Grant Armstrong Email & Phone Number

United States

Recognizing the role of organized code in the research ecosystem to enhance reusability, reproducibility, and collaboration, I lead the restructuring of our code bases utilizing the GitHub Teams environment.Leveraging features such as pull requests, issues, project boards, and teams, I created a collaborative environment that facilitated communication, code review, task tracking, and professional documentation. This not only improved the overall organization of our research activities but also provided a transparent platform for team members to communicate and share our projects with external stakeholders. As part of these foundational efforts, I actively led and mentored an undergraduate student in a cutting-edge research project utilizing the Git workflow. This hands-on mentorship not only facilitated their professional growth but also contributed to the overall technical skill development within the team. This inclusive approach to research collaboration accelerated our individual projects and fostered a dynamic and supportive environment to contribute to the growth of the lab as a whole.

Bethlehem, Pennsylvania, United States

The Lehigh Informatics Lab aims to use machine learning models to reveal the biochemical mechanisms of binding specificity, which is how proteins recognize and bind their partners. Binding specificity depends on the molecular characteristics of proteins, such as their shape, charge, and hydrophobicity, and is crucial for the regulation and function of many biological processes, such as signal transduction, enzyme catalysis, and gene expression.My research examined how sequence mutagenesis affected hydrogen bonding patterns for 7000 single and multi-point mutations of 346 protein-protein complexes in the Skempi v2 data set. I used a computational geometric algorithm to measure the volumetric profile of interacting hydrogen bond acceptors and donors at the PPI interface. Together with engineered electrostatic and steric features, my team built a random forest decision tree machine learning model to estimate the effect of those mutations on binding affinity. Our results are awaiting publication but show state-of-the-art performance compared to similar methods

▪ Oversaw residential halls of upper-class students and helped them navigate to professional, social, and academic success by providing support for individual development.▪ Coordinated monthly events and signage to foster an inclusive community and connect residents to on-campus resources.▪ Navigated the interpersonal needs of students and helped them resolve conflicts

Addressed the issue of incorrect covalent bond recognition that can occur when RDKit and MDAnalysis convert molecules into their native software representations.MDAnalysis does not have built-in functions for sanitizing molecules, checking valences, or removing small fragmetns, as RDKit does.Therefore, these erroneous covalent bonds remain in the MDAnalysis universe unless explicitly handled. RDKit, while capable of identifying atoms with incorrect valences, lacks a sophisticated mechanism for managing such instances. This can lead to crashes during the conversion of a pdb file to an RDKit molecule.My program establishes a custom method for managing incorrect valences between and within protein residues. It uses amino_acid_connectivity.json, a dictionary containing the "ground truth" covalent connectivity of all 20 amino acids and the protein backbone, to screen the local bonded neighborhood of an atom RDKit flags as having incorrect valence, and subsequently removes the incorrect connections from BOTH the RDKit molecule and MD Universe objects.

Tarrytown, New York, United States

During my time at Regeneron, I took the initiative to propose and execute a groundbreaking project focused on applying geometric deep learning to the early-stage development of therapeutic antibodies. I played a pivotal role in implementing a scalable end-to-end data preprocessing and model tuning pipeline that leveraged cutting-edge techniques such as multi-headed Graph Attention (GAT), custom encoder and decoder architectures, and multi-GPU distributed training. I exhibited a great degree on independence and reached out to various departments within the company whenever there was a scientific need for the project. The collaborations I built were beneficial for my team and will contine into the future.Early results from experimentation with the complete model indicated near chemical accuracy with a MSE of 1.86 (kcal/mol)^2 at inference time when predicting binding affinity. As the model development process is inherently iterative, we extract insights from each round of experimentation and refine the model in subsequent phases.

Combing sports analytics and unsupervised machine learning, I developed an automated Expectation-Maximization (EM) algorithm tailored for soccer match role labeling. This innovative approach harnessed the power of probabilistic modeling to achieve precise player role identification within time series data, and effectively automates an otherwise time consuming task. The EM algorithm, served as a powerful tool to iteratively minimize the 2D probability distribution describing each of the 11 players on the field.With this approach, I conceptualized team dynamics as a multi-agent interaction problem, where high-level positioning decisions of agents indicated role-role interchange and player flexibility independent of player identity. This distinction between player identity and role proved to be a key element in accurately capturing the nuanced strategies employed during soccer matches.Beyond its immediate application to situational awareness, I demonstrated the algorithm's versatility and potential for future advancements. The envisioned applications extended to deep imitation learning for the development of intelligent game-playing agents. This innovative work not only showcased the algorithm's effectiveness in the realm of sports analytics but also laid the foundation for further exploration into cutting-edge applications at the intersection of machine learning and sports strategy.

Graph Neural Networks for Cheminformatics: Explainable AI▪ Built proficiency with deep learning models by interpreting technical papers, workshopping open-source code, tuning hyperparameters responsible for the model state, and adapting the model for continuous and discrete predictions ▪ Implemented GradCAM class activation mapping to explain the predictions of a molecular GCNN trained on a dataset containing qualitative toxicity measurements for 8k compounds on 12 different targets, including nuclear receptors and stress response pathways ▪ Predicted log solubility directly from SMILES molecular structure for 1128 compounds, and generated heat maps which show both positive and negative contributions from nodes, i.e. atoms, for the final prediction

Bethlehem, Pennsylvania, United States

The Wonpil Im Research Group has a long-standing focus on computational chemistry and is the key developer of the CHARMM-GUI environment for the automated construction of input files for molecular dynamics (MD) simulations. The Charmm36 force field is a set of parameters and equations that describe the interactions and motions of atoms and molecules in MD simulations. It is effective at calculating the forces and energies acting upon organic atoms commonly found in living organisms, but it lacks accurate parameters for many inorganic atoms, such as silver (Ag), which otherwise have elusive behaviors.Silver nanoclusters are a class of nanoparticle that are known to have cytotoxic effects on gram-negative bacteria, which have an outer membrane that protects them from many antibiotics and can thus develop resistance. While the effects of Ag nanoparticles on gram-negative bacteria have been observed experimentally, a biophysical explanation for the mechanism of cytotoxicity has not been reached by consensus.I conducted individual research to generate novel CHARMM force field parameters using ab-initio calculations of frequency and potential energy scans of structural motifs conserved between 20 different silver nanoclusters obtained from the CCDC database. With these parameters in place, it becomes possible for the scientific community to model the behavior of Ag nanoclusters in the presence of the gram-negative outer membrane.

I performed a comparative physicochemical analysis on 18 all-atom, model membrane systems with native-like lipid compositions corresponding to eukaryotic, bacterial, and archaebacterial membranes, as well as three single-component lipid bilayers. To gain deeper insight into the influence of sterols and lipid unsaturation on the structural and mechanical properties of these bio-membranes, I organized meetings with academic experimentalists and completed trajectory analysis on over 1 microsecond of simulation time. As a core first author, I drafted the manuscript and completed revisions, delegated the completion of milestones to team members, and visualized data for a scientific journal audience.

Master'S Degree, Data Science

Activities and Societies: President Machine Learning ClubTook courses in Convex Optimization, Accelerated Computing with NVIDIA GPUs.

Cognitive Science

Summer 2020 Non-Degree Seeking Student

Self-elected study of: ORGANIC CHEMISTRY I COMPUTER SCIENCE I - Programming and Problem Solving with C++ MA 244 - Intro to Linear Algebra

Bachelor Of Science - Bs, Cognitive Science, 3.54

Activities and Societies: The Student Organization for Cognitive Science (SOCS) Independent Study in Cognitive Neuroscience under.