Building machine learning models and neural networks trained on structural information of drug targets to predict antimicrobial resistance

点击翻译按钮获取中文摘要

The proposed project focuses on training machine learning models using protein structural, chemical and evolutionary features of relevant antibiotic targets to predict antimicrobial resistance (AMR) conferred by Mycobacterium tuberculosis (Mtb). Whilst many researchers are using genetic features to predict resistance, we have previously demonstrated that traditional machine-learning models trained on structural and biophysical features of RNA polymerase can robustly and accurately predict the effect that a missense mutation confers on rifampicin susceptibility. However, these models are inherently unable to predict the effect multiple mutations can have, thereby constraining usable mutation data to a subset of the available mutation data, and thus limiting the clinical applicability of the models. The primary goal of the DPhil project is to address this. The student will have access to the dataset of around 70,000 clinical TB samples amassed by the international CRyPTIC project which was led by Oxford and is reporting its main findings through a series of publications. CRyPTIC collected 15,211 samples, each of which was whole genome sequenced and the susceptibility of 13 antibiotics measured using a 96-well broth microdilution plate. A limitations of this dataset was the lack of resistance to new compounds, such as bedaquiline. One of the CRyPTIC partners has recently provided c. 1000 high-value samples that are extensively resistant and the initial aims (Y1) of this DPhil are to analyse this additional dataset, including retraining previously developed machine learning models, as well as developing a rigorous statistical analysis pipeline to enable continuous robust and easily accessible performance assessment and benchmarking. This will facilitate the primary objective of the project; developing graph convolutional neural networks (gCNNs) featurised with structural and chemical data to predict AMR conferred by multiple mutations against first- and second-line anti-TB compounds. The hypothesis underlying this approach is that the topology of gCNNs can accurately capture all the information from a resistant allele, thereby allowing machine learning models to be efficiently trained and permitting protein targets with high levels of genetic variability to be considered for the first time. A logical extension, time permitting, would be incorporating dynamic data pulled down from molecular dynamics trajectories into the feature sets and assessing the impact on model performance. Aside from gCNNs being a more intuitive architecture to represent structural data than conventional convolutional neural networks (CNNs), gCNNs also preserve the concepts of the atom and the chemical bond until the final layers of the network, thereby preserving spatial information of the drug target. This allows for interrogation of atom embeddings to boost model attribution, a concept particularly relevant in clinically applicable molecular diagnostics. Although the use of structural data to predict AMR is still a relatively new approach, the real novelty of this methodology is that to date the field of AMR prediction has been largely unable to benefit from neural networks trained on sufficiently large datasets, and particularly neural networks trained on structural, physiochemical, and spatial information of the drug target. Furthermore, equivariant graph neural networks (which arguably show the most potential) are extremely new (2021), and with regard to structural modelling problems, have mostly been adopted by groups focussing on binding affinity prediction, not AMR prediction. This project would fall within the following EPSRC research themes: AI and data science Antimicrobial resistance Biological informatics Biophysics Clinical technologies Software engineering