Advances in machine learning are transforming antibody engineering, yet key challenges remain: sparse training data, unreliable in silico performance metrics, and slow experimental feedback. In this talk, I will present recent work addressing these limitations. First, I will show how generalisable models can be built from limited and heterogeneous data to predict antibody biophysical traits, and how we applied this framework to nanobody thermostability with the development of NanoMelt. Second, I will examine common pitfalls in evaluating models for epitope-specific antibody design and propose alternative metrics that better capture practical performance. Finally, I will introduce SpyBLI, a rapid kinetic screening platform that delivers quantitative binding data in under 24 hours, without requiring purification. Together, these advances demonstrate how predictive algorithms and streamlined experimental pipelines can be combined to accelerate the design–make–test cycle in antibody discovery.

mRNA-based vaccines and therapeutics are increasingly popular and used for a variety of conditions. A key challenge in designing these mRNAs is sequence optimization. Even small proteins or peptides can be encoded by a vast number of mRNA sequences, each affecting properties such as expression, stability, and immunogenicity. To facilitate the selection of optimal sequences, we developed CodonBERT, a large language model (LLM) specifically for mRNAs.

AlphaFold2 has transformed structural biology with groundbreaking advances in protein structure prediction. However, despite these advances, many challenges remain. In this talk, I will present a progress update on AlphaFold2 and share insights gained from OpenFold, an optimized and trainable variant of AlphaFold2. I’ll also explore potential paths to address the limitations of AlphaFold-like systems.

Developing effective monoclonal antibody (mAb) therapies requires optimizing multiple properties, known as 'developability,' to ensure they can progress through the development pipeline. However, there is limited understanding of how the characteristics (redundancy, predictability, sensitivity) of developability parameters (DPs) in human-engineered antibodies compare to those in natural antibodies. We analyzed 86 DPs across two million antibody sequences, finding key differences in the predictability and sensitivity of sequence- and structure-based DPs. Our findings reveal that human-engineered antibodies occupy a narrower space within the natural antibody landscape, offering a foundation for more precise mAb design.

An Antibody Drug Conjugate (ADC) is a targeted cancer therapy that combines a cancer-specific antibody with a cytotoxic drug to deliver treatment directly to tumors, minimizing damage to healthy cells. However, ADCs pose several developability challenges due to their hybrid modality involving both small and large molecules. In this talk, I'll explore computational techniques that transform ADC evaluation and enhancement by predicting critical biophysical properties such as stability, solubility, and viscosity, thereby streamlining the design process. I'll highlight key methodologies and real-world applications, demonstrating the benefits of advanced in silico models in ADC development.

The adaptive immune response relies on B-cell receptors (BCRs) for pathogen neutralization, yet designing BCRs de novo remains challenging due to structural complexity. Here, we introduce Antibody Generative Pretrained Transformer (AbGPT), a fine-tuned model from a foundational protein language model. Using a tailored generation and filtering pipeline, AbGPT generated 15,000 high-quality BCR sequences, effectively capturing the intrinsic variability and conserved regions critical to antibody design.

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.

I will present recent work on interpretable deep learning models for predicting immune receptor binding specificity, including TITAN and DECODE, which jointly model T cell receptor (TCR)–epitope interactions and extract binding-associated motifs. I will also show how machine learning can help identify disease-associated TCRs in rare autoimmune conditions such as giant cell arteritis. Finally, I will discuss our ongoing efforts to integrate deep generative structural models to capture the flexibility of T and B cell receptors, enabling structure-informed therapeutic design.

Many researchers are trying to replicate the success of machine learning (ML) in computer vision and natural language processing in modeling biophysical systems. As a discipline, ML heavily relies on low-cost empirical benchmarks to guide algorithm development, but available benchmarks for biophysical applications have major shortcomings. Drawing inspiration from mutational landscape models, we propose Ehrlich functions, a new class of test functions for biophysical sequence optimization algorithms.

This presentation will explore the development of machine learning models for protein property prediction. Focusing on customized protein language models for the NANOBODY modality, it will demonstrate the implementation of a downstream thermostability predictor. Using this example, the talk will emphasize the critical importance of meaningful data splits in model training and validation.

Adaptyv accelerates data generation for training and validating AI models with a high-throughput lab that companies can access via our software interface and API. We empower protein-design teams to validate their AI models many times faster than before, without the need to run in-house wet labs. We're partnering with dozens of companies—from techbio startups to major pharma—and generated lab data for thousands of novel proteins.

We are leveraging protein language models to prioritize antibody candidates with superior biophysical properties, streamlining the path from discovery to functional reagents. By applying these models to high-throughput data from our integrated antibody platform—including antigen design, yeast display, sequencing, and characterization—we can rapidly select leads with enhanced developability. This talk will highlight examples and comparisons between different models and languages.

The IMMREP23 competition evaluated TCR-pMHC interaction prediction methods from 53 participating teams submitting 398 sets of predictions. Results showed reasonable performance for "seen" pMHC targets but near-random performance for "unseen" peptides, highlighting an unsolved generalization challenge. Here we discuss what has been learned from detailed analysis of predictions and provide insights for improving future benchmarks by carefully addressing biases in dataset construction.

Building on DTU’s in-house expertise in generative protein binder design, this new project at the DTU Arena for Life Science Automation (DALSA) aims to establish a fully automated closed-loop pipeline in collaboration with Novo Nordisk. By integrating AI-driven design with automated build, test, and learn stages, the pipeline will significantly accelerate the discovery and optimization of therapeutic candidates. The project showcases DALSA’s role as an open-access hub for innovation, demonstrating how cutting-edge automation can transform life science R&D and bridge academic and industrial efforts in drug development?

• Closing the loop in therapeutic discovery: How automation and AI can transform binder design from months to days, enabling rapid iteration and optimisation.
• Industry-academic co-development: Lessons from the DALSA–Novo Nordisk collaboration on aligning priorities, timelines, and infrastructure for mutual benefit.
• From proof-of-concept to scalable platform: The challenges of moving from isolated automation steps to a fully integrated closed-loop system.
• Rethinking validation bottlenecks: Why 'build' and 'test' stages—not just design—are now the limiting factors, and how automation addresses this.