Enhancing MRI inputs with filter-derived activation maps yielded modest yet meaningful performance gains. Although preliminary, our findings illustrate how combining unsupervised representation learning with supervised classification can uncover imaging biomarkers. This approach offers a promising step toward more interpretable, data-efficient, and clinically relevant machine learning models in neuro-oncology.