Composition-agnostic prediction of self-assembly in multicomponent amphiphile mixtures from molecular structure

TL;DR

This work tackles the challenge of predicting self-assembly in multi-component amphiphile mixtures by extending the critical packing parameter (CPP) to mixtures and building a composition-agnostic framework that predicts morphology directly from molecular structure. A large DPD-based dataset spanning 2–5 component systems is generated, and twelve combinations of feature representations and architectures are evaluated to identify representations that capture intramolecular and intercomponent interactions. The results show that architectures incorporating a fully connected graph convolutional network (GCN) layer, particularly the GCN–GCN design, achieve superior performance and strong extrapolative capabilities, predicting CPP for five-component and unseen molecular species even when trained only on lower-component data. This composition-agnostic approach enables efficient virtual screening and provides a foundation for rational design of complex amphiphilic materials, with potential extensions to concentrations and solvent conditions for broader applicability.

Abstract

Predicting self-assembly in multi-component amphiphilic systems is challenging due to the complexity of intercomponent interactions and the combinatorial growth of possible formulations. In this study, we develop a unified machine-learning framework that directly predicts self-assembly behavior from the molecular structures of constituent components, independent of the number or identity of those components. We extend the critical packing parameter (CPP) to multi-component systems and generate a large dataset of self-assembled morphologies using dissipative particle dynamics (DPD) simulations. By systematically evaluating twelve combinations of feature extraction methods and model architectures, we find that models incorporating a fully connected graph convolutional network (GCN) layer achieve superior performance, with the GCN-GCN architecture accurately capturing both intramolecular relationships and intercomponent interactions. Notably, this model exhibits strong extrapolative capability: it accurately predicts CPP values for five-component mixtures even when trained only on systems with fewer components, and it maintains high accuracy for mixtures composed of molecular species that are entirely absent from the training data. These results demonstrate that a composition-agnostic predictive framework can enable efficient virtual screening and provide a foundation for the rational design of complex amphiphilic materials.

Figures (8)

• Figure 1: Particle model used in the simulations. (a) Coarse-grained DPD beads used in this study: water bead as solvent (cyan), hydrophobic tail bead (red), weakly hydrophilic head bead (green), and hydrophilic head bead (blue), labeled as W, T, wH, and H respectively. (b)Representative modeled amphiphilic molecules composed of three types of DPD beads: hydrophobic tail (red), weakly hydrophilic head (green), and hydrophilic head (blue).
• Figure 2: Calculation of CPP for multi-component systems: The CPP for each component is calculated and then averaged with weights given by the number of molecules $N_i$ in the cluster. An example for a 3-component system is shown.
• Figure 3: Summary of the calculated CPP values for the simulated self-assembled structures. (a) Representative snapshots of self-assembled structures for systems with $n_\text{c} = 1$ and $n_\text{c} = 3$, together with the corresponding molecular models. These examples show that the computed CPP values are consistent with the expected morphologies. (b) Histograms of the CPP values for each component, binned at intervals of 0.1. Across systems with $n_\text{c} = 1-5$, the CPP distributions cover the full range from 0 to 1, demonstrating the diversity of the CPP data generated for subsequent machine-learning analysis.
• Figure 4: Two types of methods for encoding molecular structures used in this study: (a) Modified-SMILES, a method that linearly transforms coarse-grained molecular models, used as input for NN, CNN, and GRU algorithms; (b) Graph representation, a method that represents molecular structures as graph data using a feature matrix and an adjacency matrix, used as input for GCN algorithms.
• Figure 5: Schematic overview of the machine learning architectures employed in this study to predict the critical packing parameter (CPP) of multi-component systems from molecular representations. (a) single-stream architecture: all component features are concatenated into a single input array before feature extraction. (b) multi-stream architecture: each component is independently processed, and the extracted features are subsequently integrated for prediction. (c) interactive multi-stream architecture: inputs of each component are connected through a fully connected graph convolutional network (GCN) layer to explicitly learn intercomponent interactions. The feature extraction layers in these architectures can utilize various machine learning algorithms, including neural networks (NN), convolutional neural networks (CNN), gated recurrent units (GRU), and graph convolutional networks (GCN). CPP denotes the predicted critical packing parameter.