In this study, a deep learning framework for tumor prognostic risk assessment was constructed around the joint modeling of dynamic genomic signals associated with target genes and structured clinical features. The genome input was GRO-seq data from Huh7 Mock cells and ADAMTS5e overexpressing cells, and the sequencing quality was maintained above Q38 overall and close to Q40 locally. groHMM identification results show that the density of the transcript part reaches 0.7197, and the FivePrimeFP signal is 0.4125 near the transcription start site. Differential analysis detected 312 upregulated and 145 downregulated de novo transcripts. Functional enrichment mainly involved apoptosis regulation, cell junction assembly, adhesion junction organization and receptor inhibitory activity. Site level analysis further confirmed that RUNX1 and LOC100507412 had definite changes, while the global RNA Pol II pause index was generally stable. The above genomic descriptors were co-coded with clinical variables such as age, stage, pathological index, treatment record and follow-up outcome, and the prognostic risk score was output after cross-modal fusion. This framework emphasizes dynamic feature compression, joint representation learning and reproducible calculation process, which provides a computational basis for subsequent tumor prognosis stratification and model expansion.
