Leveraging Machine Learning-Guided Molecular Simulations Coupled with Experimental Data to Decipher Membrane Binding Mechanisms of Aminosterols

Stefano Muscat, Silvia Errico, Andrea Danani, Fabrizio Chiti, Gianvito Grasso

TL;DR

This paper introduces a powerful computational framework combining machine learning (Deep-TICA) with enhanced molecular dynamics simulations (OPES) to accurately predict how molecules like aminosterols interact with and insert into cell membranes. This approach successfully deciphers the complex binding mechanisms of these neuroprotective compounds and provides experimental validation, paving the way for better drug design.

What Problem Does This Paper Address?

Understanding how molecules interact with and cross cell membranes is fundamental for many applications, from drug delivery to understanding disease mechanisms. However, accurately simulating these processes, especially rare events like membrane insertion, is computationally challenging. Traditional molecular dynamics (MD) simulations often struggle to capture the full free energy landscape due to the high energy barriers involved. This makes it difficult to predict binding affinities and detailed insertion pathways. The authors specifically focus on aminosterols, natural compounds that protect neurons from toxic protein aggregates implicated in neurodegenerative diseases like Alzheimer's and Parkinson's, highlighting the need to understand their interaction with neuronal membranes.

Methods

The core of this work lies in a novel computational strategy that integrates several advanced techniques:

  • Coarse-Grained (CG) Molecular Dynamics: To make simulations computationally feasible for complex systems like lipid bilayers, the authors employ the Martini 2.2p force field, which represents groups of atoms as single "beads." This significantly reduces the number of particles and computational cost.
  • On-the-Fly Probability Enhanced Sampling (OPES): This is an advanced sampling method that builds upon metadynamics. Instead of defining a fixed bias potential, OPES dynamically adjusts the bias to better sample the system's free energy landscape, allowing for more efficient exploration of rare events like membrane insertion.
  • Time-Lagged Independent Component Analysis (TICA): TICA is a linear dimensionality reduction technique used to identify the slowest, most important dynamic modes of a molecular system. These modes are often good candidates for collective variables (CVs) that describe complex processes.
  • Deep-TICA: This is the key innovation. Deep-TICA combines TICA with deep neural networks (NNs). It uses TICA to identify potential slow modes from simulation data and then employs a feed-forward neural network to learn a non-linear combination of these modes, creating a highly effective and accurate collective variable (CV) that can distinguish between different states of the system (e.g., molecule in water vs. molecule in membrane).

The process involves running multithermal OPES simulations to generate diverse trajectories, then training a Deep-TICA neural network on this data to create a specialized CV for aminosterol membrane insertion. This CV is then used in further OPES simulations to calculate the free energy landscape.

Datasets

The primary "dataset" used for training the Deep-TICA model consists of the molecular dynamics trajectories generated from the OPES expanded multithermal simulations of the aminosterol-membrane systems. These simulations provided the necessary data to identify the slow dynamics and learn the optimal collective variable.

Key Insights and Ideas

  • Deep-TICA as a powerful CV generator: The study demonstrates that Deep-TICA can effectively learn complex, non-linear collective variables that go beyond simple linear combinations, significantly improving the ability to describe intricate molecular processes like membrane insertion.
  • Aminosterols have distinct binding preferences: Both trodusquemine (TRO) and squalamine (SQ) show a preference for the membrane environment over the aqueous phase, with SQ exhibiting a slightly higher affinity. This is attributed to their structural differences, particularly the length and charge of the polyamine tail.
  • Membrane insertion is a multidimensional process: Simple Z-distance CVs are insufficient to fully capture the nuances of membrane insertion. The Deep-TICA CV, by incorporating multiple molecular descriptors, provides a more accurate representation of the free energy landscape, including the energy barrier.
  • Experimental validation is crucial: The computational predictions of binding affinity were remarkably well-aligned with experimental measurements using fluorescently labeled aminosterols and liposomes, lending strong credibility to the computational framework.

Key Results

  • The Deep-TICA framework successfully generated a collective variable that accurately distinguished between aminosterols in the aqueous environment and those inserted into the lipid bilayer.
  • Free energy calculations revealed that SQ has a slightly higher affinity for the membrane than TRO, with calculated ΔG values of -21.27 ± 1.02 kJ/mol for SQ and -16.07 ± 1.35 kJ/mol for TRO (using Z-distance as CV).
  • When using the Deep-TICA CV, the calculated ΔG values were -22.75 ± 0.6 kJ/mol for SQ and -16.71 ± 1.1 kJ/mol for TRO, showing good agreement with the Z-distance results but providing a more nuanced representation of the free energy landscape.
  • Experimental binding assays confirmed SQ's higher affinity for liposomes (KD = 0.013 ± 0.002 mg/mL) compared to TRO (KD = 0.034 ± 0.004 mg/mL), translating to ΔG values of -27.36 ± 0.69 kJ/mol for SQ and -24.93 ± 0.47 kJ/mol for TRO. These experimental values are in good agreement with the computational predictions.

Limitations and Future Work

  • Concentration Discrepancy: The MD simulations were conducted at a concentration approximately 200 times higher than the experimental conditions. While the framework still provided accurate relative affinities, this difference in absolute concentration could be a factor in future studies.
  • Force Field Accuracy: While the Martini force field is efficient, its inherent approximations might contribute to the observed discrepancies in free energy values between computational and experimental results (though within typical error ranges).
  • Future Work: The authors suggest that this computational framework is highly adaptable and can be integrated with other advanced sampling techniques (e.g., bias exchange, parallel tempering) to potentially achieve faster convergence. They also envision its application to designing descriptors specific to membrane protein binding pockets, expanding its utility in drug discovery.

Summary updated

Abstract

Understanding the molecular mechanisms of the interactions between specific compounds and cellular membranes is essential for numerous biotechnological applications, including targeted drug delivery, elucidation of the drug mechanism of action, pathogen identification, and novel antibiotic development. However, estimation of the free energy landscape associated with solute binding to realistic biological systems is still a challenging task. In this work, we leverage the Time-lagged Independent Component Analysis (TICA) in combination with neural networks (NN) through the Deep-TICA approach for determining the free energy associated with the membrane insertion processes of two natural aminosterol compounds, trodusquemine (TRO), and squalamine (SQ). These compounds are particularly noteworthy because they interact with the outer layer of neuron membranes, protecting them from the toxic action of misfolded proteins involved in neurodegenerative disorders, in both their monomeric and oligomeric forms. We demonstrate how this strategy could be used to generate an effective collective variable for describing solute absorption in the membrane and for estimating free energy landscape of translocation via on-the-fly probability enhanced sampling (OPES) method. In this context, the computational protocol allowed an exhaustive characterization of the aminosterol entry pathway into a neuron-like lipid bilayer. Furthermore, it provided accurate prediction of membrane binding affinities, in close agreement with the experimental binding data obtained by using fluorescently labeled aminosterols and large unilamellar vesicles (LUVs). The findings contribute significantly to our understanding of aminosterol entry pathways and aminosterol-lipid membrane interactions. Finally, the computational methods deployed in this study further demonstrate considerable potential for investigating membrane binding processes.
BibTeX
@article{muscat2024leveraging,
  author = {Stefano Muscat and Silvia Errico and Andrea Danani and Fabrizio Chiti and Gianvito Grasso},
  title = {{Leveraging Machine Learning-Guided Molecular Simulations Coupled with Experimental Data to Decipher Membrane Binding Mechanisms of Aminosterols}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2024},
  doi = {10.1021/acs.jctc.4c00127},
}
Figure