Deep neural network model for highly accurate prediction of BODIPYs absorption

Abstract

A possibility to accurately predict the absorption maximum wavelength of BODIPYs was investigated. We found that previously reported models had a low accuracy (40-57 nm) to predict BODIPYs due to the limited dataset sizes and/or number of BODIPYs (few hundreds). New models developed in this study were based on data of 6000-plus fluorescent dyes (including 4000-plus BODIPYs) and the deep neural network architecture. The high prediction accuracy (five-fold cross-validation room mean squared error (RMSE) of 18.4 nm) was obtained using a consensus model, which was more accurate than individual models. This model provided the excellent accuracy (RMSE of 8 nm) for molecules previously synthesized in our laboratory as well as for prospective validation of three new BODIPYs. We found that solvent properties did not significantly influence the model accuracy since only few BODIPYs exhibited solvatochromism. The analysis of large prediction errors suggested that compounds able to have intermolecular interactions with solvent or salts were likely to be incorrectly predicted. The consensus model is freely available at https://ochem.eu/article/134921 and can help the other researchers to accelerate design of new dyes with desired properties.

More Is Not Always Better: Local Models Provide Accurate Predictions of Spectral Properties of Porphyrins

“…The models with the best statistical parameters were chosen to create the consensus models as average of these individual models. We observed the same effect as in the previous study [47], namely that solvent parameterization did not provide significantly better results. For example, the mean difference between RMSE of consensus models for prediction with the parameterization of solvent and without it was 0.6 nm for the JOUNG set which was within the standard mean error of the model (Table S2).…”

“…A 5-fold crossvalidation was used to estimate accuracy of developed models. The initial calculations were performed with and without parameterization of the solvent using procedure described elsewhere [47]. The models with the best statistical parameters were chosen to create the consensus models as average of these individual models.…”

Accurate prediction of 11B NMR chemical shift of BODIPYs via machine learning

“…These models based on descriptor-less methods made a good attempt at prediction of absorption maximum wavelengths of BODIPYs in a previous study. 50…”

“…39 At the first stage, the Deep Neural Network (DNN), 41 Random Forests Regression (RFR), 42 eXtreme Gradient Boosting (XGBoost), 43 Transformer Convolutional Neural Fingerprint (Transformer-CNF), 44 Transformer Convolutional Neural Networks (Transformer-CNN) 45 and SchNet 46 were chosen as ML algorithms. Previously, these methods showed a good quality of predicting the spectral properties (absorption maximum wavelengths and molar absorption coefficient) of dipyrromethene compounds including BODIPYs [47][48][49][50] and thermal properties of ionic liquids. 51,52 RFR is the most flexible and easy-to-use algorithm based on classification and regression trees (CART).…”

Beyond Woodward–Fieser Rules: Design Principles of Property-Oriented Chromophores Based on Explainable Deep Learning Optical Spectroscopy

“…However, designing specific property-oriented molecular structures is challenging because of the lack of a complete understanding of structure–property relationships; consequently, structural modifications using functional groups to achieve products with desired properties rely mainly on the scientists’ experience. Recently, deep learning methods have been attracting significant interest as innovative approaches for understanding the structure–property relationships of molecules and accurately predicting their properties. − Additionally, deep learning methods have been successful not only in estimating the molecular properties − but also in predicting the influence of the surrounding interactions on the molecular properties . For example, optical properties, such as spectral position and bandwidth, photoluminescence quantum yield (PLQY), and fluorescence lifetime, are largely affected by the surrounding solvent molecules.…”