We’re running a high-throughput wet lab for protein designers. To date, we’ve synthesized and tested over 10,000 different proteins and performed binding assays against more than 50 different targets. In this talk, we’ll share insights on what makes some targets more difficult than others, how target properties influence protein design strategies, and what types of experimental data are most useful for improving machine learning models in this domain.

Therapeutic antibody design is a complex multi-property optimization problem that traditionally relies on expensive searches through sequence space. Here, we introduce “lab-in-the-loop antibody design,” a new approach to antibody design that orchestrates generative machine learning models, multi-task property predictors, active learning ranking and selection, and in vitro experimentation in a semi-autonomous, iterative optimization loop. In this talk, we will discuss considerations for scaling model-driven design across complex modalities and targets.

Amgen is applying an integrated approach to therapeutic antibody engineering, combining modern computational protein design, predictive and generative AI, and high-throughput experimentation. Purpose-built ML models enable rapid generation and selection of therapeutic candidates with favorable developability indicators, and a tightly linked in silico–experimental workflow enables early and iterative feedback, improving how therapeutic candidates are designed, evaluated, and advanced across the discovery pipeline. We continuously adapt this workflow to more diverse therapeutic modalities and evolving automation technologies.

Protein structural generative models have been applied to a wide range of design tasks. As part of our efforts toward designing functional proteins involving structural dynamics, we developed an all-atom model, Protpardelle, that leverages a language model–based architecture to generate 3D structures. Design decisions and training datasets guide the behaviors of ML models, and we developed an evaluation metric, called SHAPES, to reveal the biases in state-of-the-art generative models. Many challenges remain, but structure generative design points to new ways to create novel functional proteins.

This talk will present an overview of Sanofi's state-of-the-art computational pipeline for de novo design of VHH therapeutics. The integration of the computational workflow with customized wet-lab processes for efficient molecule discovery and optimization will be discussed. A use case demonstrating the application of the pipeline for the computational design of VHH building blocks against therapeutic targets will be shared.

This young scientist meet-up is an opportunity to get to know and network with members of the BioLogic Summit community. This session aims to inspire the next generation of young scientists with discussion on career preparation, work-life balance, and mentorship.

This discussion group will convene for a discussion of applying language model architecture to protein structural features.

Much recent work on multimodal ML/AI for protein design has focused primarily on building larger models. We take an alternative approach, with experiments designed for better multi-modal models—and vice versa. We show that joint modeling allows intelligent integration of alternating rounds of ML-guided display selections and active-learning driven multi-objective optimization of antibodies produced in high throughput via cell-free synthesis, yielding highly developable binders unattainable via either modality alone.

ML/AI-designed epitope scaffolds show promise for ligand discovery in challenging targets, with recent efforts emphasizing the maintenance of binding sites during design. We discuss two studies where we demonstrate the successful creation of scaffolds showing significant ligand binding, with or without an empirically determined protein structure. The scaffolds show robust soluble expression in both bacterial and mammalian systems (> 30 mg/L), have proper disulfide bond formation confirmed by MS, are monodisperse by aSEC, have high alpha-helical content as predicted, and show double-digit nM binding to native ligand by BLI. Further structural characterization of the protein-ligand complexes is underway, and lessons learned from the design process are discussed.

Accurately modeling biomolecular interactions is a central challenge in modern biology. While recent advances have substantially improved our ability to predict biomolecular complex structures, these models still fall short in predicting binding affinity and generating accurate de novo designs. Here, we present the Boltz model series, open-source models for structure, binding affinity, and design that provide a robust and extensible foundation for both academic and industrial research.

An AI-guided pipeline will be discussed that integrates AlphaFold2, ProteinMPNN, and live-cell screening to convert conventional antibody sequences into functional intrabodies for use inside living cells. Our approach optimizes antibody frameworks while preserving epitope-binding regions, rescuing 18 previously nonfunctional sequences—including a panel for imaging histone modifications. This method offers a scalable, cost-effective route to intrabody development and opens new doors for live-cell imaging and functional studies.

• Structural biology to map epitope footprints—how important is it still to define epitopes/paratopes experimentally for precise targeting—is AI already there?
• Structure prediction and computational design—machine learning or physics-based?
• Antibody challenges—structure prediction, dynamics, design, diversity, glycan shields, breadth
• When will we have an AI designed therapeutic/vaccine?

De novo protein design traditionally overlooks conformational dynamics and desolvation—factors critical for protein function. We introduce a computational workflow that integrates these properties at the earliest design stages. By analyzing molecular dynamics and calculating desolvation energies, our approach more effectively identifies viable candidates than static methods. This strategy boosts the predictive power of design tools, significantly improving the success rate for developing stable and functional proteins.

Influenza vaccines tend to induce strain-specific antibodies against the hemagglutinin (HA) protein that protect against a narrow range of strains, thus requiring updated annual vaccines for ongoing protection. Rare, broadly reactive antibodies that recognize a diverse range of influenza HAs have been isolated from humans. We explore these broadly reactive antibodies as templates for designing universal vaccines for influenza.

Generating a library of antibody designs that is epitope-specific and highly developable is extremely challenging. There have been few, if any, studies rigorously combining and comparing design approaches to identify an optimal toolbox. To systematically address this problem, we compared exhaustive combinations of classical and deep learning-based methods according to their relative success rates in yeast display binding measurements. This allowed us to identify optimal workflows for similar design efforts.

In developing therapeutic antibodies, the heavy chain is often prioritized while appropriate pairing of the light sequence is important for functionality. We introduce LICHEN, a heavy-chain-conditioned light sequence generation tool that enables collaborative design by leveraging computational capabilities alongside experimental expertise. LICHEN generates valid and diverse light sequences which are a fit for the heavy sequence, as demonstrated with high expression yields and retained affinity in vitro.

Computational models are now widely used to predict critical antibody properties such as binding affinity, immunogenicity, and developability. While most AI-driven methods have been tailored for conventional monoclonal antibodies, the therapeutic landscape is increasingly dominated by complex multispecific formats. This talk will address this gap by focusing on property modeling for bi- and trispecific antibodies.

The growing interest in bi- and multispecific therapeutic proteins stems from their unique modes of action. While significant progress has been made in predicting developability for standard antibodies, these complex formats present ongoing research challenges, highlighting the need for improved tools. This presentation will discuss various in silico approaches (and their associated remaining challenges) aimed at predicting the critical features required for designing developable multispecific drug candidates.

Design of multispecific biologics typically involves connecting different subunits that bind different targets. These non-natural asymmetric formats present major developability challenges, including poor expression, suboptimal stability, high immunogenicity, charge asymmetry, and manufacturing difficulties. Here, we show AI usage for designing multibodies: natural, symmetric IgG antibodies that are multispecific and highly developable. We present numerous examples of multibodies with experimental data suggesting that multibodies can solve many challenges of therapeutic multispecifics.

AstraZeneca’s InSiDe (In Silico Developability) platform provides cross-pipeline insights into antibody developability risk via machine learning models for non-specific binding, self-association, chemical liabilities, etc. As interest in multi-specific therapeutics grows, so does the need for in silico developability predictions for these complex formats, posing additional challenges relative to conventional antibodies. Here, we identify gaps where computational approaches used for conventional antibodies are insufficient and discuss approaches to overcome these hurdles.