Drug discovery is undergoing a historic transition driven by artificial intelligence. The DeepMind ecosystem — AlphaFold 2 and 3, AlphaMissense, AlphaGenome— and its therapeutic arm, Isomorphic Labs, have moved the field from static structural prediction to the rational design of ligands; the Isomorphic Drug Design Engine (IsoDDE), unveiled in February 2026, more than doubles the accuracy of AlphaFold 3 on out-of-distribution protein–ligand complexes and identifies cryptic and allosteric pockets from amino-acid sequence alone. In parallel, Yann LeCun’s critique of autoregressive models and his bet on large world models (LWMs) built on joint embedding predictive architectures (JEPA) —channelled since 2026 through AMI Labs— offers the conceptual framework needed to overcome the “snapshot” limitation of current structural predictors. In this Perspective I argue that, in HER2-positive breast cancer —where resistance to antibody–drug conjugates such as trastuzumab deruxtecan, to tyrosine-kinase inhibitors and to bispecific antibodies is governed by conformational dynamics and clonal evolution—, the convergence between the reasoning of large language models (LLMs) and the physical-chemical simulation capacity of LWMs could, between 2027 and 2030, enable the design of a new generation of therapeutic agents. I explicitly distinguish between today’s demonstrated capabilities, ongoing developments and prospective projections, and I propose a roadmap for Latin American academic clinical oncology to take part in this transformation as a co-developer rather than a late adopter.