Pin-Kuang Lai and Lateefat Kalejaye Create DeepViscosity, a Powerful AI Tool to Improve Antibody Drug Development
Stevens assistant professor and Ph.D. student teamed up with AstraZeneca to help researchers develop more accessible life-saving therapies
Through a groundbreaking new study, Pin-Kuang Lai, assistant professor in the Department of Chemical Engineering and Materials Science at Stevens Institute of Technology, and Lateefat Kalejaye, chemical engineering Ph.D. candidate, Class of ’27, have partnered with global pharmaceutical leader AstraZeneca to help overcome one of the toughest challenges in therapeutic drug development.
Published in mAbs, a top journal in the field of therapeutic antibodies, their findings introduce DeepViscosity, a powerful artificial intelligence model to predict whether an antibody solution will be too thick (viscous) to be delivered by an under-the-skin (subcutaneous) injection. It’s a critical milestone toward creating injectable antibody medicines that are safe, stable and simple for patients to use.
“We have created a powerful, accurate tool that can predict viscosity, which is one of the biggest barriers to patient-friendly therapies,” Lai said. “DeepViscosity has the potential to improve efficiency, reduce the risk of late-stage failures and even enable re-engineering of problematic antibodies through early mutations. It’s an innovative and proactive approach to therapeutic design.”
It’s clear why viscosity matters
In contrast with the hassle and stress of intravenous (IV) infusions such as chemotherapies that must be given over a period of time in hospital settings, self-administered under-the-skin injections like insulin for diabetes can fit more seamlessly into the lives of patients dealing with chronic conditions. Being able to pick up the prescribed medication from the pharmacy and take it independently at home can greatly improve treatment adherence and quality of life.
But antibodies are large, complex biological molecules. Drugs using these germ-fighting proteins often can’t balance the thick consistency of high concentrations needed for subcutaneous delivery with the thin profile required to pass through a small needle. They become too thick, too difficult to develop and produce, and too painful or even impossible to administer.
And for drugs still in the early design phase, the lab data hasn’t been available to predict which antibodies will have these issues.
Until now.
DeepViscosity, the Stevens-AstraZeneca team’s innovation, was trained on a massive 229 antibody sequences, the largest-known dataset of these molecular blueprints, through 102 deep learning networks. The diverse dataset allowed Lai and Kalejaye to train and rigorously validate the model.
Using just the antibody’s sequence, DeepViscosity can rapidly screen and predict — with close to 90 percent accuracy — whether the solution at high concentration will be an easy-to-inject low-viscosity option suitable for subcutaneous delivery, or a high-viscosity therapy that may be unsuitable for this route
“The goal is to screen out problematic antibodies early in the pipeline before wasting time and other resources,” Kalejaye explained. “That way, we can either redesign the antibody or choose a better candidate.”
AstraZeneca provided the antibody sequences with the experimental viscosity data used to train the model, as well as guidance from extensive experience in drug development.
“Our lab is predominantly computational, and our colleagues at AstraZeneca served as our experimental counterpart,” said Lai. “They also shared perspectives into the modelling and training techniques from a data science perspective. They offered valuable insight into real-world developability challenges, which helped shape the model’s design to ensure it would be truly useful in an industry setting.”
This isn’t the first time AstraZeneca has worked closely with Stevens. Notably, Stevens alumna Pam Cheng '92 M.Eng. '95, AstraZeneca executive vice president, shared her insights during Stevens’ Dean’s Lecture Series, including a 2020 presentation on the company’s COVID-19 vaccine trials.
Collaboration — through thick and thin
DeepViscosity is freely available, inviting scientists, developers and others to improve and expand the tool.
Already, AstraZeneca and biotechnology company Apoha are using DeepViscosity in their drug development pipelines. The Stevens team also aims to extend the model to cover other types of therapeutic proteins.
This project highlights Stevens’ strengths in artificial intelligence, data science and biomedical engineering — all key pillars of the university’s strategic vision. By applying advanced machine learning to real-world healthcare problems, the team demonstrated how Stevens research can bridge science and industry to improve lives.
“This project was key in shaping my scientific development and advancing my goal to enhance computational approaches in pharmaceutical research,” Kalejaye said. “Leading this team gave me the unique opportunity to translate academic research into industry applications, seeing how this tool can empower formulation scientists to design better therapeutics early in development. It deepened my passion for drug development and reinforced the importance of collaborative, cross-disciplinary research in solving complex therapeutic challenges.”
Lai has shared this work at major conferences, including the American Association of Pharmaceutical Scientists National Biotechnology Conference as the keynote speaker, and the PEGS Boston Protein Engineering & Cell Therapy Summit. Kalejaye will present this work at the American Institute of Chemical Engineers’ fall 2025 meeting. This industry attention highlights that DeepViscosity is more than just a research project. It’s a breakthrough in the design of faster, smarter, more patient-friendly therapeutics to meet patients’ real-world needs.