Nanoparticles for targeted drug delivery
/
The Heller lab at MSKCC has discovered that poorly soluble kinase inhibitors mixed with specific indocyanine dye excipients will spontaneously form nanoparticles with very high (90% by mass) drug loadings, and that these dyes specifically target certain tumors while maintaining high blood stability [Nature Materials 2018]. These indocyanine nanoparticles (INPs) avoid both off- and on-pathway toxicities while delivering high quantities of targeted kinase inhibitors directly to tumors, solving several therapeutic challenges in kinase inhibition.
The ability to encapsulate inhibitors within targeted, soluble nanoparticles with high drug loading holds enormous potential for accelerating the process of novel therapeutic development, enabling compounds that are potent inhibitors but may have poorer ADME-Tox properties to be rapidly translated into effective therapies via nanoparticle encapsulation. Several challenges must be overcome to enable this to become a standard approach in accelerating the development of new therapies:
What is the mechanism of nanoparticle formation? Is it driven by equilibrium processes (thermodynamic free energies) partitioning between nanoparticle formation, soluble drug and dye, and amorphous solids? Or is it inherently a kinetic phenomenon?
What is the structure of the nanoparticle? How homogeneous or heterogeneous is it?
How can we best exploit available data and known physics of nanoparticle formation to predict which inhibitors a given excipient (dye) will encapsulate, and how we can modify the excipient to stably encapsulate a given inhibitor at blood pH but rapidly dissolve in lower pH environments in late endosomes?
We are working with the Heller lab to address several aspects of these questions to enable INP encapsulation to become a standard tool in the drug formulations toolbox for accelerating the development of new therapies.
⚗️TUNING DYE ENCAPSULATION AND STABILITY
We are working with the Heller lab to tune both which drugs can be encapsulated within INPs, as well as to modulate the stability in blood and rapid drug release within cells through the synthesis of excipients based on indocyanine dye derivatives. Because the pH of the blood (pH ~7.4) and late endosomal compartment (pH 4.5-6.5) likely plays a role in favoring stability in blood and disassembly after endocytosis, we need to accurately model pH-dependent free energies of transfer between relevant phases. To do this, we aim to use alchemical free energy calculations to compute transfer free energies of drug and excipient between various phases (pure amorphous drug or excipient phases, nanoparticle, and buffer), and coupling these approaches with constant-pH and counterion sampling methodologies.
⚗️KINETIC MECHANISM OF NANOPARTICLE FORMATION
illustration of critical nucleus size in nucleation theory. Image from mmarso
Nanoparticle formation may be an inherently kinetic, rather than thermodynamic, phenomenon. In that case, we need to build a model of the kinetics of formation of these nanoparticles in order to develop predictive quantitative models.
Using the OpenPathSampling package (based on the GPU-accelerated, Python-enabled OpenMM molecular simulation framework) we developed in collaboration with the Bolhuis and Noé laboratories, we aim to study the kinetic mechanism of nanoparticle formation to understand how critical nucleation events occur using transition path and interface sampling techniques. By identifying the critical nucleus—the minimum assembly of drug and excipient for which the nanoparticle formation process becomes a downhill process as more drug or excipient is added, which may be much smaller than the nanoparticles—we can develop predictive models that allow us to understand how modifying either component will affect the potential to form nanoparticles and their formation rates.
⚗️MACHINE LEARNING FOR RAPID PREDICTION OF NANOPARTICLE FORMULATIONS
While our initial quantitative structure-nanoparticle assembly potential (QSNAP) models demonstrated machine learning models can effectively predict the ability of an excipient to encapsulate small molecules, extending these models to make accurate predictions for many drug-excipient pairs is an active area of research.
Because INP nanoparticles can be so easily synthesized, the protocol can easily be scaled to screen many combinations of drugs and excipients using automation to generate large datasets useful for training machine learning models. Recently, collaborator Daniel Reker and colleagues did this to test 1440 formulations, from which a random forests machine learning model were built to allow the screening of millions of potential combinations. We are working to scale up this high-throughput screen to our in-house fully automated laboratory to inexpensively generate very large datasets using active learning techniques.
To exploit active learning, we must develop machine learning methods capable of predicting their own uncertainty to direct the synthesis of new compounds and excipients to test from very large virtual synthetic libraries, like Enamine REAL space (which has 14B compounds that can be made for ~$100/compound with an >80% success rate). Inspired by work by Connor Coley in exploring uncertainty quantification (UQ) techniques in the context of message-passing networks in quantitative structure-activity relationships [arXiv:2005.10035], we are experimenting with this class of models to predict nanoparticle formation propensity for drug-excipient pairs using packages developed in-house (gimlet and pinot).
examples of indocyanine dyes that efficiently form nanoparticles
SOFTWARE
open path sampling : A Python library to facilitate path sampling algorithms.
gimlet : Graph Inference on MoLEcular Topology. A package for modelling, learning, and inference on molecular topological space written in Python and TensorFlow.
pinot : Probabilistic Inference for NOvel Therapeutics
COLLABORATORS
Daniel Heller (MSKCC): Experimental nanoparticle synthesis
Peter Bolhuis (Universiteit van Amsterdam): Transition path/interface sampling methodologies
Daniel Reker (MIT): Machine learning for nanoparticle formulation
PERSONNEL
Mehtap Isik (TPCB Graduate Student)
PUBLICATIONS
Yosi Shamay, Janki Shah, Mehtap Işık, Aviram Mizrachi, Josef Leibold, Darjus F. Tschaharganeh, Daniel Roxbury, Januka Budhathoki-Uprety, Karla Nawaly, James L. Sugarman, Emily Baut, Michelle R. Neiman, Megan Dacek, Kripa S. Ganesh, Darren C. Johnson, Ramya Sridharan, Karen L. Chu, Vinagolu K. Rajasekhar, Scott W. Lowe, John D. Chodera, and Daniel A. Heller. Nature Materials 17:361, 2018. [DOI] [PDF] [Supporting Info] [nano-drugbank]