Optimisation & Developability

Multispecific antibodies present unique engineering challenges driven by geometry, flexibility, and format diversity, often in data-limited settings. This talk describes practical and scalable structural modeling strategies to guide rational engineering multispecific antibodies across different formats. Automated model construction and conformational sampling are used to analyze avidity, binding geometry, and surface properties, enabling design decisions that improve functional performance while addressing developability considerations early in discovery.

Biparatopic antibodies are a subclass of bispecific antibodies which can simultaneously engage two epitopes on a single antigen species. This property unlocks novel target engagement and pharmacology, which we explore using in vitro binding and functional assays. Mathematical models were established which accurately predict the behaviour of these antibodies, requiring only SPR-derived affinities as inputs, potentially streamlining their optimisation during lead discovery.

Small interfering RNAs (siRNAs) are increasingly developed as antibody–oligonucleotide conjugates (AOCs) to improve targeted delivery. Chemical modifications and antibody conjugation, however, pose major challenges for all-atom modelling. Here we summarize computational strategies for AOC–siRNA systems, with an emphasis on all-atom molecular dynamics simulations. We highlight available force-field options and structure-building workflows and note the lack of dedicated software for conjugation chemistries.

Combinatorial approaches were useful to achieve multi-level optimization of Interleukin-2-derived immunomodulators, ideally suited for particular applications. Functional fine-tuning, through engineering of individual interfaces with each component of the multi-chain receptor system, resulted in responses biased to the desired immunological outcome. Phage display-based filtering steps prevented distal effects on non-targeted interfaces, and removed typical cytokine intrinsic biophysical liabilities, giving rise to a set of highly developable candidates with potent anti-tumor activity.

This talk highlights how we leverage large-scale datasets together with domain and data science expertise to analyse and understand the drivers of CHO expression performance of complex biologics. By linking sequence- and format-level features to downstream outcomes, we establish an approach that integrates manufacturability considerations early into lead selection, which leads to improved robustness, reduced uncertainty, and overall strengthened R&D productivity.

To date most foundation model work in the protein design space has focused on sequence and structure. Yet for therapeutics applications, designed molecules must also demonstrate a suite of favorable biophysical properties to be “developable.” This makes an antibody developability foundation model a highly sought-after goal, yet recent publications and public competitions have shown that existing public datasets are insufficient. At BigHat we have recently pointed our platform, which leverages synthetic biology and active learning to rapidly generate data and improve AI/ML models, at this problem. BigHat’s foundational developability model lets us provide quantitative answers to the question of how much and what kind of data is needed to improve multiple developability properties for diverse therapeutic formats. We present several case studies leveraging BigHat’s developability model to rapidly improve key developability properties for diverse antibody formats.

MeLLOw is an active learning framework for multi-property antibody sequence optimisation. It combines diverse protein encodings (BLOSUM, language models, and structure-aware neural networks) with an ensemble of more than 21 ML models and evolutionary search to select variants for experimental validation. Training on both experimental and in silico data, MeLLOw closes the loop between computation and the wet lab. A web-based GUI enables non-computational scientists to configure, launch, and analyse campaigns without writing code.

30 years’ cumulative experience has allowed the early identification of manufacturable standard mAb therapeutics. To allow a similar capability for more complex formats we have performed a wide range of developability assays (17 methods, 27 variables), coupled with long-term (5˚C) and accelerated (25˚C and 45 ˚C) stability studies on 44 next-generation mAb scaffolds comprising canonical reference mAbs, multispecifics, and Fc-fusions variants of five well-characterised mAb therapeutics.