Benchmarking AI-generated structural ensembles of membrane proteins against physics-based modelling

Technical Summary: Benchmarking AI-generated Structural Ensembles of Membrane Proteins

Problem Statement
Understanding the structural dynamics of proteins is essential for elucidating function and developing therapeutics, yet experimentally determining complete conformational ensembles remains challenging. This difficulty is exacerbated for membrane proteins due to their hydrophobicity and the complications introduced by non-native reconstitution environments. While Molecular Dynamics (MD) simulations offer a powerful tool for sampling conformational landscapes, capturing rare states often requires substantial computational resources and long runtimes. Furthermore, enhanced sampling methods can be technically demanding and prone to bias. Conversely, while AI-based structure prediction tools like AlphaFold have revolutionized static structure prediction, they typically generate a limited range of microstates, often failing to capture critical functional conformations (e.g., closed ion channels) without specific adaptations. Although the Biomolecular Emulator (BioEmu) has shown promise in generating conformational ensembles for soluble proteins, its ability to accurately model the complex landscapes of membrane proteins—particularly those absent from its training data—remains unverified.

Methodology
The authors evaluated the capability of BioEmu (version 1.2.0) to model the conformational dynamics of the bacterial rhomboid intramembrane protease, GlpG, a model six-pass membrane protein.

• Data Generation: 1,000 structural samples of full-length GlpG were generated using BioEmu. Unphysical structures were filtered, leaving 969 valid models.
• Clustering and Analysis: The models were clustered using FoldSeek with varying template modeling thresholds (0.7 for full-length, 0.98 for the transmembrane core only, GlpG ΔN) to identify conformational groups.
• Quantification: Key structural metrics were calculated, including Root Mean Square Fluctuation (RMSF) to identify variable regions, and lateral gate quantification (distance between Transmembrane helices 2 and 5) to distinguish between open and closed states.
• Benchmarking: The BioEmu ensemble was compared against:
  1. Experimental X-ray crystal structures (both open and closed states).
  2. A previously published dataset of 10 µs of MD simulations (seeded from 2IC8 and 2NRF structures) for GlpG.
  3. MD simulations of a human rhomboid (RHBDL1) where rare open states are difficult to sample.
• Resource Assessment: Computational costs (wall time and estimated carbon emissions) were calculated for generating the BioEmu ensemble versus running equivalent MD simulations on a single GPU/CPU node.

Key Results

• Conformational Diversity: BioEmu successfully generated a diverse ensemble of GlpG conformations. While the soluble N-terminal domain exhibited high variability (with ~5.6% of models placing the domain in physically implausible positions within the lipid bilayer), the transmembrane core displayed biologically relevant dynamics.
• Lateral Gate Sampling: The ensemble captured a spectrum of lateral gate states. Approximately 12.8% of models occupied an "open" state (lateral gate distance >1.4 nm), while others represented "closed" states. Specific clusters corresponded to distinct functional states:
  • Cluster 1: Predominantly represented intermediate/partially open states.
  • Cluster 2: Consisted of fully open states with significant outward movement of TM5 and lifting of the L5 loop.
  • Cluster 8: Represented a fully closed state with a narrow lateral gate (1.16 nm), closely resembling inhibitor-bound crystal structures.
• Comparison with MD: Principal Component Analysis (PCA) revealed that the BioEmu ensemble overlapped significantly with the conformational space sampled by 10 µs of MD simulations, particularly regarding the outward movement of TM5. However, BioEmu did not fully reproduce the extreme tails of the MD distribution (i.e., the most open or most closed states observed in specific MD seeds). Notably, BioEmu outperformed short (500 ns) MD simulations in sampling rare open conformations for both GlpG and RHBDL1.
• Computational Efficiency: Generating 1,000 BioEmu samples took approximately 40 minutes on a single workstation. In contrast, achieving comparable sampling via MD (10 µs total) would have required over 500 hours on the same hardware. This represents an approximate 900-fold reduction in computational time and estimated carbon emissions.

Significance and Claims
The authors conclude that BioEmu serves as a rapid and computationally efficient approach for generating plausible conformational ensembles for relatively large, multi-pass membrane proteins. The study demonstrates that AI-based ensemble generation can access rare functional states (such as the open lateral gate of rhomboids) at a fraction of the cost of conventional MD simulations.

However, the paper maintains a modest stance regarding the limitations of the method:

1. Incomplete Landscape: BioEmu does not yet fully capture the entire conformational landscape or the specific equilibrium weighting of states observed in long-timescale MD simulations.
2. Physical Artifacts: The lack of an explicit lipid bilayer context in BioEmu's prediction process can lead to physically implausible placements of soluble domains relative to the membrane.
3. Complementary Role: The findings suggest that AI-based ensemble generation should complement, rather than replace, conventional molecular modeling. MD remains necessary for applications requiring precise characterization of conformational energetics, relative state populations, and the identification of specific rare states that may lie outside the current AI sampling distribution.

Ultimately, this work provides proof-of-principle that AI tools trained largely on soluble proteins can be extended to model the dynamics of membrane proteins, offering an accessible entry point for exploring membrane protein dynamics where experimental structures are lacking.