Researchers at the Institute for Research in Biomedicine in Barcelona have successfully demonstrated a novel artificial intelligence methodology for designing selective therapeutic compounds. Led by Dr. Patrick Aloy of the Structural Bioinformatics and Network Biology Lab, the team engineered a predictive and generative AI system that identifies new chemical entities based on desired cellular outcomes rather than predefined molecular targets. Traditional drug discovery typically relies on isolating a specific protein to modulate, a strategy that fails when disease mechanisms remain unclear or lack defined targets. The Barcelona team inverted this paradigm by adopting phenotypic discovery. Instead of targeting a molecule, the system was trained to generate compounds that produce a specific observable effect in designated cell populations while sparing others. To construct the foundational database, the researchers screened over eleven thousand existing chemical compounds across eight distinct cellular models, including six pancreatic cancer cell lines and two control lines. Bioactivity data from these tests trained machine learning models that predicted cellular responses with significantly higher accuracy than traditional structure-based similarity assessments. These predictive models were subsequently integrated into a generative AI framework designed to propose novel molecular candidates optimized for dual selectivity: high efficacy against target cells and minimal impact on non-target or healthy cell profiles. Experimental validation confirmed the system’s efficacy. Multiple AI-designed molecules successfully replicated their intended phenotypic functions, demonstrating selective activity against specific cancer lines while showing reduced effects on control cells. Crucially, these compounds outperformed those identified through conventional high-throughput screening and exhibited structurally novel scaffolds distinct from existing pharmaceutical libraries. Although the research remains in the early discovery phase, the methodology establishes a scalable pathway for accelerated compound identification in complex disease states. By decoupling drug design from the requirement of a known molecular target, this approach expands therapeutic development possibilities for pathologies with poorly characterized biological mechanisms. The integration of predictive biology with generative chemistry marks a significant advancement in computational drug discovery, offering a more targeted and efficient alternative to traditional screening workflows.