Research

Bridging physics-based modeling, artificial intelligence, and cryo-EM to map protein conformational landscapes

Molecular landscape visualization of membrane proteins and conformational dynamics — An ion channel sampling distinct functional states and drug-bound conformations across an energy landscape of hills and valleys

The central goal of my research is to develop data-driven, physics-informed methodologies for characterizing macromolecular motions and understanding how drugs modulate these dynamics at atomic resolution.

Physics-based molecular dynamics (MD) simulations provide a powerful framework for following the time evolution of macromolecules and linking molecular motions to biological function. However, conventional simulations remain computationally limited in accessing the timescales over which many functionally important conformational changes occur.

At the same time, artificial intelligence has transformed our ability to predict static protein structures. Tools such as AlphaFold have demonstrated the remarkable potential of data-driven approaches for structural biology. Yet, predicting and characterizing the conformational dynamics that underlie protein function remains a major challenge, in part because high-quality datasets describing molecular motion at atomic resolution remain scarce.

Experimental approaches such as cryogenic electron microscopy (cryo-EM), however, offer an exciting opportunity to address this limitation. Modern cryo-EM experiments can capture millions of heterogeneous molecular snapshots, providing an exceptionally rich source of information about protein conformational landscapes. Extracting meaningful structural and dynamical information from these data remains challenging because of experimental noise, conformational heterogeneity, and limitations in current computational methods.

Integrating Three Perspectives

Physics-based modeling

Provides high-resolution molecular priors and mechanistic insight into how proteins move and respond to ligands.

Artificial intelligence

Enables the analysis and integration of large-scale experimental datasets beyond the reach of traditional methods.

Cryo-EM

Provides direct experimental observations of molecular heterogeneity and motion in near-native conditions.

My research brings together the complementary strengths of physics-based modeling, artificial intelligence, and cryo-EM. By integrating these three perspectives, I aim to develop new computational frameworks for mapping protein conformational landscapes, connecting structure to function and dynamics, and understanding how small-molecule drugs reshape these landscapes.

Ultimately, my goal is to build quantitative approaches that bridge molecular simulation, experimental structural biology, and AI, enabling a more complete understanding of how macromolecular dynamics govern biological function and how they can be modulated therapeutically.

Selected Examples

Capturing unknown protein states with physics-based sampling

Enhanced molecular dynamics sampling can reveal functionally important conformations that are invisible to experiment, including cryptic drug-binding pockets and activation intermediates of pentameric ligand-gated ion channels.

Cryptic pocket opening and binding of a stimulant derivative in a vestibular site of the 5-HT3A receptor — Cryptic vestibular pocket opening in the 5-HT_3A receptor, discovered by enhanced-sampling simulations

Discovering cryptic pocket opening and binding of a stimulant derivative in a vestibular site of the 5-HT_3A receptor
N. Haloi, E. Karlsson, M. Delarue, R. J. Howard, E. Lindahl
Science Advances, 11, eadr0797
Adaptive sampling-based structural prediction reveals opening of a GABA_A receptor through the αβ interface
N. Haloi, S. E. Lidbrink, R. J. Howard, E. Lindahl
Science Advances, 11, eadq3788

Deep learning for cryo-EM maps of protein–ligand complexes

CryoLigATE is a deep-learning framework that enhances the resolvability of cryo-EM densities in protein–ligand complexes, improving the interpretation of bound small molecules.

CryoLigATE deep-learning framework for enhancing cryo-EM maps of protein-ligand complexes — Overview of the CryoLigATE deep-learning framework

CryoLigATE: enhancing the resolvability of cryo-EM maps in protein–ligand complexes using deep learning

N. Haloi, R. J. Howard, E. Lindahl

bioRxiv (2026)

Modeling alternative states from cryo-EM densities

Combining AlphaFold2-based models with density-guided simulations makes it possible to build structures in alternative functional states directly from cryo-EM data.

Modeling cryo-EM structures in alternative states with AlphaFold2-based models and density-guided simulations

T. Shugaeva, R. J. Howard, N. Haloi†, E. Lindahl†

Communications Chemistry, 8:317