(S-013) Proteosis-Targeting Chimera Efficacy Prediction Using a Deep-Learning–QSP Model

Sungwoo Goo, PhD. – Department of Bio-AI convergence – Chungnam National University; Jina Kim, M.S. – Department of Bio-AI convergence – Chungnam National University; Soyoung Lee, Prof. – College of Pharmacy – Chungnam National University; Sangkeun Jung, Prof. – Department of Computer Science and Engineering – Chungnam National University; Jung-woo CHAE, Prof. – College of Pharmacy – Chungnam National University; Jae-mun Choi, CEO – Calici Co.; Hwi-yeol Yun, Prof. – College of Pharmacy – Chungnam National University

Objectives: This study aimed to develop and validate an integrated computational framework to quantitatively predict the efficacy of PROTAC molecules. The goal was to overcome the limitations of time-consuming experimental screening and simple binary classification models [1, 2] by accurately estimating key pharmacodynamic parameters: the half-maximal degradation concentration (DC50) and the maximal degradation level (Dmax). This predictive capability is intended to aid in the efficient prioritization and design of novel PROTAC candidates.

Methods: The framework combines deep learning with Quantitative Systems Pharmacology (QSP). First, a Convolutional Neural Network (CNN)-based model, DeepCalici [3], predicted the dissociation constants (Kd) for the binding of a PROTAC molecule to its target protein (POI) and its recruited E3 ligase, using 3D structural inputs. These predicted Kd values, along with target protein degradation rates (kdeg,P obtained from ProteomicsDB [4] or averaged) and initial estimates for other parameters (E0, kcat, α), were fed into the mechanistic QSP 'Hook Model' equations [5] to calculate initial DC50 and Dmax values. To refine these predictions, a Deep Neural Network (DNN) was trained using curated experimental data sourced from PROTAC-DB [6]. PROTAC-DB is an online database containing structural and experimental data for PROTACs; specifically, reported DC50 and Dmax values were utilized. The data was filtered to include entries with Dmax ≥ 10% and DC50 between 10 pM and 10 μM, yielding a curated set of PROTACs, their targets, and E3 ligases. This curated data, along with various chemical/biochemical features of the PROTACs and predicted Kd values, was used to train the DNN, which generated adjustment coefficients (β) for the Hook model parameters.

Results: The DeepCalici model showed reasonable performance in predicting binding affinities (R²=0.72). The integrated Deep Hook model demonstrated strong predictive performance for DC50, achieving an R² of 0.764 on the training set and 0.547 on the test set, indicating its ability to predict compound potency across orders of magnitude. However, the prediction accuracy for Dmax was significantly lower (R²=0.313 for training, R²=-0.180 for test), suggesting the model struggled to capture the variability influencing maximal degradation levels, possibly due to uncaptured experimental context. The study also confirmed that using the full PROTAC structure for Kd prediction yielded better results than using isolated components.

Conclusions: The integrated Deep Learning-QSP modeling approach provides a promising framework for quantitatively predicting PROTAC efficacy, particularly for DC50. The successful prediction of DC50 demonstrates the model's potential for in silico screening and prioritization. The limitations in Dmax prediction highlight the significant influence of experimental conditions and cellular context, which are not fully captured or standardized in the current dataset [6]. Future improvements require incorporating standardized experimental data. This hybrid strategy, combining mechanism-based QSP [5] with data-driven deep learning [3], offers a powerful tool to accelerate the discovery of effective PROTACs.

Citations: [1] Li, F., Hu, Q., Zhang, X., Sun, R., Liu, Z., Wu, S., Tian, S., Ma, X., Dai, Z., Yang, X., Gao, S., Bai, F.: Deepprotacs is a deep learning-based targeted degradation predictor for protacs. Nature Communications 13(1) (2022) https://doi.org/10.1038/s41467-022-34807-3
[2] Bartlett, D.W., Gilbert, A.M.: Translational pk-pd for targeted protein degradation. Chemical Society Reviews 51(9), 3477-3486 (2022)
[3] Choi, J.-M.: A SYSTEM FOR DISCOVERING NEW DRUG CANDIDATES AND A COMPUTER PROGRAM THAT IMPLEMENTS A PLATFORM FOR DISCOVERING NEW DRUG CANDIDATES. 1020220022030, February 2022
[4] Schmidt, T., Samaras, P., Frejno, M., Gessulat, S., Barnert, M., Kienegger, H., Krcmar, H., Schlegl, J., Ehrlich, H.-C., Aiche, S., Kuster, B., Wilhelm, M.: Proteomicsdb. Nucleic Acids Research 46(D1), 1271–1281 (2017) https://doi.org/10.1093/nar/gkx1029
[5] Haid, R., Reichel, A.: A mechanistic pharmacodynamic modeling framework for the assessment and optimization of proteolysis targeting chimeras (protacs). Pharmaceutics 15(1), 195 (2023) https://doi.org/10.3390/pharmaceutics15010195
[6] Ge, J., Li, S., Weng, G., Wang, H., Fang, M., Sun, H., Deng, Y., Hsieh, C.-Y., Li, D., Hou, T.: Protac-db 3.0: an updated database of protacs with extended pharmacokinetic parameters. Nucleic Acids Research 53(D1), 1510–1515 (2025)

Keywords: PROteolysis TArgeting Chimera, Quantitative Systems Pharmacology, Deep Learning