Machine Learning-Enhanced Nanoparticle Design for Precision Cancer Drug Delivery

Abstract

In recent years, nanomedicine has emerged as a promising approach to deliver therapeutic agents directly to tumors. However, despite its potential, cancer nanomedicine encounters significant challenges. The synthesis of nanomedicines involves numerous parameters, and the complexity of nano-bio interactions in vivo presents further difficulties. Therefore, innovative approaches are needed to optimize nanoparticle (NP) design and functionality, enhancing their delivery efficiency and therapeutic outcomes. Recent advancements in Machine Learning (ML) and computational methods have shown great promise for precision cancer drug delivery. This review summarizes the potential use of ML across all stages of NP drug delivery systems, along with a discussion of ongoing challenges and future directions. The authors first examine the synthesis and formulation of NPs, highlighting how ML can accelerate the process by searching for optimal synthesis parameters. Next, they delve into nano-bio interactions in drug delivery, including NP-protein interactions, blood circulation, NP extravasation into the tumor microenvironment (TME), tumor penetration and distribution, as well as cellular internalization. Through this comprehensive overview, the authors aim to highlight the transformative potential of ML in overcoming current challenges, assisting nanoscientists in the rational design of NPs, and advancing precision cancer nanomedicine.

Keywords: biomaterial; drug delivery; machine learning; nanomedicines.

Figure 1

a) The workflow of applying machine learning (ML) to nanomedicine. Created with BioRender.com. b) Overview of ML‐enhanced nanomedicine drug delivery Processes. This figure illustrates the integration of ML in optimizing various stages of nanomedicine drug delivery. ML contributes significantly to the synthesis and formulation phase by predicting and optimizing parameters such as nanoparticle (NP) shape, size, solvent accessible surface area (SASA), zeta potential, polydispersity index (PDI), surface chemistry, encapsulation efficiency, and drug loading and release profiles. In the protein adsorption and blood circulation phase, ML predicts protein adsorption on NP surfaces as well as key pharmacokinetic parameters. During the extravasation to the tumor microenvironment (TME), ML quantifies tumor vascular permeability, helping us understand the mechanisms that facilitate NP transport across vascular barriers. Reproduced with permission.^{[ 7 ]} Copyright 2017, Springer Nature. In the tumor penetration phase, ML elucidates the relationship between tumor heterogeneity and the accumulation and distribution of NPs, aiming to enhance delivery efficiency. For cellular internalization, ML predicts cellular uptake, identifies key biomarkers and predicts mRNA transfection. Reproduced with permission.^{[ 7 ]} Copyright 2017, Springer Nature. Finally, ML assists in bioactivity assessment by analyzing large‐scale clinical datasets and predicting therapeutic outcomes, thereby providing a comprehensive evaluation of the efficacy and safety of nanomedicine formulations. Created with BioRender.com.

Figure 2

Integrating ML with high‐throughput experimentation to identify high drug‐loading nanoformulations. a) Molecular Dynamics simulation (MD) of drug–excipient systems quantified non‐covalent interaction potentials, which served as input for an ML model. This model used these interaction potentials and the molecular properties of drugs and excipients to predict drug–excipient pairs likely to form NPs. b) Left: Schematic of high‐throughput experimental workflow for creating NPs using nanoprecipitation and rapid Dynamic Light Scattering assessment. Right: High‐throughput testing of drug–excipient combinations, with a color gradient indicating NP size reduction compared to the unformulated drug. c) This figure visualizes the interaction network between drugs (red dots) and excipients (blue dots) as predicted by a computational model. The purple edges highlight specific drug–excipient formulations (I–‐VI) that were further characterized, such as Sorafenib and Glycyrrhizin (I). Node size corresponds to the number of predicted excipients or drugs that can form NPs with the respective compound. Gray lines indicate potential interactions between drugs and excipients. The bottom dashed box shows MD and Transmission Electron Microscopy (TEM) results on the left, and Kaplan–Meier analysis on the right. MD simulations map non‐covalent interactions between drugs (black van der Waals spheres) and excipients (green spheres). TEM images show NPs formed by drug–excipient co‐aggregation. Kaplan–Meier analysis shows mice treated with sorafenib–glycyrrhizin NPs have longer morbidity‐free survival compared to oral sorafenib and glycyrrhizin‐only control. Reproduced with permission.^{[ 44 ]} Copyright 2021, Springer Nature.

Figure 3

Bayesian Optimization (BO) combined with Deep Neural Networks (DNNs) optimizes the synthesis of NPs to achieve target spectral properties. Left: The two‐step optimization framework includes an initial loop (runs 1–5) where BO (blue points) samples the parameter space to train a DNN. In the subsequent loop (runs 6–8), the DNN (orange points) samples the parameter space to validate its regression function. Right: The suggested conditions from BO and DNN are tested on a droplet‐based microfluidic platform. The absorbance spectrum of each droplet is measured and compared to the target spectrum using a loss function before inputting to BO, while the fully resolved absorbance spectrum is provided to DNN. Reproduced under terms of the CC‐BY license.^{[ 59 ]} Copyright 2021, Mekki‐Berrada et al., published by Springer Nature.

Figure 4

ML‐assisted accurate prediction of molecular optical properties upon aggregation. The figure illustrates the multi‐modal approach using qualitative and quantitative molecular descriptors, combined with 5 ML algorithms: Logistic Regression, K‐Nearest Neighbor, Gradient Boosting Decision Tree, Random Forest (RF), and Neural Network. The model achieved the best performance with 93.83% accuracy. The approach was experimentally validated with three new molecules, predicting new structures superior to human perception. The rightmost figure shows molecule 2 correctly predicted with Aggregation‐Induced Emission (AIE) properties, validated by photoluminescence intensity at 578 nm. Reproduced under terms of the CC‐BY license.^{[ 80 ]} Copyright 2021, Liu et al., published by Wiley‐VCH.

Figure 5

ML predicts protein adsorption to carbon nanotubes (CNTs). a) The Random Forest Classifier (RFC) workflow used a dataset combining Liquid Chromatography‐Tandem Mass spectrometry protein corona composition data with UniProt protein properties data. The dataset was split into 90% training and 10% test data. b) Positive features (i) and negative features (ii) influencing protein corona formation on (GT)15‐SWCNTs. SWCNT, single‐walled carbon nanotube. c) Protein corona dynamics for (GT)15‐SWCNTs. (i) The figure shows the concentration of desorbed Cy5‐(GT)15 ssDNA versus time by experimental results from a corona exchange assay, where ssDNA desorption serves as a proxy for protein adsorption onto the SWCNTs. (ii) Comparison of endstate‐desorbed ssDNA with the RFC‐predicted in‐corona probability for (GT)15‐SWCNTs. Proteins are predicted by the RFC to be in the corona (probability > 0.5; blue‐green colors) or out of the corona (probability < 0.5; purple‐pink colors). Reproduced under terms of the CC‐BY license.^{[ 69 ]} Copyright 2022, Ouassil et al., published by American Association for the Advancement of Science.

Figure 6

ML quantifies tumor vascular permeability for rational NPs design. a) The ML‐based image segmentation models were developed by training on annotated images of vessel and protein nanoprobe distributions in tumor tissues. The extracted tumor vascular features were then used to quantify the permeability of the vessels for subsequent analysis. The rightmost diagram shows high permeability tumors using passive extravasation and low permeability tumors using active transendothelial transport. b) Left: Diagrammatic representations of ferritin nanocages (FTn) and its variants. Right: Diagram illustrating a strategy that enhances transcytosis in endothelial cells by stimulating Golgi‐dependent exocytosis. c) Left: Tumor growth curves in tumor‐bearing mice with various treatments. Right: Kaplan–Meier survival curve of treated tumor‐bearing mice. Reproduced with permission.^{[ 46 ]} Copyright 2023, Springer Nature.

Figure 7

ML predicts the tumor accumulation of nanomedicines. a) Schematic of the experimental protocol to identify tumor‐tissue biomarkers correlating with nanomedicine accumulation in tumors. Tumor accumulation of PHPMA, a prototypic polymeric nanocarrier, was assessed using computed tomography–fluorescence molecular tomography in three mouse models with varying tumor targeting. b) Correlation analyses were performed using 23 TME features related to vasculature (red), stroma (green), macrophages (blue), and cellular density (grey). Gradient Tree Boosting‐based ML ranked feature importance, identifying blood vessel and Tumor‐Associated Macrophages (TAMs) densities as key predictive features. The rightmost two images show the correlation of these densities with total liposomal DXR tumor accumulation. c) Histopathological biomarker product score workflow for predicting nanomedicine tumor targeting. (d) Means of blood vessel and TAM product scores plotted against means of liposome tumor targeting, demonstrating that biomarker product scoring effectively identifies breast cancers as tumors with low nanomedicine accumulation. Reproduced under terms of the CC‐BY license.^{[ 54 ]} Copyright 2024, May et al., published by Springer Nature.

Figure 8

a,b) ML predicts NP accumulation in individual micrometastases. a) Light‐sheet imaging captured individual channels for nuclei (DAPI), cancer cells (Ki67), blood vessels (GSL‐1), and NPs (darkfield). b) ML image segmentation enabled detailed analysis of 3D microscopy images. A predictive model of NP delivery to micrometastases was created using 3D imaging data and physiological characteristics. c) ML predicts NP tumor distribution. The leftmost image shows a 3D U‐87 MG tumor with 25 µm radius sampling regions. (i) Close‐up of the tumor with a 25 µm radius sample region. (ii) 3D image with NP boundaries overlaid on tumor features: blue (25 µm radius sphere), white (NPs), red (blood vessels), green (macrophages). Data from blood vessels and macrophages were used to train a Logistic Regression model to classify 25 µm radius regions as NP‐positive or NP‐negative. a, b) Reproduced with permission.^{[ 67 ]} Copyright 2019, National Academy of Sciences. c) Reproduced with permission.^{[ 70 ]} Copyright 2023, American Chemical Society.

Figure 9

ML predicts cellular internalization of carbon nanoparticles (CNPs) in specific cell Types. An artificial neural network (ANN) model was trained using the design parameters of eight CNPs to predict their cellular internalization. The model utilized the particle size, zeta potential, and surface chemistry of the eight CNPs, along with the inhibitor and the targeted cell, as input parameters. It assigned a quantitative value to each CNP based on its cellular internalization via the specific endocytosis pathway. Reproduced with permission.^{[ 58 ]} Copyright 2020, American Chemical Society.

Figure 10

Using DNA barcoding, high‐throughput sequencing, and ML to identify key factors in NP internalization. a) A schematic of the nanoPRISM assay illustrates the incubation of fluorescently labeled NPs with pooled cancer cells, followed by Fluorescence‐Activated Cell Sorting and DNA barcode sequencing for analysis. b) Omics integration identified predictive biomarkers of cellular uptake. Native expression of the lysosomal transporter SLC46A3 was found to be a key factor in predicting NP–cell interaction. Univariate analysis and ML showed that SLC46A3 expression is strongly inversely correlated with liposome association. c) Whole‐animal fluorescence images were captured 24 h after intratumoral injection of LIPO‐0.3% PEG* NPs in mice with low and high SLC46A3 expression. Reproduced with permission.^{[ 45 ]} Copyright 2022, American Association for the Advancement of Science.

Figure 11

Using ML and combinatorial chemistry to predict ionizable lipid mRNA transfection efficiency. a) Structures of chemical compounds in four components used as 4CR Ugi reactants for creating a combinatorial library of 384 ionizable lipids. b) A schematic diagram illustrating the process of synthesizing and screening a lipid library to generate data for training ML models. c) ML ranked the top structures in each component, followed by wet‐lab validation of lipids using these top structures. d) This figure shows representative In Vivo Imaging System images of mouse organs taken 6 h after intravenous administration of lung‐targeted mLuc lipid nanoparticles (LNPs) containing 119‐23 and C12‐200 (0.25 mg mRNA per kg mouse). Region of Interest 1 demonstrates signal intensities of 3.129 × 10⁸ (i) and 5.319 × 10⁷ (ii). Reproduced with permission.^{[ 47 ]} Copyright 2024, Springer Nature.

Figure 12

Using ML and Genetic Algorithm (GA) to discover NPs with selective cytotoxicity to cancer cells. a) Schematic depicting the screening of selectively cytotoxic NPs using an ML‐reinforced GA. b) ML‐reinforced GA was validated by reproducing experimental results of selective cytotoxicity of ZnO NPs on HepG2 and hepatocyte cell lines. Reproduced with permission.^{[ 156 ]} Copyright 2023, Wiley‐VCH.

Figure 13

ML predicts the immune activity of spherical nucleic acids (SNAs) functioning as cancer vaccines. a) The 11 structural parameters, organized by core, antigen, and oligonucleotide property categories, resulted in a potential design space of 3072 variants. DOPC, 1,2‐dioleoyl‐sn‐glycero‐3‐phosphocholine; DOPE, 1,2‐dioleoyl‐sn‐glycero‐3‐phosphoethanolamine; PO, phosphodiester; PS, phosphorothioate. b) Following the creation of a library of SNA‐based cancer vaccines, high‐throughput mass spectrometry was employed to characterize their cellular immune activity. These data were subsequently used to train an ML algorithm for predicting the immune activity of each SNA in the library. Reproduced with permission.^{[ 159 ]} Copyright 2019, Springer Nature.