TYC Mini-Symposium Bio Interest Group: Chris Oostenbrink, BOKU & Franca Fraternali, UCL

Drinks reception will be held in Physics E3/7 ground floor

Use of machine-learned potentials in QM/MM settings: the BuRNN methodology – Chris Oostenbrink
Institute for Molecular Modelling and Simulation, Department of Natural Sciences and Sustainable Resources, BOKU University, Vienna, Austria

In hybrid quantum mechanics / molecular mechanics (QM/MM) approaches, the molecular system is partitioned into regions that are treated at different levels of theory. At the interfaces between these regions, artifacts may occur. Examples are an overpolarization of the QM region due to near partial charges in the MM region, the lack of polarization in the MM region or unbalanced interactions between particles in the different regions, leading to an intrusion of MM particles into the QM region, or an accumulation or depletion of QM particles if particles are allowed to change character.

We have recently introduced a buffered embedding scheme, in which a buffer region between the inner (QM) and outer (MM) region is defined for which the interactions are computed both at the QM and MM level. This comes at the cost of introducing a second QM-calculation at every timestep of the simulation. The use of neural networks to describe molecular potential energies, allows for an elegant solution to this problem. We train a neural network directly on the difference between the two QM calculations, ensuring that the network reproduces the QM-interactions of the inner region, with itself and with the buffer region as well as the polarization of the buffer region due to the inner region. Any remaining artifacts largely cancel in the trained differences and are far removed from the inner region of interest. The use of the Buffer Region Neural Network (BuRNN) approach, furthermore, allows us to apply alchemical free-energy calculations at the QM-level of theory. In this presentation, I will demonstrate our most recent advances with BuRNN.

1. Lier, B., Poliak, P., Marquetand, P., Westermayr, J., Oostenbrink, C. (2022) BuRNN: Buffer Region Neural Network Approach for Polarizable-Embedding Neural Network/Molecular Mechanics Simulations. J Phys Chem Lett 13, 3812-3818. doi: 10.1021/acs.jpclett.2c00654
2. Crha,R., Poliak, P., Gillhofer, M., Oostenbrink C. (2025) Alchemical Free-Energy Calculations at Quantum-Chemical Precision. J. Phys. Chem. Lett 16, 863–869. doi: 10.1021/acs.jpclett.4c03213

Predicting cognate pairing of heavy and light immunoglobulin chains using single-cell antibody repertoire data – Franca Fraternali – Division of Biosciences, Department of Structure and Molecular Biology, University College London

The formation of stable antibodies through compatible heavy (H) and light (L) chain pairing is essential for both the natural maturation of antibody-producing cells in vivo and the engineered development of therapeutic antibodies ex vivo. Here, we introduce ImmunoMatch, a novel machine learning framework designed to decode the molecular principles underlying antibody chain pairing. Leveraging an antibody-specific language model, ImmunoMatch is trained on paired H and L sequences from single human B cells, enabling it to differentiate between cognate H-L pairs and randomly paired sequences.The application of ImmunoMatch is crucial in understanding how V(D)J usage preference drives differences in chain pairing propensities, and how this in turn affects in vivo antibody repertoire formation. Furthermore, ImmunoMatch opens up avenues to optimise chain pairing and facilitate in silico antibody design: while this has recently received much attention, research effort focuses almost exclusively on antigen affinity.

Guo D, Dunn-Walters DK, Fraternali F+ , Ng JCF+ . (2025). ImmunoMatch learns and predicts cognate pairing of heavy and light immunoglobulin chains. bioRxiv, doi: https://doi.org/10.1101/2025.02.11.637677

Ng JCF(*,+), Montamat Garcia G(+), Stewart AT(+), … Fraternali F(*). (2024). sciCSR infers B cell state transition and predicts class-switch recombination dynamics using single-cell transcriptomic data, Nature Methods, 21(5):823-834, doi: https://doi.org/10.1038/s41592-023-02060

Guo D, Ng JCF, Dunn-Walters DK, Fraternali F. VCAb: a web-tool for structure-guided exploration of antibodies. Bioinform Adv. 2024 Sep 24;4(1):vbae137

Bio: Professor Franca Fraternali is currently Chair of Integrative Computational Biology at UCL, and Head of the Institute of Structural and Molecular Biology.

Her group research focuses on the study of physical interactions of proteins and their interaction networks by combining information theoretic methods, statistical analyses and molecular simulations. Recently the group’s research introduced AI methods and multiscale approaches bridging atomistic and cellular protein function in Computational Systems Immunology.

Web-site: https://fraternalilab.github.io/