Update on AZ's in silico developability automated screening pipeline for biologics. How should it be used? Is it helping pipeline projects? What are the challenges and successes that we have encountered so far?

Designing developable antibodies requires an integrated approach that combines high-throughput data collection, predictive modelling, and rational framework design. We present a data-driven workflow for VHH-Fc engineering, combining wet-lab measurements with in silico predictions of stability and developability. By analysing the correlation between predicted and experimental readouts, we evaluate the utility of sequence-based models for estimating developability features.

Antibody design and property prediction are frequently hampered by data scarcity. Here, we describe DyAb, a model that addresses this issue by leveraging pair-wise representations to predict property differences. When trained on binding affinity datasets containing as few as 100 labels, DyAb generates novel antibody candidates with high binding rates, improving affinity by up to 50-fold. We discuss DyAb's general utility for therapeutic property optimisation in low data regimes.

The flexible BEAT platform enables 5 or more functional modules to be combined into a single molecule with excellent manufacturability and developability. This has been clinically validated by generating ISB 2001, a trispecific BCMA and CD38 T-cell engager advancing in the clinic to treat relapsed/refractory Multiple Myeloma, with superior efficacy, low immunogenicity in humans and good pharmacokinetics. The biophysical plus functional screening and precision engineering required to generate ISB 2001 will be described.

Developability assessment (DAS) is a core element in identification of developable drug candidates at biopharmaceutical industry. Retrospective analysis of internal programs has revealed gaps in the DAS concept, particularly in detecting critical structural-functional relationships that link critical quality attributes (CQA) findings to efficacy and safety parameters. The extensive time and material costs associated with studies to elucidate structural-functional relationships often push these activities into the Development stage, typically post-Phase I.The presentation will demonstrate how integrating machine learning, automation, focused forced degradation, and multi-attribute methodology (MAM) has enabled Novartis to establish a platform process for obtaining structural-functional relationship information during the DAS stages.

Antibody humanisation remains a pivotal strategy in the development of therapeutic antibodies, reducing immunogenicity while retaining antigen specificity and affinity. We present LifeArc case studies of the humanisation of mAbs leading to licensed candidates and those entering the clinic. Antibody engineering approaches we have employed in humanisation, including CDR grafting, framework region modification, and de novo design. By integrating these strategies, we enhance the safety profiles of therapeutic antibodies, maintain functional characteristics while enhancing human content, reducing immunogenicity, and enhancing developability.

In vitro assessments for predicting pharmacokinetics (PK) of biotherapeutics can identify risks earlier in discovery, reducing the need for extensive in vivo characterization. The clearance of antibodies with diverse sequence and biophysical characteristics was assessed in hFcRn Tg32 mice. In particular, in vitro measures of polyspecific interactions showed the highest correlations to clearance. Beyond its use in screening, the polyspecificity reagent can be applied in a flow cytometric assay that identifies, and counter selects polyspecific antibodies during both discovery and candidate optimization. Additionally, a computational approach that combines multiple in vitro measurements with a multivariate regression model was developed, improving the correlation to PK compared to any individual assessment.

We developed a high-throughput panel of low-material developability assays to assess key risks in 860 VHH-Fc constructs. The resulting dataset, covering multiple stability and interaction properties, supports predictive model building to enable sequence-based in silico screening and streamline selection of stable drug candidates.

• Foundational models for VHH representation
• Examples of ML-based methods for VHH building block property prediction
• Translating building block properties into multi-specific formats & examples of ML-based methods for multi-specific VHH developability predictions
• Addressing the data gap for complex multi-specific formats: Data generation strategies for “fit-for-purpose” and “AI-ready” data?

Unwanted immunogenicity can negatively affect the safety and efficacy of biological drugs. Computational tools can be employed during the early stages of drug discovery and development to screen libraries and prioritize drug candidates with reduced immunogenicity risks. We will introduce ImmunoScreen, AstraZeneca's in silico tool for immunogenicity assessment, within the context of the developability screening workflow. Additionally, we will discuss efforts to predict the anti-drug antibodies incidence in clinic.