PMIpred: A physics-informed web server for quantitative Protein-Membrane Interaction prediction

Abstract

MotivationMany membrane peripheral proteins have evolved to transiently interact with the surface of (curved) lipid bilayers. Currently, methods toquantitativelypredict sensing and binding free energies for protein sequences or structures are lacking, and such tools could greatly benefit the discovery of membrane-interacting motifs, as well as theirde novodesign.ResultsHere, we trained a transformer neural network model on molecular dynamics data for>50,000 peptides that is able to accurately predict the (relative) membrane-binding free energy for any given amino acid sequence. Using this information, our physics-informed model is able to classify a peptide’s membrane-associative activity as either non-binding, curvature sensing, or membrane binding. Moreover, this method can be applied to detect membrane-interaction regions in a wide variety of proteins, with comparable predictive performance as state-of-the-art data-driven tools like DREAMM, PPM, and MODA, but with a wider applicability regarding protein diversity, and the added feature to distinguish curvature sensing from general membrane binding.AvailabilityWe made these tools available as a web server, coined Protein-Membrane Interaction predictor (PMIpred), which can be accessed athttps://pmipred.fkt.physik.tu-dortmund.de.

When Data Are Lacking: Physics-Based Inverse Design of Biopolymers Interacting with Complex, Fluid Phases

“…The observed accuracy of neural networks trained by EVO-MD data has motivated us to launch the protein−membrane interaction prediction (PMIpred) server (https://pmipred.fkt. physik.tu-dortmund.de/), 19 which utilizes a transformer model trained by physics-based generation (EVO-MD) of over 54000 curvature-sensing peptide sequences to predict the membraneinteraction behavior of peptide sequences (Figure 5A). Moreover, PMIpred enables users to (i) scan protein structures (PDB files) for the presence of membrane-binding domains, (ii) quantify curvature sensing, (iii) discriminate between membrane binding and curvature sensing (see Figure 5B for an example), and (iv) quantify the contribution of each individual amino acid to membrane binding, thereby easing the design of point deleterious mutations.…”

“…Subsequently, a follow-up analysis to determine raft affinity in native protein sequences could be streamlined by employing a neural network trained in a manner similar to PMIpred. 19 However, the main challenge will be in translating the affinity derived from simplified in silico models of lipid rafts to the complex and dynamic lipid rafts in vivo. Molecular Recognition in Structureless Proteins.…”

“…Overview of PMIpred: on-the-fly prediction of peptide–membrane interaction using physics-based inverse design. Adapted with permission from ref . Copyright 2023 The Authors, preprinted by Cold Spring Harbor Laboratory.…”

“…16,17 This leads to a search space which is highly localized around unique solution sequences, aligning with the key−lock hypothesis for molecular recognition. In contrast, interactions with fluid phases involve weak, transient, and dynamic interactions with individual constituents of the fluid phase such as, for example, a lipid membrane 15,18,19 or a protein liquid condensate. 9,20−25 These interactions feature search spaces characterized by smoothness and a high degree of degeneracy in relevant physicochemical properties.…”

“…Moreover, in structure-focused protein design, even minor alterations can have a significant impact within the tightly sculpted angstrom-resolution world of protein cavities. , This leads to a search space which is highly localized around unique solution sequences, aligning with the key–lock hypothesis for molecular recognition. In contrast, interactions with fluid phases involve weak, transient, and dynamic interactions with individual constituents of the fluid phase such as, for example, a lipid membrane ,, or a protein liquid condensate. ,− These interactions feature search spaces characterized by smoothness and a high degree of degeneracy in relevant physicochemical properties. In other words, due to the smoothness of the search space, the optimum is represented by a potentially large set of equally valid and degenerate solutions, rather than a single “holy grail” sequence that outperforms all others.…”