Yehor Mishchyriak

SAWNERGY is a high-performance toolkit for turning molecular-dynamics trajectories into machine-learning–ready graph data. The project is open-source and available though . Documentation is hosted on my website.
SAWNERGY builds residue interaction networks (RINs) from cpptraj outputs, extracts attractive and repulsive interaction channels, prunes and normalizes them into transition matrices, samples massive random and self-avoiding walks in shared memory, and trains skip-gram/SGNS embeddings for every frame. The entire pipeline is reproducible, Zarr-backed, and designed for HPC workloads. The result: clean, compact vector representations of protein conformational dynamics -- ready for downstream tasks like structural clustering, frame classification, and functional prediction. What you are seeing in this short clip above is the residue interaction network of the full length p53 protein (on the right) constructed and visualized by SAWNERGY, and on the left is its embedding, learned from interaction patterns and visualized by SAWNERGY. Note the visual resemblance of both images despite the fact that the right one is x,y,z coordinates of amino acid residues and the left one is network embedding deduced entirely from interactions data and no spacial information.

PureML is a lightweight deep-learning framework built entirely on NumPy. It’s designed to feel instantly familiar to PyTorch users while staying transparent, minimal, and easy to read. The library includes a complete autodiff engine, modular layers, activations, losses, optimizers, data utilities, and a clean neural-network API — everything you need to prototype models without extra overhead.

It’s fast to learn, simple to extend, and built for teaching, experimenting, and understanding models at every level rather than relying on a black box. PureML is open-source and available through . Documentation can be found on my website or through PyPI.

Unlike TensorFlow and PyTorch, which rely on massive multi-language codebases, PureML is intentionally 100% NumPy-based and fully transparent. Despite that simplicity, it remains very fast thanks to extensive vectorization throughout the system. It’s ideal for rapid prototyping, education, and situations where you want full clarity into every stage of the computation -- and the codebase is straightforward to read, modify, and extend.

M2F (microbiome-to-function) is a modular toolkit that turns messy UniProt records into clean, machine-learning-ready data. It automates large-scale UniProtKB mining with efficient, rate-limited REST calls, cleans and normalizes free-text annotations using targeted regex extraction, and encodes both sequences and ontology terms into meaningful numerical representations. All results are stored compactly in a single Zarr ZipStore, making datasets easy to share and reconstruct. The entire pipeline -- from HUMAnN outputs or accession IDs to tidy feature tables -- is fully reproducible and designed for seamless use in ML models. Code is available at this repo.

This poster, created for the Summer 2024 research symposium at Wesleyan, represents the starting point of my work at ThayerLab. Since then, I've made significant progress and refined the project's direction. See my project named "SAWNERGY".

My three friends and I developed the entire system from scratch over an intense 24-hour period for a hackathon held at Wesleyan. We recorded this demo video literally 10 minutes before the projects were judged, which is why the quality isn't the best. That said, focusing all our effort on project development rather than cosmetics paid off: we won the internship prize track!
(link to the hackathon)

I worked on this project as part of my audio-visual machine learning class. Code as well as the final write-up can be found here.