Deep generative models have made significant advances in drug discovery. However, generating drug-like molecules directly from protein sequences remains challenging due to limited feature extraction, weak contextual understanding, and suboptimal molecular activity modeling. To address these issues, we propose DrugGAN-MSM, a GAN-based generative model driven by protein sequences. Masked training and self-attention mechanisms are employed to enhance the extraction of global and local sequence features. Furthermore, multi-objective optimization is integrated into the discriminator to improve both chemical validity and biological activity of the generated molecules. Experimental results show that DrugGAN-MSM outperforms existing methods in molecular generation quality and activity optimization, demonstrating its effectiveness for data-driven drug design.