Terahertz (THz) metasurfaces offer precise control over electromagnetic wave propagation, showing great promise in sensing, imaging, and communications. Recently, deep learning has accelerated metasurface design by efficiently modeling the complex relationship between structures and electromagnetic responses. This work presents two key contributions: (1) employing deep learning methods to achieve both forward prediction and inverse design of metasurfaces, thereby improving design efficiency and accuracy; and (2) realizing a sensitive response to the glutamic acid absorption peak at 1.22 THz by tuning metasurface structures, confirming its feasibility for biosensing. The results demonstrate that deep learning–driven metasurface design not only overcomes limitations of conventional approaches but also expands the application potential of THz metasurfaces in biosensing.