Antibodies are molecules that exhibit remarkable flexibility, a property that arises from several structural features, including hinge regions, elbow linkers, interdomain interfaces between Ig-fold domain pairs, and the loops within the paratope. Collectively, these elements contribute to the ability of antibodies to adopt various conformations, thereby enabling them to recognise and bind to a wide array of antigens with high specificity. In this study, we demonstrated that, in solution, the binding-competent structure of antibodies is typically the dominant conformation, even though this is often distinct from the structures observed in their unbound state. This suggests that antibodies inherently adopt a structural ensemble in which the binding-competent form is energetically favorable when free in solution. These findings have significant implications for understanding antibody behavior and designing tools to predict their structures computationally. Specifically, they reveal both opportunities and challenges for AI-driven antibody structure prediction, as accurate modelling must account for the dynamic and flexible nature of these biomolecules. The ability of AI models to incorporate the conformational adaptability of antibodies will be critical for improving predictions, particularly when resolving structures relevant to antigen binding. This work underscores the need for innovative approaches that leverage both experimental data and computational advancements to address the complexity of antibody flexibility in structural biology.

The physiological behaviour of peptides is intrinsically linked to their folding, that is a critical aspect for their activity and it is hard to replicate in silico, accurately. One of the main challenges in achieving this is the selection of appropriate system parameters. This study aims to identify the optimal combination of force field and solvation model for molecular dynamics simulations of peptides, evaluating the impact of different combinations on predicting peptide secondary structures. A systematic evaluation of eight peptides, comprising two helical peptides, three β-hairpins, and three intrinsically disordered peptides, was carried out using accelerated molecular dynamics (aMD) simulations. These simulations employed three force fields (ff99SB, ff14SB, and ff19SB) in combination with two explicit solvation models (TIP3P and OPC) and they were initiated from both extended and misfolded conformations. The results show which could be the best option for each class of simulated peptides.

Recent advancements in deep learning have significantly improved the success rate of computational protein design tools, making them attractive for integration into drug discovery pipelines. However, the rapid pace at which new tools are being published makes it challenging to choose the best combination of tools and settings, i.e. design strategy. To support both our internal design tasks and benchmark new tools, we developed a highly scalable and modular platform, consisting in a series of modules for backbone and sequence generation, in-silico structure prediction for validation, and developability predictions. The modular architecture makes it easy to onboard new tools and the platform can generate up to 2.5 M designs per day. To evaluate the best strategies for structure-based de novo binder design, we used the platform to perform a large-scale benchmark of two backbone design tools and six sequence design tools, generating more than 25M designs in total on 12 different targets. The benchmark includes a systematic exploration of therapeutically relevant protein lengths (15-95 aa), the impact of hotspot information, and the predicted biophysical properties of designed sequences with interesting learnings that will be shared at the conference.

This talk presents the application of pharmacophore modeling to protein-protein interactions, specifically antibody-antigen (Ab:Ag) interfaces. The number and types of pharmacophore features at Ab:Ag interfaces are surveyed over a diverse collection of Ab:Ag complexes. The Ab:Ag pharmacophore queries are used to perform searches on databases of 3D antibody structures.

The results show that pharmacophore queries can be used to efficiently search and pose 3D antibody structures and that pharmacophore searching represents a high-throughput virtual screening (HTVS) alternative to protein-protein docking.

A. Kötter¹, E. Kober², E. Lorent^1
1BioAIM, Digital R&D Large Molecule Research, Sanofi-Aventis Deutschland GmbH, Industriepark Höchst, D-65926 Frankfurt am Main, Germany
²Large Molecule Research, Sanofi-Aventis Deutschland GmbH, Industriepark Höchst, D-65926 Frankfurt am Main, Germany
³Large Molecule Research, Sanofi, Technologiepark-Zwijnarrde, B-9052 Gent, Belgium

Developability properties of biologics drug solutions represent an important optimization parameter during drug discovery. Properties like the diffusion interaction parameter kD and the aggregation onset temperature agg measured at low concentration are known to be correlated to high concentration colloidal properties like opalescence or viscosity of mAb solutions. Thus, predictive ML models for properties like kD and agg can help to de-risk the selection of molecules during the discovery process. Here, we show that simple ML models based on sequence-based descriptors of multispecific VHH molecules can accurately predict kD and Tagg, despite the complexity of this modality.

A.K., E.K, E.L. are employees of Sanofi and may hold shares and/or stock options in the company.

Antibodies are an important class of medicines, whose efficacy is driven by specific target binding. Given the therapeutic relevance, there have been multiple attempts to predict antibody-antigen binding affinity computationally. I will discuss our findings on how training data influences the success and selection of machine learning strategies to tackle this challenge, ranging from antigen-specific to generalizable and zero-shot affinity prediction.

Significant advancements have been achieved in predicting the developability properties of antibodies through AI/ML methods, enabling the in silico design of extensive sequence sets. Although these techniques are effective for standard monospecific antibodies, they frequently fall short when applied to more complex next-generation antibodies, such as multispecifics and antibody-drug conjugates (ADCs). This presentation will highlight key lessons learned and specific applications of physico-chemical property prediction strategies for evaluating and optimizing bispecifics and ADCs.