Title:	Conformational sampling, scoring, and docking using AI
DNr:	Berzelius-2026-297
Project Type:	LiU Berzelius
Principal Investigator:	Björn Wallner <bjorn.wallner@liu.se>
Affiliation:	Linköpings universitet
Duration:	2026-11-01 – 2027-05-01
Classification:	10610
Keywords:

Abstract

Proteins are key players in virtually all biological processes and accomplish many of their functions as part of larger protein complexes. Unfortunately, compared with structure determination of individual proteins, structural characterization of protein-protein interactions is substantially more challenging and remains a major problem in structural biology. Transient interactions are particularly difficult to study because they exist only for short periods of time and are therefore difficult both to detect and to characterize at molecular resolution. Yet such interactions play key roles in regulating complex signaling networks that determine normal cell fate and are implicated in diseases including cancer, autoimmunity, cardiovascular disease, and neurological disorders. In this proposal, we aim to use and further develop AlphaFold-based AI methods to study protein-protein interactions involving dynamic and disordered partners. Over the last year, we have adapted AlphaFold to more extensively sample conformational space, making it possible to capture alternative conformational states related to protein dynamics and function. We will continue developing these approaches and adapting them to newer generations of AlphaFold, including AF3. In particular, we will explore strategies for improving the sampling and identification of less abundant conformational states. We will also develop an open-source PyTorch implementation of AF3 within OpenFold, enabling us to train and fine-tune customized models for our specific applications. In parallel, we will initiate several new machine-learning projects that move beyond the direct application of AlphaFold. These include DistoMove, which aims to predict protein conformational movements; DockQPredict, which will predict the quality of protein-protein interaction models; and a new pipeline for predicting transient protein-protein interactions. Together, these projects will involve large-scale training, inference, and systematic evaluation of new neural-network models. Access to the GPU resources on Berzelius will therefore be essential both for developing new machine-learning models and for advancing our AlphaFold-based approaches to protein dynamics and protein-protein interactions.