<p>Evolution has optimized proteins over time by the incorporation of precise and context-specific amino acid substitutions adapted to structural and functional demands. We have reconceptualized this principle using deep learning to engineer monoclonal antibodies (mAbs) targeting immune checkpoints PD-1 and LAG-3. These two checkpoints are targeted synergistically in combination immunotherapy to minimize cancer cell evasion. From the established antibodies, the best set was selected based on their clinical validation. These served as templates to improve binding affinity and therapeutic potential in the heterogeneous tumor microenvironment. To guide antibody design, we formulated inverse modeling pipeline using message passing graph neural network for protein sequence design given a fixed backbone structure. This led to the prediction of functionally viable substitutions at the receptor-antibody interface. Resulting variant models were filtered based on physicochemical accuracy, evolutionary feasibility, empirical validation, geometric complementarity and machine learning guided mutation prediction, ensuring structural integrity and enhanced performance. In addition, thermostability and immunogenicity analyses of the filtered ones were carried out. Ultimately, the top candidates were subjected to molecular dynamic (MD) simulations leading to post simulation trajectory analysis including stability, interaction and energy decomposition analysis. After a robust computational evaluation, seven variants exhibited improved network stability and superior binding as compared to their respective references. Moreover, we have also added negative control to reinforce the novelty and importance of our framework. Our results establish a robust and scalable framework to design ICIs and underscores potential leads having improved binding, concertedly targeting PD-1 and LAG-3, paving the path for next-generation immunotherapy.</p> Graphical abstract <p></p>