Machine Learning for Protein Engineering Part 2

This presentation examines how AI-driven computational frameworks—combining sequence, structural, and biophysical data—can predict nonspecific binding and developability risks in therapeutic antibodies and related formats. By integrating protein language models, inverse folding approaches, and dynamic structural features (simulated or ML-derived), we demonstrate how these tools identify molecular liabilities, and in turn how these models can impact the DMTA cycle by enhancing hit selection, guide optimisation, and de-risk development.

Designing antibodies for pre-selected epitopes is hampered by scarce structural data and slow, costly experimental validation or noisy high-throughput screens. I will present a data-efficient workflow that fuses sequence and structural embeddings with data augmentation to design antibodies predicted to engage user-defined epitopes, addressing practical questions on minimal data requirements and useful in silico metrics. To close the Design-Make-Test loop, I will introduce SpyBLI—a kinetic screen that reports k_on, k_off, and K_D bypassing purification and concentration-determination steps, giving DNA-to-data in less than 24 hours. Benchmarking across affinities spanning six orders of magnitude confirms its reliability, enabling rapid, quantitative validation of epitope-specific antibodies.

Antibodies bridge adaptive and innate immunity through their constant (Fc) domains, yet most of Fc sequence space remains unexplored due to experimental constraints. To address this, we developed a machine learning-guided platform for Fc-engineering. By integrating the screening of synthetic Fc-libraries with next-generation sequencing and deep learning, we can accurately predict antibody functional activity from sequence, and computationally design antibodies with bespoke functional profiles, unlocking new possibilities for precision immunotherapy.

Machine learning optimizes biotherapeutics by evaluating antibody developability, focusing on stability, functionality, and safety. In silico assessments streamline discovery by deselecting problematic antibodies early. Predictive models can optimise libraries toward designs with fewer liabilities. Evaluating these models hinges on new, representative data, with an emphasis on generalisation to novel paratopes. Deliberate data partitioning and appropriate evaluation metrics are critical to achieving this and are the focus of this talk.

Machine Learning accelerates drug design using surrogate models to predict key molecular properties. An iterative experimental feedback introduces distribution shifts in design cycles. Domain generalisation classifies antibody–antigen stability across five design cycles. Foundational models and ensembling, improve performance on out-of-distribution data. A public codebase and dataset extend the DomainBed benchmark for realistic shift simulation.

Language models learn powerful representations of protein biology. We introduce a new foundation model suite that directly investigates scaling effects for protein generation. We then apply this for applications in antibody and gene editor design.

The AIntibody competition was launched to benchmark real-world performance, and potential value, of artificial intelligence (AI) models in antibody discovery through a blinded, prospective experimental design. In the inaugural challenge, 33 organizations submitted 527 antibody sequences responding to three tasks focused on RBD, the most studied protein in history: (1) in silico affinity maturation from early round NGS selection outputs, (2) affinity ranking from NGS selection outputs, and (3) de novo generative design of CDRs based on selection outputs. All sequences were expressed as full-length IgGs and experimentally tested for binding affinity using surface plasmon resonance (SPR) and developability. The affinities of the highest affinity antibodies were further validated by KinExA. This talk will provide an overview of the NGS data used in the competition and final results of the competition.

In traditional structural biology, the intrinsic technical difficulties associated with the experimental structural characterisation of biological macromolecules have frequently imposed reductionist approaches limiting the investigations of individual proteins. The rapid determination of protein structures starting from their sequences, assured by computational approaches based on machine learning, allows now the simultaneous elucidation of structure-function relationships in entire families. Illustrative examples will be provided for proteins (KCTDs/CHCHD4) involved in key physio-pathological processes.

Considerable unmet need exists in autoimmune, inflammatory, and allergic indications with underlying etiology related to immunoglobulins. IgG, IgE, IgM, and IgA can each play a role in disparate disease processes, and an ability to precisely target only the immunoglobulin isotype involved is crucial in striking the desired balance between efficacy and safety. Seismic has achieved this using its structure-augmented AI/ML IMPACT platform creating a Swiss Army knife of Ig degrading therapeutics.

We will challenge the paradigm of selection from large universal libraries to obtain binding proteins rapidly and efficiently. When it comes to linear epitopes, we can exploit the periodicity of peptide bonds and create a completely modular system, based on a binding protein design that shares the same periodicity. Here, we present our progress on a binding protein system that is modular and complementary to a given peptide sequence.

Advances in protein design have enhanced our ability to engineer proteins with defined properties, functions, and structures. Here, we integrate computational protein design with structural biology to develop targeted vaccines and therapeutics for two major global health threats: influenza and malaria. For malaria, we present computationally designed multi-epitope scaffolds that combine epitopes from the circumsporozoite protein (CSP) and the blood-stage antigen PfEMP1. Informed by cryo-EM structures, we designed 10 scaffolds that bind with low nanomolar affinity to broadly neutralizing antibodies (bnAbs) targeting both major/minor CSP repeats and PfEMP1. These scaffolds also engage inferred germline antibodies, underscoring their potential as germline-targeting vaccine candidates.For influenza, we designed hemagglutinin (HA) lower stem scaffolds that bind with low nanomolar to picomolar affinity to stem-directed bnAbs such as 1D03 and 1C06. Structural analysis of 1D03-HA complexes (H1, H2) and related mAbs revealed that the HCDR1 loop is a key determinant of cross-group breadth. Guided by these insights, we engineered antibodies capable of binding H1, H2, and H5 subtypes with high affinity.Additionally, we optimized the central stem bnAb CR9114 for enhanced breadth across group 1 and group 2 Has, including H3, using a combination of language models and physics-based interface design.Together, our work highlights the power of integrating structural data, machine learning, and physics-based modeling to inform rational antibody and immunogen design for infectious disease targets.

Protein-protein interactions (PPI) are essential for most biological processes governing life. Using a geometric deep-learning framework on protein surfaces, we generated fingerprints capturing key interaction features. As a proof of concept, we designed de novo protein binders targeting proteins and protein-ligand complexes. These novel interactions could act as protein therapeutics, enhance biosensing, and enable the construction of new synthetic pathways in engineered cells.