(I-341) A large-scale machine learning analysis of inorganic nanoparticles in preclinical cancer research

Friday, October 13, 2023

3:30 PM – 4:30 PM PDT

Location: Exhibit Hall - Row I - Poster # 341

Presenting Author(s)

Zilu Zhang (he/him/his)

PhD Candidate
Department of Biomedical Engineering, Duke University
Durham, North Carolina, United States

Co-Author(s)

BM

Bárbara B. Mendes

Postdoctoral Researcher
ToxOmics, NOVA Medical School, Faculdade de Ciências Médicas, NMS|FCM, Universidade NOVA de Lisboa, United States
JC

João Conniot

Postdoctoral Researcher
ToxOmics, NOVA Medical School, Faculdade de Ciências Médicas, NMS|FCM, Universidade NOVA de Lisboa, United States
DS

Diana P. Sousa

PhD Student
ToxOmics, NOVA Medical School, Faculdade de Ciências Médicas, NMS|FCM, Universidade NOVA de Lisboa, United States
JR

João M. J. M. Ravasco

Research Scholar
ToxOmics, NOVA Medical School, Faculdade de Ciências Médicas, NMS|FCM, Universidade NOVA de Lisboa, United States
AL

Andżelika Lorenc

Research Scholar
Instituto de Investigação do Medicamento (iMed), Faculdade de Farmácia, Universidade de Lisboa, Av. Prof. Gama Pinto; Department of Biopharmacy, Ludwik Rydygier Collegium Medicum in Bydgoszcz, Nicolaus Copernicus University in Toruń, United States
TR

Tiago Rodrigues

Assistant Professor
Instituto de Investigação do Medicamento (iMed), Faculdade de Farmácia, Universidade de Lisboa, Av. Prof. Gama Pinto, United States
JC

João Conde

Assistant Professor
ToxOmics, NOVA Medical School, Faculdade de Ciências Médicas, NMS|FCM, Universidade NOVA de Lisboa, United States

Primary Investigator(s)

DR

Daniel Reker

Assistant Professor
Department of Biomedical Engineering, Duke University
Durham, North Carolina, United States

Introduction:

Inorganic nanoparticles have become an important tool as cancer therapeutics, diagnostics, and theranostics. However, notwithstanding the achievements that have been made, it remains challenging to objectively, comprehensively, and systematically define key design requirements to ensure the success of inorganic nanomedicines in cancer treatment. To discover such design rules, we have curated the world’s largest database of inorganic nanoparticles in preclinical cancer research and use statistics and machine learning to capture trends in nanoparticle design, correlate nanoparticle properties with treatment outcomes, and provide predictive tools that can guide the creation of safe and efficacious nanoparticles for cancer drug delivery (Figure 1a).

Materials and Methods:

A database integrating data from 707 publications was established through a semi-automated search with manual data curation from the PubMed database, initiated using the search terms “in vivo” + “tumor” + “cancer” and “gold nanoparticles” or “silica nanoparticles” or “iron oxide nanoparticles”, covering the period from January 2007 to June 2020. Data was manually extracted from these studies except for biodistribution values that were digitized from manuscript figures using WebPlotDigitizer. For machine learning, we filtered features present in less than 60% of the database, and only studies were kept that had all the selected features annotated. Categorical features were transformed using either one-hot encoding or ordinal encoding depending on the inherent relationship between categories within each feature. Discriminative machine learning models were built and evaluated using repeated 10-fold cross validation. Model interpretation was investigated using SHAP (SHapley Additive exPlanations) analysis on the random forest classifier. Linear discriminant analysis was performed to reduce the dimensionality of the data to 2D for clustering analysis.

Results, Conclusions, and Discussions:

Metallic nanoparticles cored with gold or iron oxide have been consistently occupying over 65% of published studies each year, along with an increasing number of non-metallic and alloyed nanoparticle systems. Noticeable uniformity in physical characteristics such as medium sizes (10-100nm, 45.0%), spherical geometries (81.5%), and negatively charged surfaces (50.8%) indicated a discernable bias in the current nanoparticle design, possibly driven by prior success stories and material or biological requirements. Active targeting (35.4%), however, constantly lagged behind passive targeting (64.6%) as the minor choice of targeting strategy (Figure 1b, c). Tumor volume reduction exhibited a skewed distribution, with over 50% of studies demonstrating tumor shrinkage of more than 90%. Surprisingly, this outcome was accompanied by only a low delivery efficiency of less than 7.5% ID/g in over half of the studied nanoparticles, indicating an unmet need and a substantial scope for more effective control of desired biodistribution (Figure 1d, e). Beyond these observations based on statistical analysis, we designed machine learning algorithms capable of forecasting nanoparticle effectiveness directly from their properties. The predictive model suggested, for instance, that photothermal therapy and theranostics involving photoacoustic imaging modalities were strongly correlated with significant tumor shrinkage, and superior treatment outcomes were often reported from studies using breast cancer models. Cluster analysis revealed that rod-shaped geometry, photodynamic therapy, and combination therapies were particularly successful approaches to achieve high tumor reduction (Figure 1f-i).

Taken together, we here compiled the world’s largest inorganic nanoparticle database from over 700 preclinical studies published in the past decades. We performed large-scale statistical analysis that revealed both established and emerging patterns in inorganic nanoparticle design, unveiled quantitative relationships between selected features and tumor reduction using robust machine learning algorithms, and clarified predictive mechanisms via explainable artificial intelligence techniques to guide nanotechnology research at the intersection of life sciences and data analytics. Overall, our analyses provide a blueprint for tracking, reflecting on, and contributing to the development of nanotechnology in a data-driven manner, and reinforce the importance of standardized operating and reporting procedures in the interest of high-quality data sharing and knowledge transfer.