Felix Jost1, Clemens Giegerich2, Henrik Cordes1
1Sanofi, Translational Medicine Unit, Research Pharmacometrics, 2Sanofi, Translational Medicine Unit, Disease modeling
Objectives: Sequence based pharmaceuticals represent a significant fraction of new therapeutics. During lead identification, optimizing in vivo clearance is crucial for achieving favorable pharmacokinetic (PK) profiles. Current approaches for predicting PK parameters often rely on in vitro assays or empirical correlations, which can be resource-intensive with limited accuracy. This study aims to develop a machine-learning (ML) approach that predicts PK profiles of monoclonal antibodies (mAbs) directly from their amino acid sequences, potentially reducing the need for extensive in vivo testing while accelerating candidate selection. Methods: We compiled PK from mAbs tested in preclinical studies. The dataset covered a wide range of in vivo clearances including slow and fast clearing compounds. Each mAb was represented as a vector of pooled features derived from sequences using multiple complementary approaches. The numerical representations were transformed using self-attention layers to capture dependencies across the sequences. A physics-informed neural network (PINN) [1] was constructed using a two-compartmental PK model trained on the observed vs. predicted concentration-time data. Hyperparameter optimization was performed to determine the optimal model architecture [2]. The overall performance was evaluated using multiple metrics to facilitate comparison with related approaches in literature. [3,4,5] Results: The final model architecture was capable to adequately described the concentration-time profiles with high correlation between observed and predicted concentrations. In addition, the identified model was capable of adequately predicting key PK characteristics including volume of distribution (Vss) and a wide range of clearance (CL). The developed approach successfully predicted the in vivo clearance of approximately three quarters of the test dataset within a two-fold error margin. This performance is comparable with traditional QSAR approaches, while in addition providing the advantage of predicting complete concentration-time profiles rather than just individual PK parameters. The sequence-based approach demonstrated successful separation of fast and slow clearing mAbs, which is critical for candidate selection during drug development. Conclusion: This proof-of-concept study demonstrates that sequence-based ML methods can predict PK profiles of mAbs with sufficient accuracy. The integration of advanced sequence representation techniques with empirical PK modeling provides a powerful framework for early-stage predictions of mAbs. This approach has several potential applications in drug discovery, including compound ranking for in vivo testing, prioritization, and early human dose predictions. Challenges for this approach remain generalizing the concept for more complex antibody constructs and other sequence-based pharmaceutics. Thus, other representation methods to investigate transferability of the approach to other biotherapeutic modalities such as nanobodies, RNA, and peptides will be investigated. Overall, the shown concept represents a promising step towards in silico prediction of PK profiles having the potential of reducing animal testing while accelerating the development of new therapeutics.
1. Raissi M, set al. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J Comput Phys. 2019;378:686-707. 2. Beckers M, et al. DeepCt: Predicting pharmacokinetic concentration-time curves and compartmental models from chemical structure using deep learning. 2024. doi:10.26434/chemrxiv-2024-vg9h7 3. Bräm DS, et al. Introduction of an artificial neural network–based method for concentration-time predictions. CPT: Pharmacomet Syst Pharmacol. 2022;11:745-754. 4. Grebner C, et al. Application of Deep Neural Network Models in Drug Discovery Programs. ChemMedChem. 2021;16:3772-3786. 5. Führer F, et al. A deep neural network: mechanistic hybrid model to predict pharmacokinetics in rat. J Comput-Aided Mol Des. 2024;38:7.
Reference: PAGE 33 (2025) Abstr 11381 [www.page-meeting.org/?abstract=11381]
Poster: Methodology – AI/Machine Learning