We develop and apply machine learning models to simultaneously optimize multiple drug-like properties of biologics, including antibodies and enzymes. We present generative models that harness both the design of functional proteins and the prediction of drug-like properties to engineer therapeutically developable proteins. We produce and experimentally characterize these designs for fitness, function, and developability, exploring the synergy of these methods in a generalized multi-objective optimization pipeline for biologics.

Biophysical characterization of biologics is resource-intensive and requires extensive wet-lab experimentation. To scale such efforts to millions of candidate molecules, protein language models offer a promising approach for predicting experimental outcomes computationally. However, the performance of current ML models is constrained by the limited availability of large, high-quality training datasets. Here, we introduce a purpose-designed, information-rich dataset, tailored to train ML models for predicting nanobody developability with improved accuracy.

Due to their highly engineered formats and complex effector mechanisms, multispecific antibodies (including T-cell engagers [TCEs]) require context-specific training datasets to power ML models. We provide insights into our highly integrated and automated experimental and computational infrastructure that enables our cycle-based, multi-objective optimisation of multispecifics. We highlight our success deploying this infrastructure to develop a pipeline of highly potent and selective TCEs.

Dr. Jiangyan Feng earned her Ph.D. in Computational Biology from the University of Illinois at Urbana-Champaign, where she specialized in molecular dynamics simulations to investigate protein conformational dynamics and applied bioinformatics for machine learning model development. She currently works at Eli Lilly, where her research focuses on antibody engineering and the application of in silico methods to accelerate biologics discovery and optimization.

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.

• Public data availability and ease of screening designs and generating experimental data
• Deimmunization and humanization methods for antibodies vs. non-antibody biologics
• Computational and deep learning methods for functional engineering and de novo design
• Methods for developability predictions and multi-objective optimization​

• Data scarcity as the key constraint to progress in antibody model development
• High-throughput assays: Current limitations and future potential
• Cross-disciplinary insights from medicinal chemistry and related domains​

There is a severe lack of protein-protein interaction (PPI) structural data, especially for antibody-antigen systems. We present an approach to create thousands of putative PPI structures, wherein we computationally design many de novo PPI structures and subsequently validate them in vitro with AlphaSeq. We show that the validated putative structures form a dataset on which we can train downstream machine learning models to yield improved model performance.

The AI Structural Biology Network brings together pharmaceutical companies and cutting-edge federated learning technology to leverage the collective data of multiple parties to improve AI-driven drug discovery. In this session, we will show results on how we improved OpenFold3—an open-source reproduction of AlphaFold3—with the proprietary structural data of the participating pharma companies.

The OpenBind consortium represents a transformative opportunity to address the critical data bottleneck for evolving co-folding models. This talk will highlight the progress made for routine collection of structural-affinity pairs at a scale never attempted in the public domain. This challenge requires coordinating diverse scientific work packages: from AI-assisted compound design, automated synthesis, to high-throughput structural and affinity measurements, and finally data dissemination and blind challenges.

Sequence-only PLMs lack spatial context and miss critical folding, interface, and environment-dependent cues, while structure-prediction and docking methods are too slow and underperform on antibody and VHH complexes. NextGenPLM bridges this gap with a modular, scalable design that fuses pretrained PLMs with multimodal inputs—from raw sequences and functional assays to high-resolution structures—via spectral contact-map embeddings. Using results from internal campaigns, we will demonstrate its potential for rapid, data-driven biologics discovery.

• Active learning strategies for antibody optimization
• Closing the loop: Integrating experimental feedback into ML models 
• Multi-objective optimization in antibody DMTA 
• Speed vs. thoroughness: Designing efficient DMTA cycles​

• Strengths and limitations of the current systems
• Practical techniques to evaluate these systems
• Best practices for enabling agents to effectively utilize external tools and APIs
• Planning for complex problem decomposition and effective action sequencing
  

Despite advances in sequence and structure prediction, capturing functionally relevant protein dynamics at scale remains a major challenge. We present BioEmu,  deep learning system that emulates protein equilibrium ensembles by generating thousands of statistically independent structures per hour on a single GPU. Trained on over 200 milliseconds of molecular dynamics (MD) simulations, static structures, and experimental stability data, BioEmu captures complex motions like cryptic pocket formation and domain rearrangements. It predicts relative free energies within 1 kcal/mol accuracy and jointly models structural ensembles and thermodynamic properties, offering mechanistic insights and a scalable, cost-effective path to understanding and designing protein function

We present a dual-framework approach for therapeutic antibody design that combines EvolveX, a structure-based CDR design pipeline using FoldX, with MadraX, a PyTorch-integrated differentiable force field. This physics-informed strategy enables data-efficient, interpretable protein engineering. EvolveX achieved >1000-fold affinity gains with structural and functional validation, while MadraX bridges deep learning and biophysics by allowing gradient flow through energy functions, enhancing design in data-scarce contexts.

Many proteins' biological function rely on micro- to millisecond dynamics, but large-scale data to study these motions is minimal. By curating >100 NMR relaxation datasets, we noticed that an observable hides in plain sight in >10,000 proteins in the BMRB. We trained Dyna-1 on this observable and found that it also predicts µs–ms motion directly measured in NMR relaxation experiments. Dynamics linked to biological function (e.g., enzyme catalysis, ligand binding) are particularly well-predicted.

Antibodies exhibit structural flexibility often central to their function. Here, we introduce two approaches to model conformational ensembles of antibodies at scale. First, we developed a high-throughput MD workflow that reproduces ensemble metrics observed in all-atom simulations and experimental datasets. Using this pipeline, we simulated over 150,000 antibodies to investigate flexibility profiles. Second, we trained a deep learning model on MD data and demonstrate its ability to predict key conformations.

Despite advances in AI-driven protein design, fully de novo generation of antigen-binding antibodies without extensive experimental screening has remained challenging. Here, we introduce Chai-2, a multimodal generative framework that designs antibodies (VH/VL), nanobodies (VHH), and miniprotein binders zero-shot, guided solely by target structures and epitopes. Evaluated across 52 structurally novel protein targets lacking preexisting binders, Chai-2 achieves a ~16% overall binding hit rate—over 100-fold higher than prior computational approaches—identifying binders for ~50% of targets within a single experimental round, when testing no more than 20 designs each. For miniproteins, Chai-2 yield a 68% success rate, routinely producing picomolar-affinity binders.

Antibodies represent crucial therapeutics but traditional discovery methods are labor-intensive and limit high-throughput analysis. We developed an approach combining structural analysis and bioinformatics to infer heavy and light chain sequences from cryo-EM maps of serum-derived polyclonal antibodies bound to antigens. Using ModelAngelo for automated structure-building, we accelerated sequence determination and identified matches in B cell repertoires via ModelAngelo-derived hidden Markov models. Benchmarking against non-human primate HIV vaccine trials showed reduced analysis time from weeks to under one day with higher precision. Validation with murine influenza vaccination sera revealed multiple protective antibodies, enhancing discovery workflows for vaccine development and therapeutic applications.

Protein language models (PLMs) offer a fast and cost-effective alternative to traditional structure-based descriptors, which are computationally expensive and time-consuming. By generating encodings directly from sequence data, PLMs bypass the need for resolved protein structures. This talk explores whether PLMs can match the accuracy of traditional methods while providing a more efficient solution. We discuss comparative analyses that evaluate PLMs' precision and highlight their potential to accelerate drug discovery.

We present a lightweight, fully sequence-based convolutional neural network for predicting miniprotein expression. Trained on curated in-house data labeled for production fidelity, the model operates as a binary classifier and generalizes across topologies with >90% accuracy. Now integrated into all drug discovery pipelines, it enables efficient downsampling of designed proteins, reduces synthesis waste, and exemplifies the value of automated design-build-test loops in AI-driven protein design.