Abstract C053: Acrivon Therapeutics’ generative ensemble model (KaiSR) accurately predicts and expands proprietary, actionable kinase-substrate relationships globally for the human kinome

Abstract Protein phosphorylation, a post-translational modification regulated by protein kinases, plays a crucial role in controlling most aspects of biological processes. However, understanding how the more than 500 human protein kinases selectively phosphorylate their specific substrates and contribute to disease-driving signaling pathways in the intact cell remain a critical challenge. The challenge becomes particularly acute in the analysis of large-scale mass spectrometry-based phosphoproteomic datasets, where most phosphorylation sites remain unassigned due to the lack of annotated kinase-substrate relationships (KSRs). To address this, Acrivon Therapeutics has developed a proprietary generative KSR prediction model (“KaiSR”) that is an ensemble of multiple transformer models fine-tuned on protein language model ESM-2. Each fine-tuned model reaches area under the precision-recall curve (AUPRC) values ranging from 83% to 92% on their respective holdout datasets. KaiSR can perform zero-shot predictions with high accuracy on unseen kinases as demonstrated through its performance on data from public and in-house kinase perturbation experiments. The predicted substrates from our model were further cross validated through in-house in vitro recombinant kinase assays on selected kinases. We have successfully applied KaiSR, a fully integrated and core component of Acrivon’s Predictive Precision Proteomics (AP3) Generative Phosphoproteomics platform, for streamlined, differentiated drug discovery, and the rational design and prioritization of novel compounds through pathway-based structure-activity relationship (SAR) in the intact cell. AP3 was recently used for designing the in-house discovered, potentially first- and best-in-class compound ACR-2316, a WEE1/PKMYT1 dual inhibitor currently in a phase 1b trial. Using KaiSR, we further confirmed the rationally AP3-designed activation of CDK1, CDK2, and PLK1 through potent and balanced inhibition of WEE1 and PKMYT1 in pre-clinical models with ACR-2316, which resulted in superior single agent efficacy of ACR-2316 via robust pro-apoptotic mitotic catastrophe. As an example of another application, we demonstrated KaiSR’s utility in identifying novel drug targets through a combined analysis of kinase activity inference based on our expanded KSRs on tumor-normal paired phosphoproteomic data from Clinical Proteomic Tumor Analysis Consortium (CPTAC) and patient matched survival data. In conclusion, Acrivon’s AP3 platform and generative AI model KaiSR uncover kinome-wide signaling networks generating unprecedented insights enabling a unique approach to target identification, streamlined drug discovery, and clinical development. Acrivon continues to leverage KaiSR and the full suite of its AP3 Generative Phosphoproteomics platform to expand and progress its proprietary pipeline of novel therapeutic candidates. Citation Format: Corey K. Xu, Bryan Thornlow, Ignacio A. Diez, Ahmed Youssef, Valentina Siino, Portia Lombardo, Zachary Best, Reina Improgo, Lei Shi, Magnus E. Jakobsson, Spring Y. Liu, Helen Nilsson, Caroline Wigerup, Joon Jung, Kristina Masson, Peter Blume-Jensen. Acrivon Therapeutics’ generative ensemble model (KaiSR) accurately predicts and expands proprietary, actionable kinase-substrate relationships globally for the human kinome [abstract]. In: Proceedings of the AACR-NCI-EORTC International Conference on Molecular Targets and Cancer Therapeutics; 2025 Oct 22-26; Boston, MA. Philadelphia (PA): AACR; Mol Cancer Ther 2025;24(10 Suppl):Abstract nr C053.