The sequences are designed via a two-step approach. First, a novel transformer-based cross-attention GNN is employed and trained in multiple datasets. The architecture enables explicit drug-target interactions due to wiring of ligand and protein representations. The related manuscript is currently under review. During inference, we map the atom-residue interactions and identify high-attention binding pockets learned by the model. We use the identified residues to design complexes that could bind strongly to these specific residues. Three complementary strategies are applied: (i) attention-gradient following to trace high-attention paths, (ii) attention-weighted mutations that preferentially alter low-attention residues while preserving high-attention positions, and (iii) randomized sampling of high-attention regions to enhance diversity. Each candidate is evaluated with a composite score combining attention quality and consistency per residue. Structural proxies (ipSAE and iPTM), alongside an estimated ΔG and a confidence metric are also used.