Fine-tuned machine learning models for the discovery of dye nanoparticles with enhanced lung delivery

Abstract

Drug delivery using self-assembling drug-excipient nanoparticles offers a scalable strategy to improve the solubility, bioavailability, and biodistribution of poorly soluble therapeutics. However, progress in the development and translation of these materials has been limited by the lack of safe excipients capable of stabilizing drug-rich nanoparticles that allow for extrahepatic delivery. Here, we present an adaptive machine learning-guided workflow for excipient discovery that prospectively prioritizes candidate dyes and identifies clinically relevant drugs compatible with dye-based nanoparticle formulations. Guided by model predictions, we deployed semi-automated synthesis together with in vitro and in vivo experimental validation to evaluate and characterize the resulting nanoparticles. We identified two FDA-approved dye excipients and rediscovered a third dye as favorable excipients. Fine-tuning our machine learning models enabled us to identify 35 previously unreported nanoparticle formulations utilizing these dyes. Importantly, in proof-of-concept in vivo biodistribution studies, these dye excipients enabled the encapsulation of the anticancer drug sorafenib with distinct extrahepatic biodistribution profiles, including preferential accumulation in the lung. Taken together, this work establishes an adaptive machine learning-driven framework for excipient discovery and demonstrates that small-molecule dye excipients can serve as modular regulators of nanoparticle biodistribution. These findings expand the molecular and computational toolboxes for drug-excipient nanoparticle design and prototype the excipient selection process as a potentially powerful strategy to tune in vivo drug delivery.

Introduction

Clinical successes of nanoparticle-based medicines have demonstrated the critical role that nanoformulations can play in achieving safe and effective drug delivery [1], [2], [3]. These advances underscore the ability of nanoparticle systems to overcome fundamental delivery barriers such as poor drug solubility, payload instability, and inappropriate bioavailability and biodistribution [2]. However, broader nanoparticle translation remains constrained by fundamental material challenges, including reliance on complex, multicomponent formulations that often result in low drug loading [4]. Additionally, a central challenge in nanoparticle delivery is the limited control over tissue accumulation, which often results in disproportionate, non-specific uptake in clearance organs such as the liver and kidneys, reducing therapeutic efficacy and increasing the risk of off-target toxicities [5]. Together, these limitations highlight the need for simpler, modular nanoparticle systems that enable high drug loading while also offering greater control over their in vivo biodistributions.

Synthesis of drug-excipient co-aggregating nanoparticles represents one emerging strategy to address these challenges [6], [7], [8]. In these drug-excipient nanoparticles, small-molecule self-aggregating drugs are stabilized by small-molecule excipients, resulting in a suspension of small, uniform particles [9]. These nanoparticles are an emerging drug delivery platform due to their simple synthesis that is compatible with high-throughput experimentation, very high drug loading often close to or even exceeding 90%, and established computational design workflows that guide material selection and optimization [7], [8]. However, a major challenge in the development and translation of these materials is the small number of available small-molecule excipients capable of stabilizing them, which hinders our ability to effectively encapsulate several drugs [10].

Additionally, several of the previously identified excipients, such as Congo red, are not yet approved for human use and are under scrutiny for potential concerns around carcinogenicity, genotoxicity, and exposure toxicity [11]. This raises safety concerns that may limit their clinical translation as inactive ingredients [12]. Notably, prior studies have demonstrated certain dyes such as Congo red [8], [13], [14], IR783 [6], [15], [16], and Evans Blue [17] can stabilize drug-excipient nanoparticles. These findings suggest that dye molecules represent a promising class of small molecule stabilizers capable of forming nanoparticles with high drug loading compared to traditional polymer or surfactant-based carriers, which often require an enrichment of carrier material [18], [19]. Therefore, there is an urgent need to identify additional excipients with precedent for human use that can form nanoparticles with drugs that are currently limited by formulation challenges.

Here, we implemented an integrated computational and experimental platform to identify novel nanoparticle-forming excipients from compounds with FDA approval status. Using iterative train-predict-test cycles, we fine-tuned our machine learning model to prospectively prioritize promising excipients and to guide the automated design of new drug-excipient nanoparticles. This approach identified the FDA-approved excipients Brilliant Blue FCF and Fast Green FCF as effective dye stabilizers for nanoparticle formation. In addition, the model rediscovered Brilliant Blue G250, an FDA-approved dye previously evaluated for nanoparticle formation that had so far only shown limited potential [15]. Brilliant Blue FCF and Fast Green FCF are approved food dyes, while Brilliant Blue G250 is approved for internal limiting membrane staining in ophthalmic procedures, indicating prior human use and regulatory precedent for these materials [15], [20], [21], [22], [23], [24].

To accelerate experimental validation, we integrated machine learning with semi-automated robotic synthesis and high-throughput nanoparticle characterization via dynamic light scattering. Iterative model refinement improved predictive performance and enabled the identification of 35 previously unreported nanoparticle formulations across multiple therapeutics. Subsequent analytical, in vitro, and in vivo characterizations support the translational potential of these nanoformulations. Notably, proof-of-concept biodistribution studies of sorafenib nanoparticles formulated with these dyes revealed distinct tissue distribution profiles and pharmacokinetic parameters with enhanced delivery to the lung, suggesting that rational excipient selection can serve as a strategy to modulate nanoparticle biodistribution.

Together, this work establishes an adaptive machine learning-guided framework for excipient discovery within drug-excipient nanoparticle systems. By integrating computational prioritization with automated synthesis and experimental validation, we expand the repertoire of FDA-approved excipients capable of stabilizing drug-rich nanoparticles. Additionally, our proof-of-concept in vivo findings suggest that rational excipient selection might serve as a modular strategy for tuning nanoparticle biodistribution and pharmacokinetics for these materials, thereby enabling greater control over the biological fate of these drug-rich nanoformulations and advancing their translational potential.

Download the full article as PDF here Fine-tuned machine learning models for the discovery of dye nanoparticles with enhanced lung delivery

or continue reading here

Lauren A. Onweller, Rebeca T. Stiepel, George A. Cortina, Joseph R. Laforet, Ivan Spasojevic, Ping Fan, Daniel Reker, Fine-tuned machine learning models for the discovery of dye nanoparticles with enhanced lung delivery, Journal of Controlled Release, Volume 398, 2026, 115226, ISSN 0168-3659, https://doi.org/10.1016/j.jconrel.2026.115226.

See our next webinar: