Machine learning-based biophysical models of antibody properties and function
Project Description
Antibodies are proteins that play a key role in the immune response against pathogens by binding specifically to the pathogen. The body is able to modify sections of antibodies through mutations to improve the specificity and affinity of the binding to unseen pathogens. Consequently, the design of antibodies as possible therapeutic tools that can bind to specific targets (e.g., pathogens, cancerous cells) is an area of highly active research. However, which molecular and structural properties determine the specific binding of antibodies to protein targets remains unclear, thus hampering our understanding of the mutational effects in the immune response and impeding progress in the rational design of antibodies. In this project, we will develop machine learning methods to predict antibodies’ functional properties related to their binding that are informed by biophysical modelling of their sequence and structure as captured by graph-based representations of antibody-protein interactions. In particular, the aim of these models is to predict and characterize single-site mutations that can improve antibody binding to specific targets without compromising other biophysical properties, with potential applications in antibody design.
Main objectives of the project
To achieve our main goal, we will leverage various machine learning approaches which we have developed, and we will develop new ones to exploit the increasing amount of data on antibodies and their cognate target proteins. Schematically, the objectives and tasks of the project will be:
To train a model that can capture long-range complex dependencies between amino acids at different sites along the antibody/protein amino acid chain, and which, accounting for such dependencies, can provide a single-site measure of amino acid importance to target binding. For this task we will build upon a transformer-like architecture [1].
To build biophysically informed models of antibody-protein interactions that can give graph representations of such interactions and antibody/protein structures, summarizing and distilling relevant biochemical and structural information. We will employ different graph construction techniques, from geometric graphs that capture packing to biophysical models of energetic interactions to higher-order models (akin to simplicial complexes) that capture many-body interactions in the structure. We will then explore strategies to machine-learn refinements of such graph representations (e.g., GCNs or GNNs) by optimizing the task of predicting antibodytarget binding.
To set up a multi-task learning framework whereby different prediction tasks are performed jointly (e.g., predicting structural flexibility, binding specificity, binding affinity etc.). Such a framework will rely on graph neural networks and will be designed to obtain single-site predictions of importance to target binding that account for long-range correlations between sites as well as multiple structural and biochemical constraints. These predictions will be key to estimate in silico the effect of mutations and set up a computational framework to guide mutation-based antibody design in the laboratory. Ongoing collaborations with the Imperial Department of Chemistry, as well as with the LiverpoolImperial AIChemy UKRI Hub in AI will allow us to establish links to experimental antibody design for validation and further development.
Existing background work
The field of machine learning approaches to biophysical modelling and design of immune-related proteins like antibodies has witnessed growing activity recently [2]. The supervisors’ group has recently published a machine learning method to predict antibody binding affinity to specific targets that leverages jointly a modelling framework capturing antibodies’ structural fluctuations upon binding and convolutional neural networks [3]. The ongoing research is focussing on biochemically informed graph-based representations of antibody structures [4,5] and on combining them to neural network architectures to model how structural flexibility contributes to antibodies’ functional properties related to target binding.
Details of Software/Data Deliverables
The coding and data developments during the project will consist of well curated computational pipelines comprising:
Algorithms to produce graph-based representations of protein data combining structural and biochemical information;
Machine learning architectures (transformers, graph neural networks) taking protein data as input and performing different learning tasks (potentially in a multi-task learning setting).
The software deliverable will consist of the python packages made freely available e.g. via github (like we did for Ref. [3]) and usable through a webserver (like we did for Ref. [5]). Such software will allow a potential user to: pre-process custom antibody/protein data of interest to produce inputs to the machine learning methods and graph-based representations that can be used for further analysis; evaluate on them the predictions of the machine learning methods; re-train/fine-tune the machine learning architectures on the custom data; extract insights and analyze the predictions for e.g. antibody design purposes.
References:
[1] Leem, Mitchell, Farmery, Barton, Galson. Deciphering the language of antibodies using selfsupervised learning, 2022. Patterns, 3(7). [2] Bravi. Development and use of machine learning algorithms in vaccine target selection, 2024. npj Vaccines, 9(15). [3] Michalewicz, Barahona, Bravi. ANTIPASTI: interpretable prediction of antibody binding affinity exploiting Normal Modes and Deep Learning, 2024. Structure, 32: 1-13. [4] Song, Barahona, Yaliraki. Bagpype: A python package for the construction of atomistic, energy-weighted graphs from biomolecular structures, 2021. [5] Amor, B., Schaub, M., Yaliraki, S. et al. Prediction of allosteric sites and mediating interactions through bond-to-bond propensities, 2016. Nat Commun, 7:12477.
