Dengue virus poses a serious threat to global health and there is no specific therapeutic for it. The antibody response in dengue infection and immunisation can be deconvoluted with high-throughput sequencing and artificial intelligence methods. Machine learning applied to sequencing data identifies rare and underrepresented dengue-specific antibodies.

Foundation models have had a transformative impact across many fields, through their ability to learn from large unstructured datasets, and to be fine-tuned for specific tasks. Here I will introduce the Alchemab discovery platform, which blends computational and experimental approaches at every step. Three foundation models trained on antibody sequences (AntiBERTa, AntiBERTa2-CSSP and FabCon) enable many steps in our discovery process. Our AntiBERTa models achieve state-of-the art predictions in amino-acid-level and sequence-level tasks, while FabCon can generate human-like and developable antibody sequences.

The identification of the specific functions of each protein is essential for understanding the underlying mechanisms of life and developing novel treatments against deadly diseases. Large language models (LLMs) have emerged as a reliable tool for uncovering hidden knowledge in sequence-based data. In this seminar, I’ll present our work on protein foundation models, which employ LLMs and other deep-learning architectures to embed proteins in high-dimensional vector spaces and learn their functional properties proficiently. To accomplish this, we utilized multi-modal learning, wherein the amino acid sequence, molecular interaction, and text-based data are integrated to construct a holistic representation learning model.

The design and discovery of early-stage antibody therapeutics is time- and cost-intensive. I will present an end-to-end machine learning–driven single-chain variable fragments (scFv) design framework that uniquely combines large language models, Bayesian optimisation, and high-throughput experimentation. The method enables rapid and cost-effective design of thousands of scFvs across all complementary determining regions. The designed antibodies exhibit strong binding affinities, at high levels of diversity, to a given antigen.

Robust datasets are essential for efficient training of machine learning algorithms, particularly in the context of affinity and epitope prediction. We have developed an iterative selection strategy for yeast equilibrium sorting paired with NGS that promotes recovery of antibody sequences with broad ranges of paratopes and affinities. Coupling these outputs with high-throughput functional screening assays has the potential to yield broadly distributed, validated sequences, ideal for model training.

Large language models (LLMs) trained on antibody sequences have shown significant potential in the rapidly advancing field of machine learning-assisted antibody engineering and drug discovery. However, current antibody LLMs often overlook structural information, which could enable the model to more effectively learn the functional properties of antibodies. In response, we introduce IgBlend, a LLM which integrates both the 3D coordinates of the backbone and antibody sequences

We challenge the paradigm of selection from large universal libraries to obtain binding proteins rapidly and efficiently. For linear epitopes, we found it to be possible to exploit the periodicity of peptide bonds and create a completely modular system, based on a binding protein design that shares the same periodicity. To reach selective and sequence-specific binding, we found it to be advantageous to combine machine learning, structural biology, and experimental evolution.

• Novel deep learning approaches for predicting MHC antigen presentation and the modeling challenges
• Interpretability and explainability of the available deep learning models
• Best practices in data preparation for machine learning of peptidomics datasets
• Antibody design by transitioning from peptide presentation to protein screening