Performing computational molecular dynamics (MD) simulations of small-molecule systems has become one of the most prominently used methods in studies of molecular structure and behavior. However, MD is a computationally expensive and time-consuming methodology because of the requirement of computing detailed interactions among atom-atom pairs. There is great interest, therefore, in reducing the time and computational power needed to approximate real-world systems. Most commonly, such efforts have employed machine learning techniques to predict extensive properties of molecular systems. Here, we propose accelerating simulations by predicting conformational changes - a prospect that has not yet been fully explored. Previous work attempted applying a linear dynamical analysis algorithm named Dynamic Mode Decomposition to MD data, which has been shown to be ineffective through a multiresolution analysis. We propose herein the use of Sparse Identification of Nonlinear Dynamical Systems (SINDy), a nonlinear model which has been shown to accurately decipher the governing equations of dynamical systems. We will be testing the effectiveness of SINDy with MD data by performing an iterative error analysis while varying the initial parameters of the dataset, thereby gaining a better understanding of how much data (and in what form) should be inputted to maximize the accuracy of a simulated SINDy model of an MD dataset. If shown to be sufficiently accurate, we then can implement SINDy simultaneously with MD in an active learning loop to save time and computational power while maintaining a high degree of predictive capability for peptide conformations. The current goal is to obtain a deeper understanding of peptide conformational changes that could, in the future, be combined with machine learning techniques to greatly accelerate classical MD simulations.

This project is supported by the UW Computational Neuroscience Center, and the DMREF Program of NSF through the MGI platform under DMR# 1629071, 1848911, and 1922020.

The goal of this project is to develop a dynamically evolving connectionist model that more closely resembles the brain through its information-processing. Over the years, AI has shifted from the first generation of feedforward systems to the use of recurrent or convolutional Neural Networks. The third and newest generation of AI models, the brain-based models, and the Spiking Neural Network (SNN), attempts to bridge the gap between Neuroscience and ML using biologically realistic models like Θ-model, LIF, Izhikevich, HR, HH. These models, however, are still a black box leaving very little control or understanding on the learning process within the system without the access to the inner structure of the network. In addition, these systems are highly inefficient, slow, and very complex due to the limitations imposed by the hardware and explicit simulation of partial differential equations. Real world problems require “flexible learning and dynamically adaptive connectionist systems” that are capable to adapt and accommodate new input in real time. Current solutions have focused on varying the weights within a system rather than focusing on how connections within the system are formed. Based on our understanding from organismal brain structures, our approach, called biomimetic information codec, .bic, is a morphologically-adaptive coding hierarchical network that form in accordance with energy minimization - driven by dissipation of "heat" generated by the training data - constructing cortices and connectome for processing of information. My first objective herein is to quantitatively compare detailed structures between biological (fly brain) and .bic. networks using a random matrix approach.

The pupillary light reflex (PLR) curve is an important point-of-care biomarker for the diagnosis of traumatic brain injury (TBI). Using PLR, first responders can determine the severity of TBI in the field and direct patients to a trauma center where staff can continually assess PLR to monitor TBI severity. Manual pupillometry, the most commonly available method for first responders and most clinicians wishing to assess PLR, is qualitative and often inaccurate. The current gold-standard device for PLR measurement is digital infrared pupillometry, but such devices are fragile and expensive. Our research team has developed a smartphone-based pupillometer (PupilScreen) with the ability to assess PLR using a standard iPhone, assisted by a cloud-based neural network. To demonstrate the feasibility of using PupilScreen in a realistic clinical setting and compare the accuracy of the device to the current clinical gold-standard, we have built an annotated dataset of the PLR in n=120 patients with TBI who are hospitalized in a neurological intensive care unit. Pupillometry is performed using the mobile device and the gold-standard digital infrared pupillometer. Pupil videos are manually annotated and used in the further training of our machine learning algorithm that generates a PLR curve for each patient. We anticipate that our technology will demonstrate accuracy in assessing the PLR that exceeds that of manual pupillometry and is at least equivalent to the gold-standard digital pupillometer. This technology has the potential to alleviate the current undertreatment of many TBI patients in the United States and abroad that results from a lack of accurate and cost-effective pupillometry equipment.

Single amino acid conformational preferences on a substrate are invaluable to our understanding of how conformational propensities are dictated by a peptide’s sequence. The efficiency of understanding SAP/SLAM (self-assembling peptides on single layer atomic materials) interfaces is of high importance. An efficient way of performing computational modeling of a peptide’s free energy landscapes is needed to predict the folded structures on solid surfaces towards designing bio/nano interfaces, a key to bioelectronics and biosensor developments. Graphene, single-atomic layer graphite, is an ideal substrate for a dodecapeptide to bind and spontaneously self-organize to form ordered biomolecular structures on the surface. For the purpose of shortening computation times of peptides at graphene interfaces, sampling each amino acid’s free energy landscape in terms of the peptide’s natural torsional configuration, such as the phi and psi angles of the peptide backbone on graphene is studied. This is due to the fact that a predetermined natural starting point, such as the peptide’s lowest energy structure will allow for quicker convergence of the system and more accurate structure prediction. In an effort to retrieve this data, here we simulate each amino acid using the enhanced sampling computational technique Metadynamics. Using the computational modeling and random sampling of the lowest energy wells, we aim to aid in the determination of low energy preferences on conformational landscapes on graphene towards more predictive design of soft bio/nano interfaces for practical implementations. 

The purpose of my research is to explore the protein-ligand binding interaction by using sequence encoding and signal analysis to process and understand amino-acid sequences. Applications of this technique would be useful in a variety of biomedical fields, but more specifically in creating a platform for a unique and streamlined vaccine candidacy process. If we can encode protein/DNA sequences using electron-ion interaction-potential, then by applying various signal processing functions we can more easily identify the meaning of these complex structures. The goal in doing this would be to relate simple amino-acids, the building blocks of our genetic sequences, to numerical values which we may manipulate. Combining these inferences with a convolutional neural network, the result would create a non-empirical and efficient method of understanding protein folding and bimolecular binding interaction, specifically through predicting pIC50 binding affinity. The methodology begins with accessing a large set of human amino-acid sequence data, transforming the data using signal processing, and learning from the data to understand similarities between sequences and predict binding affinity. The anticipated results will initially be the neural networks ability to predict binding affinity. Further results and experimentation would involve tuning the model to be more adaptive and testing new data from the SARS-Cov-2 virus. The more variety of human data the model may learn from, the more adaptive and accurate it will predict pIC50 values in an efficient, non-empirical manner.