The design of multispecific protein therapeutics presents unique challenges that remain largely unaddressed by current computational approaches. We discuss critical data gaps in this field and present strategic approaches for generating fit-for-purpose datasets specifically tailored for multispecifics. Through practical examples and case studies, we demonstrate how targeted computational and machine-learning strategies can support the optimization of next-generation multispecific therapeutics.

Protein-protein interactions are fundamental to most biological processes but remain difficult to design due to their complex structural determinants. We present BindCraft, an open-source and automated platform for de novo protein binder design that achieves high-affinity binding without experimental optimization or prior structural information. BindCraft enables the generation of functional binders against a broad spectrum of targets, including receptors, allergens, and nucleases. By integrating state-of-the-art structure prediction and design modules, BindCraft advances the field toward a “one design one binder” paradigm, establishing a generalizable framework with broad implications for therapeutic, diagnostic, and biotechnological applications.

Deep learning has accelerated protein design, but most existing methods are restricted to generating protein backbone coordinates and often neglect interactions with other biomolecules. I will present the next generation of protein design methods that include side-chain coordinates for design of more complex biomolecular function. Finally, I will show a series of applications of these algorithms to design of experimentally characterized functional proteins.

While AI has been used in bioinformatics since the early 1990s for problems such as protein secondary structure prediction, advances in AI over the last 5 years have revolutionized many areas of life from animation to bioinformatics. These changes have been driven by approaches such as protein language models and generative models used in AlphaFold for protein structure prediction. There have been several publications that use such approaches for 'ab initio' antibody design, but I for one remain skeptical. Nonetheless, there are clear applications for modern AI techniques around antibody developability and improving candidate antibody-based drugs.

Antibodies are promising protein therapeutics, but successful development requires meeting strict developability criteria. We present TherAbDesign, a machine learning method that evaluates and optimizes antibodies based on sequence alone, proposing modifications that mimic the biophysical properties of successful therapeutics. This approach circumvents computationally expensive structure prediction and physics-based calculations. We show that this method improves known developability liabilities, such as viscosity, without explicitly modeling their mechanism of action.

Recent advances in protein language models (PLMs) have demonstrated strong performance on structure and function prediction. To evaluate their performance in nanobody-related tasks, we developed a comprehensive benchmark suite, NbBench. Benchmarking of eleven models revealed that antibody language models excel in antigen-related tasks, while thermostability and affinity-related talks remain challenging across all models. We further discuss how PLMs and their benchmarks could impact on antibody research.

Infinimmune presents GLIMPSE, an antibody language model trained on proprietary human data that achieves state-of-the-art performance. We used GLIMPSE within our lab-in-a-loop platform to engineer an anti-IL-13 antibody, enhancing its drug-like properties including potency, extended half-life, affinity, stability, and manufacturability via liability removal. This work demonstrates the practical application of language models for optimizing therapeutics while maintaining their humanness, moving beyond typical proof-of-concept studies.

Supervised machine learning in antibody discovery relies on positive and negative examples, making dataset composition crucial for performance and bias. We evaluated how different negative-class definitions affect generalization and rule discovery in antibody-antigen binding using synthetic structure-based data. Models trained with negatives more similar to positives had reduced in-distribution performance but markedly better out-of-distribution generalization. Ground-truth analyses revealed that inferred binding rules shift with negative set choice, and experimental validation confirmed these findings, emphasizing dataset design for robust, biologically meaningful models.

We are developing a new strategy for designing repertoires of billions of structurally diverse and stable human antibodies. I will first describe two methods we developed for atomistic antibody design that enable this strategy and show that each method can optimize antibodies across a variety of criteria without prior mutational data. This shows that optimizing native-state energy is an excellent first approach for antibody optimisation. I will then describe a proof-of-concept universal repertoire of 500 million variants we designed and show we can reliably select highly developable and reasonably high-affinity antibodies against diverse targets.

Proof-of-concept has been demonstrated showing how AI methods can be used to design and optimize large molecules. We must shift our focus to scaling to maximize the productivity and innovation gains. This talk will cover a selection of PoCs and then how we are scaling digital biologics at Takeda across our portfolio.

At Manifold Bio, we’ve built a direct-to-vivo platform that connects AI-driven protein design to functional data from living systems. Using this approach, we generate and evaluate thousands of designed binders to novel targets simultaneously in vivo. This massively multiplexed framework has yielded functional brain shuttles capable of crossing the blood–brain barrier, and we are now extending it to other tissues to enable selective delivery of diverse therapeutics. By integrating AI design and in vivo multiplex readouts, we aim to generalize tissue targeting across the entire body.

We recently announced JAM-2, which can design antibodies with drug quality properties with high success rates. We'll discuss these results, and also share examples of what successful deployment on real drug discovery programs partnered with large pharma looks like. We'll discuss roadblocks and share practical lessons/advice for how to build teams and infrastructure to ensure AI driven biologics discovery delivers real drugs not just headlines.