Silk Fibroin Microspheres: An Innovative Approach to Improve Drug Delivery to the Lungs for the Treatment of Neonatal Respiratory Disease Syndrome (NRDS)

Overview

Neonatal Respiratory Distress Syndrome (NRDS) is a breathing disorder that is the leading cause of death in premature infants. It occurs when an infant is unable to produce enough lung surfactant, causing insufficient oxygen exchange and lung collapse. The prevailing treatment for NRDS is delivering natural animal pulmonary surfactant to the lungs to affect oxygenation and lung compliance. Known as INtubation-SURfactant Extubation (INSURE), this procedure involves an unstandardized protocol and is often inefficient and ineffective due to its inability to maximize the delivery of lung surfactant to the alveoli. Emerging studies have suggested that a more effective approach to drug delivery is administration through microspheres, as aerosolized microspheres containing the drug can maximize the surface area of the lung that is exposed to the surfactant. Specifically, Silk Fibroin (SF) best facilitates microspheres drug delivery, as SF is biocompatible and requires little resources to fabricate. However, its applications are limited due to its inconsistent batch qualities and uncontrollable degradation rates. This study introduces a protocol for designing SF microspheres containing CUROSURF®, a popular pulmonary surfactant, through the simple yet effective process of phase separation between SF and Polyvinyl alcohol (PVA). The efficacy of this protocol was optimized by iterating the procedure based on characterizations of the resulting microspheres through scanning electron microscopy, dynamic light scattering, and light microscopy. This study offers insight into a new method of drug delivery for treating NRDS in infants that addresses the long-standing issues of lung surfactant delivery experienced in today’s intensive care nurseries. Robust Weight Initialization Using Graph Hypernetworks for Efficient Adversarial Training Simon Lee Whittle School and Studios, Washington, D.C. Deep learning achieves superhuman performance on a variety of image classification tasks, but it remains susceptible to adversarial examples: adding purposefully crafted, imperceptible noise to inputs severely damages a neural network classifier’s accuracy. The lack of robustness against such perturbations is concerning given the widespread adoption of deep learning in increasingly high-risk applications with potential for human harm such as autonomous vehicles. The state-of-the-art adversarial defense is adversarial training, but it significantly increases computational cost compared to that of standard training. This work proposes to mitigate this problem through initialization of neural networks with parameters predicted by a graph hypernetwork (GHN) before adversarial training. Previous research has already demonstrated GHNs’ ability to predict network weights that produce impressive accuracy on standard image classification tasks. These findings motivate the extension of GHNs to the more difficult task of adversarial training, where parameters that are robust to a variety of adversarial examples are predicted. To this end, a GHN was trained on the DeepNets-1M dataset of diverse neural network architectures, learning to predict weights that are robust against adversarial examples generated on the CIFAR10 image dataset. Experimental results show that initialization of neural networks with GHNs significantly reduces adversarial training time compared to that of conventional gradient descent training. The adversarially optimized GHN additionally outperforms vanilla GHNs by as much as 28 percentage points in resultant robust accuracy. Through lessening the computational cost of obtaining adversarially robust models, this methodology contributes toward making robust AI more environmentally friendly and accessible.

Competition history

• JSHS 2022 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.