Generative AI-Driven Interpolation of Cross-Family Kinase Space: Discovery of Dual-Targeting Degraders of IRAK4 and TYK2

Abstract Bridging structural discontinuities between disparate protein families is a fundamental challenge in multi-target design. We present a generative artificial intelligence (AI) platform for the non-linear interpolation of cross-family kinase space to simultaneously modulate innate (interleukin-1 receptor-associated kinase 4, IRAK4) and adaptive (tyrosine kinase 2, TYK2) immunity. Utilizing a variational autoencoder−recurrent neural network (VAE−RNN) model within the POLYGON framework, we successfully navigated high-dimensional latent space through an iterative “sampling−scoring−refocusing” strategy. This approach bypassed the geometric limitations of traditional pharmacophore merging, leading to the rational discovery of a novel spirocyclic scaffold with significant structural novelty (Tanimoto similarity <0.45). Integration of WaterMap-based optimization and proteolysis-targeting chimera (PROTAC) conjugation yielded C-8-014, a potent dual-target degrader achieving nanomolar potency and efficient degradation of both kinases. This study demonstrates a validated closed-loop methodology for navigating “unreachable” regions of the kinome, providing a paradigm-shifting approach for cross-family chemical design and a powerful molecular tool for chemical biology.