Data-driven discovery of innate immunomodulators via machine learning-guided high throughput screening

Abstract

The innate immune response is vital for the success of prophylactic vaccines and immunotherapies. Control of signaling in innate immune pathways can improve prophylactic vaccines by inhibiting unfavorable systemic inflammation...

Data-Driven Discovery of Linear Molecular Probes with Optimal Selective Affinity for PFAS in Water

“…The training process first requires a unique representation of the molecules using SELFIES and optimization of a loss function to learn a continuous representation of these molecules within a low-dimensional latent space constituting the information bottleneck layer between the encoder and the decoder. This latent space provides a smooth and low-dimensional representation of the molecular design space that is well suited to the construction of surrogate structure–property models and enables a Bayesian-guided traversal and optimization of molecules within the design space. ,,, We provide a brief description of the VAE model construction and training in the Supporting Information; full details are available in Tang et al…”

“…The variational autoencoder (VAE) , model utilized in this study was previously trained on more than 1.2 M small molecules contained within the ZINC data set − plus a number of common chemical screening libraries . It was our anticipation that this pretrained VAE over a large class of small molecules would provide good representations of the 3850 linear probe molecules considered in this work as rich but interpretable featurization appropriate for our active learning search.…”

“…The variational autoencoder (VAE) 52,53 model utilized in this study was previously trained on more than 1.2 M small molecules contained within the ZINC data set 54−56 plus a number of common chemical screening libraries. 57 It was our anticipation that this pretrained VAE over a large class of small molecules would provide good representations of the 3850 linear probe molecules considered in this work as rich but interpretable featurization appropriate for our active learning search. The training process first requires a unique representation of the molecules using SELFIES 29 and optimization of a loss function to learn a continuous representation of these molecules within a low-dimensional latent space constituting the information bottleneck layer between the encoder and the decoder.…”

“…This latent space provides a smooth and low-dimensional representation of the molecular design space that is well suited to the construction of surrogate structure−property models and enables a Bayesianguided traversal and optimization of molecules within the design space. 26,28,53,57 We provide a brief description of the VAE model construction and training in the Supporting Information; full details are available in Tang et al 57 The 3850 candidate probe molecules in the design space were created and stored as SMILES strings using RDKit. 58 We use selfies 1.0.4 29 to encode each molecule into a SELFIES representation using the generated SMILES strings.…”

Exploring Optimized Organic Fluorophore Search Through Simple Experimental Data-Driven VAE

“…26,27 In response to these challenges, various generator architectures have garnered substantial interest. [28][29][30] Early work by Aspuru-Guzik et al on a SMILES-based variational autoencoder (VAE) opened avenues for optimized compound searches, albeit limited to small molecules. 31,32 Moreover, generator has been explored in ML-assisted material design as well but concentrate either on arbitrary design or theoretical properties.…”