Transcription factors (TFs) regulate chromatin structure, acting as natural archi tectural organizers and promising drug targets. To address the scarcity of known structure-related TFs, this study developed an AI model, AlphaTF-struct, to screen candidate structure-related TFs from large-scale unannotated data, with SIX5 identified as the core research target. The model constructs features based on the principle of DNA-binding colocalization of functionally similar TFs: ChIP-seq peaks are mapped to a unified genome binning framework, and their genome-wide binding distributions are represented by sparse vectors, with the feature system further refined by overlapping signatures of known architectural TFs. A two-stage positive-unlabeled (PU) learning strategy is adopted for pre cise screening. Considering heterogeneous TF binding distributions across cell lines, a distribution-corrected weighting method is designed to eliminate screen ing bias and enable cross-cell-line application.