Explore UCD

UCD Home >

Machine Learning

Modern deep-learning methods such as neural networks are of great use in modelling difficult problems due to their ability to learn complex relationships between inputs and outputs. We take advantage of these methods to allow predictions of nano-bio properties on a scale far wider than can be achieved using conventional techniques.

PMFPredictor

The potentials of mean force describing the interaction of a small chemical with a nanoparticle surface are key for prediction of its bioactivity, but require a significant amount of computational time for each pair, often taking close to a month to produce a sufficient number of potentials to predict protein affinity. To overcome this, we have developed a neural network model -- PMFPredictor -- that can predict these potentials in a fraction of the time. Our current version of PMFPredictor has parameters for over two hundred small molecules on approximately one hundred different nanoparticle surfaces, vastly improving the range of NPs for which the biocorona can be predicted and requiring only a matter of minutes to parameterise a new chemical and produce PMFs for the entire set of materials.

The paper describing the methodology: Machine-learning based prediction of small molecule–surface interaction potentials
I Rouse, V Lobaskin. Faraday Discussions 244, 306-335 (2023)

A figure showing line graphs illustrating the interaction potential between a small molecule (shown as a schematic) and a modified carbon nanotube (also pictured)

Example of an interaction potential generated using the PMFPredictor neural network for the pictured small molecule adsorbing to the surface functionalised carbon nanotube shown. The blue line indicates the results obtained from a metadynamics simulation, while the red line shows the results predicted by the neural network averaging over ten versions of the model (each shown as individual grey lines).

UANet

The UnitedAtom model enables a significant acceleration in the rate at which protein-NP adsorption energies can be calculated, but still requires on the order of minutes per protein per NP. Given the wide range of proteins and potential NPs, this clearly limits how many materials can be scanned for different species. 

The recently developed AlphaFold model for predicting the structure of a protein from its sequence  has lead to a significant increase in the number of protein structures available, such that models for the entire proteome of a wide range of species are now available. We take advantage of this to provide a large dataset for training a neural network model to predict binding energies based on the protein sequence and the target nanoparticle defined by a size, shape and material.

Contact the Soft Matter Modelling Group

UCD Physics Beech Hill, University College Dublin, Belfield, Dublin 4, Ireland.