Abstract A025: PicoGen: A structure-grounded generative AI model for drugging undruggable targets, including the TEAD–YAP axis

Abstract Despite major advances in structure-enabled and AI-driven drug discovery (AIDD), progress against protein–protein interactions (PPIs) remains a critical bottleneck—especially in oncology, where transcriptional co-activators and signaling hubs often lack defined pockets or tool compounds. A prominent example is TEAD–YAP, a central effector of the Hippo pathway and a validated oncogenic driver in mesothelioma, basal cell carcinoma, and squamous tumors. Despite extensive evidence linking this complex to tumorigenesis, it remains refractory to conventional drug design due to its shallow, featureless interface. Compounding this challenge is the field-wide scarcity of high-resolution ligand–protein structural data, a core bottleneck for training small molecule-focused AI models. Fewer than 12,000 unique ligand-bound proteins exist in the Protein Data Bank (PDB), a fraction of what is needed to train generative machine-learning (ML) models. Structural data is particularly sparse for targets like transcription factors, where ligands are rare or absent. As a result, contemporary AI models trained on legacy ligand-receptor datasets are brittle and limited in their utility against novel, undrugged targets. To address these dual limitations of data scarcity in AIDD and PPI intractability, we developed PicoGen, a foundational generative AI model trained directly on PPI surfaces. Rather than relying on ligand-bound complexes, PicoGen learns transferable features that enable structure-grounded small molecule generation even for completely unliganded targets. It benefits from a carefully curated and dramatically expanded training set that is over 50-fold larger than traditional ligand-based datasets. Free from screening library biases or human design constraints, it operates beyond pre-synthesized chemical space, enabling novel solutions for previously undruggable sites. It also accurately rediscovers known PPI-inhibitor interactions when blinded to ligand information, validating its ability to recover features relevant to drug binding from raw interface geometry. As a proof of concept, we applied PicoGen to the TEAD–YAP interface. The model generated ligands targeting TEAD to disrupt YAP binding at a site distinct from the palmitoylation pocket. Synthesized compounds were validated by differential scanning fluorimetry (DSF), confirming direct TEAD engagement. Binding was retained in a TEAD mutant lacking a functional palmitoylation pocket, confirming alternative site engagement. Structural and functional evaluation of these hits is ongoing to further elucidate their therapeutic potential. Beyond TEAD–YAP, PicoGen supports rapid hit generation across cryptic, uncharacterized or structurally complex targets, producing high-confidence ligands for historically challenging targets including MYC-MAX, MAML1 and STAT3 and other targets relevant to immune evasion. Together, these findings establish PicoGen as a powerful, structure-native, generative platform capable of unlocking elusive therapeutic targets in cancer and across disease areas. Citation Format: Nicholas Hamilton, Asanga Bandara, Steff De Graef, Stephen Weeks, Ali H. Munawar. PicoGen: A structure-grounded generative AI model for drugging undruggable targets, including the TEAD–YAP axis [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A025.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.