Optimisation-based modelling for drug discovery in malaria

Abstract

The discovery of new antimalarial medicines with novel mechanisms of action is important, given the ability of parasites to develop resistance to current treatments. Through the Open Source Malaria project that aims to discover new medications for malaria, several series of compounds have been obtained and tested. Analysis of the effective fragments in these compounds is important in order to derive means of optimal drug design and improve the relevant pharmaceutical application. We have previously reported a novel optimisation-based method for quantitative structure-activity relationship modelling, modSAR, that provides explainable modelling of ligand activity through a mathematical programming formulation. Briefly, modSAR clusters small molecules according to chemical similarity, determines the optimal split of each cluster into appropriate regions, and derives piecewise linear regression equations to predict the inhibitory effect of small molecules. Here, we report application of modSAR in the analysis of OSM anti-malarial compounds and illustrate how rules generated by the model can provide interpretable results for the contribution of individual ECFP fingerprints in predicting ligand activity, and contribute to the search for effective drug treatments.

CLigopt: Controllable Ligand Design Through Target-Specific Optimisation

“…Due to concerns of introducing biases from overfitting, we utilise two pre-defined algorithms from rdkit [22] instead of neural networks, to compute the QED and SA score. Three machine learning approaches with diverse architectures were explored to predict pIC50: (i) transformer-CNN [23], a SMILES-based Quantitative Structure-Activity Relationship (QSAR) model that learns molecular embeddings from a transformer encoder, and predicts the corresponding activity using CharNN, (ii) modSAR [24, 25, 26], a two-stage QSAR model that involves detecting clusters of molecules, and applying mathematical optimisation-based piecewise linear regression to link molecular structures to their biological activities, and (iii) DeepAffinity [27], a semi-supervised deep learning model that utilises a unified RNN and CNN model to predict binding affinity. We compare Transformer-CNN and modSAR that are trained on a subset of hDHFR inhibitors, and DeepAffinity with the provided pre-trained parameters to select the most accurate model for predicting binding affinity.…”

“…Three machine learning approaches with diverse architectures were explored to predict pIC50: (i) transformer-CNN [23] 1: The proportion of molecules that pass certain filters when sampling randomly versus when sampling using the CLaSS accepted set. More results for each individual property can be seen in Tables S2, S3 using CharNN, (ii) modSAR [24,25,26], a two-stage QSAR model that involves detecting clusters of molecules, and applying mathematical optimisation-based piecewise linear regression to link molecular structures to their biological activities, and (iii) DeepAffinity [27], a semi-supervised deep learning model that utilises a unified RNN and CNN model to predict binding affinity. We compare Transformer-CNN and modSAR that are trained on a subset of hDHFR inhibitors, and DeepAffinity with the provided pre-trained parameters to select the most accurate model for predicting binding affinity.…”