Four-dimensional computed tomography (4DCT) is required in lung radiotherapy treatment planning to track tumor motion for more accurate dose coverage. However, such acquisitions expose the patient to more radiation than a standard CT protocol. In previous works, we demonstrated the feasibility of patient-specific 4DCT generation from static 3DCT images using generative adversarial networks (GAN) conditioned on the actual patient's respiratory amplitude. Synthetic motion was achieved globally in the lung, but with yet unsatisfactory accuracy at the tumor level. This work addresses this issue by better taking into account tumor motion through tumor awareness. We condition the image-to-image GAN architecture by a static 3DCT image, a respiratory amplitude and a further condition on the tumor segmentation mask. We train the model under a combined segmentation and weighted L1 objective. We performed experiments on synthetic phantoms, where we demonstrate better tumor motion synthesis both qualitatively and quantitatively, and show preliminary results on clinical data.