Machine learning platform for determining experimental lipid phase behaviour from small angle X-ray scattering patterns by pre-training on synthetic data

Abstract

Lipid membranes are vital in a wide range of biological and biotechnical systems; they undepin functions from modulation of protein activity to drug uptake and delivery. Understanding the structure, interactions, self-assembly and phase behaviour of lipids is critical to developing a molecular undertanding of biological membrane mediated processes, establishing engineering approaches to biotechnical membrane application development. Small Angle X-ray Scattering (SAXS) is the de facto method used to analyse the structure of self-assembled lipid systems. The resultant diffraction patterns are however extremely difficult to assign automatically with researchers spending considerable time often analysing patterns ex situ from a beamline facility, reducing experimental capacity and optimisation. Furthermore, research projects will often focus on particular lipid compositions and thus would benefit significantly from a method which can be rapidly optimised for a range of samples of interest. We present a generalisable machine learning pipeline that is able to classify lipid phases based on their raw, experimental SAXS spectra, with >99% accuracy and an inference time of <60 ms, enabling high throughput on-site analysis. We achieved this through application of a synthetic data generation system, capable of building synthetic SAXS patterns from the underlying physics which dictate phase behaviour, and we also propose an extension of our system to synthetically generate co-existence phase spectra with known composition ratios. Pre-training our machine learning model on this synthetic data, and fine-tuning on experimental samples empowers the model in achieving state-of-the-art, rapid lipid phase classification, allowing researchers to be able to adapt their experiments on site if needed and hence massively accelerate high throughput lipid research.