Excitable cells produce electrical impulses generated by ion transport across their membranes via membrane proteins called ion channels, themost important family being voltage-gated channels (VGCs). VGCs are essential physiological effectors: they control transport and cellular excitability. Mutations in the genes encoding their subunits can cause channel dysfunction. Such channelopathies (hundreds of mutations) have been implicated in a variety of pathologies such as epilepsy, pain syndromes, migraines, periodic paralyses, cardiac arrhythmias, hypertension, and hypotension.
Themechanism controlling the opening and closing of the channels, and thus ion conduction, consists of a conformational transition (called activation) of their voltage-sensing domains (VSD) made up of 4 transmembrane helices (S1 to S4). The conformational change of the VSD gives rise to activation currents (gating-currents) that can be measured during electrophysiological recordings. These currents provide information on the kinetics of channel activation and are directly impacted by mutations.
Despite significant progress in electrophysiology and more recently in cryo-electron microscopy, capturing the conformational transitions of the VSD during VGC opening and closing remains a challenge. Advances in AI, notably the AlphaFold series widely used to predict protein structures, struggle to capture intermediate states of the VSD and are, of course, unsuitable for predicting transition pathways, let alone their kinetics. Molecular dynamics (MD) simulations offer an alternative, but it remains very difficult to sample these transitions as they are characterized by significant energy barriers.
In recent years, there has been an explosion in the number of adaptive sampling methods for molecular dynamics simulations integrating or using machine learning. Following these approaches, our team recently developed an innovativemethodology integrating reinforcement learning and MD, to autonomously and unbiasedly explore the conformational landscapes of complex proteins. Our AI-driven adaptive sampling approach relies on a learning agent that directs MD simulations towards unexplored regions of the conformational space, particularly those describing slow and therefore biologically important motions. The algorithm (patent filed) therefore considers a machine learning framework that accounts for physical interactions and integrates fundamental biophysical principles.
Our algorithm has been optimized for deployment on GENCI’s HPC machines, efficiently using hundreds of GPUs. Preliminary results, obtained on model systems and confirmed during calculations on the Jean Zay (A100) and Adastra MI250X (CINES) machines, demonstrate convergence efficiency superior by an order of magnitude to other methods for these model systems.
As part of the Jean Zay 4 2025 Grand Challenge, we undertook to use our algorithms to map the complete kinetics of the Kv1.2 potassium channel and one of its pathogenic mutants. Kv1.2 is a crucial VGC for neuronal excitability. Mutations of this channel are directly implicated in developmental and epileptic encephalopathies (DEE), rare childhood pathologies often refractory to treatment. This work is carried out within the framework of a HospitalUniversity Health Research project (RHU) bringing together a consortium of doctors and researchers from hospitals, universities, and the private sector, coordinated by the Imagine Institute: https://epilepsies-innov4epik.com/project/model systems.
Our team’s objective is threefold: (i) to establish atomistic kinetic atlases for the wild-type channel and its variants, (ii) to quantify how a mutation deforms the energy landscape and modifies VSD conformational transitions, and (iii) to translate these alterations into mechanisms of modulation of the channel’s electrophysiological characteristics that can be directly linked to clinical phenotypes. Ultimately, the kinetic atlases obtained could guide the development of precision therapies, by classifying mutations according to their functional effects or guiding the rational design of drugs.
The resourcesmobilized during this Grand Challenge are considerable: they enabled us to generate several hundredmicroseconds of unbiased trajectories, totaling more than 20 TB of raw data. These are then compressed, analyzed, and integrated into high-resolution Markov State Models (MSMs), providing unprecedented access to the activation kinetics of this VGC.
The first results analyzed to date for the wild-type channel show the existence of four stable macro-states (S0 to S3) between the Resting and Activated states, separated by significant energy barriers. The sequential movement of the S4 helix, detected in the simulations, corresponds to the VSD transitions measured experimentally, and the activation kinetics can be compared with electrophysiology experiments reported in the literature. For the mutant, our analyses showed that the mutation indeed induces a gain of function.
Thus, thanks to high-performance computing on Jean Zay, and as part of this “Proof of Concept” study, we demonstrated that it is possible to “predict” that a genetic mutation induces changes in activation kinetics that can be implicated in a pathology. We also refined the protocol, which can now be generalized to other VGCs and other mutations, with the prospect of more rapidly identifying therapeutic strategies for neuronal, muscular, or cardiac channelopathies. These results thus pave the way for an effective contribution of molecular modeling to precision medicine.
Beyond the scientific contribution, this project illustrates the strategic value of the Jean Zay machine: without the interconnections and massively parallel computing power of the H100 GPUs, calculations on our “local” resources would have required several years.
“This work benefited from funding from the Statemanaged by the National Research Agency under the 4th PIA, integrated into France2030, under reference ANR-23-RHUS-0002”.
Key figure :
300,000 GPU hours on Jean Zay to map the conformational states of Kv1.2 and one of its mutants.
Definitions :
Voltage-gated ion channel: Membrane protein enabling selective ion passage, essential for electrical excitability of neuronal, cardiac, and other cells.
Markov State Model (MSM): MSMs are a powerful tool that combine short and disparate MD simulations at local equilibrium to model the long-term dynamics of complex conformational changes.
Reinforcement learning: AI method training agents to make decisions in a given environment by performing actions that maximize a reward.