Data-driven discovery of molecular photoswitches with multioutput Gaussian processes

Abstract

We present a data-driven discovery pipeline for molecular photoswitches through multitask learning with Gaussian processes. Through subsequent screening, we identify several motifs with separated and red-shifted electronic absorption bands.

Calibration and generalizability of probabilistic models on low-data chemical datasets with DIONYSUS

“…Additionally, exact inference becomes computationally expensive in larger datasets, making GPs an ideal choice for low-data probabilistic predictions, as demonstrated in many previous works. 7,36,73,74 We use the GPFlow package to implement the GP. 75,76 For the MFP features, we use a kernel based on the Tanimoto distance measure commonly used for high-dimensional binary vectors, which has been implemented in Moss et al 55 The standard radial basis function (RBF) kernel was used for all other vectorvalued features.…”

“…This is set to 0.25 for the experiments. Molecules recommended for measurement are those that maximize eqn (7), i.e. X next ¼ arg max X˛X a UCB ðXÞ.…”

“…In addition to the datasets examined in this work, examples of applications in the low-data regime include design of optoelectronic materials (i.e. organic photovoltaics, 6 or photoswitching molecules 7 ), prediction of biochemical properties (i.e. olfactory response, 8,9 or mosquito repellency 10 ), and drug discovery.…”

Thermal Half-Lives of Azobenzene Derivatives: Virtual Screening Based on Intersystem Crossing Using a Machine Learning Potential

“…Further, the program is quite fast, replicating the results of spin-flip, time-dependent density functional theory (SF-TDDFT) in milliseconds through use of a transferable ML potential. This program adds to the growing collection of computational models for predicting photoswitch properties. ,− Our software and pretrained models are freely available at .…”

An Ecosystem for Digital Reticular Chemistry

“…For instance, researchers in image classification can easily compare the performance of competing approaches, as they can compare model performance on the same tasks on the same data set (e.g., ImageNet) . Over the last few years, similar benchmark data sets have been reported for generative models for molecules or quantum machine learning. − However, there is currently no widely used reference set for machine learning on metal–organic frameworks (even though the QMOF data set , makes important steps toward this goal). As a first step toward more comparable machine learning for MOFs, our package implements a consistent interface for collections of structures, along with some corresponding properties (e.g., gas adsorption or electronic properties) with which all data sets can be used via the same interface.…”