Supervised learning model predicts protein adsorption to carbon nanotubes

Abstract

The developed classifier predicts what proteins adsorb to nanoparticles and what protein features drive these interactions.

The protein corona from nanomedicine to environmental science

“…The hard corona develops gradually, is more stable and is composed of 'relatively immobile' proteins with substantial affinities to, and low tendencies to dissociate from, NP surfaces 57,60,64 . The understanding of both soft and hard coronas is crucial for inferring NP stability, function Personalized protein corona 27 ML predicts protein corona with NP physicochemical properties 128 Corona to deep pro ile plasma proteome 26 Protein corona term coined 12 NP properties change corona 57 Matrix conditions modify corona 66 Eco-corona term coined 37 Modulation of in vivo responses by corona 21 Biomarker discovery via protein corona 34 SA 33 ML predicts protein corona solely with amino sequences 40 Protein adsorption to particles 14 Leveraging protein corona for diagnostic and therapeutic approaches AI and protein corona Empirical investigations of protein corona and interactions with biological systems. Kinetically, the associations between proteins and NPs in biological fluids are governed by noncovalent interactions, such as electrostatic forces, hydrophobic forces, hydrogen bonding and π-π stacking 75 .…”

“…of NPs 40 and have predicted diseases in patients using personalized protein corona fingerprints 19,26 . Accordingly, a better understanding of the composition, pattern and decoration of biomolecules at the surface of NPs, supplemented by AI, can facilitate the development of safer and more effective nanomedicine technologies with desired biological fates.…”

Where Nanosensors Meet Machine Learning: Prospects and Challenges in Detecting Disease X

“…One way to combine all the constructed models is simply to average the predictions from each modelsuch as the creation of a random forest (RF) model using numerous independent decision tree (DT) models (Figure E-i) . For example, a RF classifier predicted 91 proteins that adsorb to single-walled carbon nanotubes with 78% accuracy and 70% precision, thereby identifying proteins with high binding affinities . Another way is to sequentially combine each constructed model, such that a new model is trained with the knowledge of the entire error that the whole ensemble has learnt so farsuch as gradient boosting techniques (Figure E-ii) .…”

“…98 For example, a RF classifier predicted 91 proteins that adsorb to single-walled carbon nanotubes with 78% accuracy and 70% precision, thereby identifying proteins with high binding affinities. 99 Another way is to sequentially combine each constructed model, such that a new model is trained with the knowledge of the entire error that the whole ensemble has learnt so far�such as gradient boosting techniques (Figure 3E-ii). 100 For example, an extreme gradient boosted tree (XGBoost) algorithm trained on fluorescent microscopy tracking of hundreds to thousands of poly(ethylene glycol)-coated polystyrene nanoparticles effectively predicted the neurodevelopmental age of rats with 86.64% accuracy by studying changes in the brain extracellular matrix.…”

From proteins to nanoparticles: domain-agnostic predictions of nanoscale interactions

“…As these methods use properties of the individual amino acids or rely on protein-specific characteristics (i.e., properties derived from sequence and residue conservation), they cannot be straightforwardly extended to molecules that lack these motifs, even when they share other physical and chemical features [1,[25][26][27]. Similarly, current ML methods for predicting nanoparticle-protein interactions use application-specific properties and are limited by small training datasets [28][29][30], which limits the cross-domain validity of the resulting ML models, and requires a new model for every application.…”