BIOVIA is proud to continue the collaboration with UC Berkeley’s Master of Molecular Science and Software Engineering (MSSE) program, now in its third year of sponsoring capstone projects that bring together industry expertise and emerging computational science talent. The first capstone in Spring 2024 produced CalMPNN, a Python module for protein sequence design and stability prediction. For Spring 2025, a new team tackled a foundational challenge in computational chemistry: using machine learning to automate atom type assignment in molecular force fields.
The Challenge
Before a molecular dynamics simulation can model how a drug interacts with its target, every atom in the molecule must be classified into a specific “atom type” that determines the physical parameters (bond lengths, angles, charges, etc.) governing how the molecule behaves. Traditionally, this classification relies on rule-based tools that can be brittle, slow on large datasets, and difficult to extend. The Spring 2025 capstone team set out to replace that approach with machine learning.
What the Team Built
The result is atoMLtype, a modular Python toolkit that learns the mapping from molecular structure to atom type in the widely used GAFF2 (General AMBER Force Field 2) framework. The team implemented and compared multiple modeling approaches, from a Random Forest baseline using hand-crafted atomic features to several Graph Neural Networks that operate directly on the molecular graph. They also developed thoughtful solutions for handling chemical symmetries in the force field’s type system, demonstrating real depth of understanding of both the underlying chemistry and the modeling challenges.
The toolkit provides a complete workflow: loading molecular datasets, training and evaluating models, and producing detailed analysis; all aimed at making atom typing faster, more scalable, and more accurate for computational chemistry pipelines.
Fresh Perspectives, Real Impact
As with the first capstone, my colleague Reed Harrison and I served as project mentors, guiding the team through the intersection of software engineering best practices and scientific rigor. These collaborations continue to demonstrate the value of pairing BIOVIA’s domain expertise with the energy and fresh thinking of Berkeley’s MSSE students..
Brandon Robello, one of the students on the team, reflected on the experience:
This project was an incredible opportunity to apply modern machine learning techniques to a real-world scientific challenge. Working with Dr. Rohith Mohan and Dr. Reed Harrison on atom-type prediction from molecular graphs sharpened my skills in model development and software engineering, while deepening my appreciation for the complexity of chemical representation.
Jeremy Millford, another team member, shared a similar enthusiasm:
Exploring the project, learning new tools, and creating an impactful deliverable was a wonderful opportunity to let our scientific curiosity and new skills run wild. The balance between the freedom to explore and the structure of a professional assignment made the experience both highly valuable and genuinely fun. I found myself proactively wrapping up coursework, so I could work on it just a bit more. I also feel like I learned a ton.
Looking Ahead
Our partnership with UC Berkeley’s MSSE program continues to be a rewarding experience for both sides; providing students with real-world challenges at the forefront of computational science while bringing innovative approaches into BIOVIA’s work in drug discovery. We look forward to continuing this collaboration and seeing what the next cohort of talented students will build.
📩Want to find out the latest news about BIOVIA events, customer stories, blogs and more? Join the BIOVIA newsletter today!
Rohith received his Ph.D. in Bioengineering from University of California, Riverside. As part of his doctorate research he studied the complement system and the molecular basis of immune-related diseases by using MD simulations for mechanistic studies, peptide design, and virtual drug screening. He joined BIOVIA in 2018 as part of the Discovery Studio R&D team for the development of modeling and simulation tools for drug design and discovery. He currently leads the Protein Modeling and Simulation team in Discovery Studio where he works on force field parameterization, free energy methods, simulation methods, and the intersection of AI/ML in these areas. Rohith is also a fellow of the BIOVIA Science Council.