Accelerating Antibody Discovery with AI
Pharmaceutical Industry
SUMMARY
Our client, a well-known pharmaceutical company, aimed to employ AI to design antibody protein sequences.
Hop brought key expertise to this project – multi-objective optimization.
In addition to solving the optimization problem, Hop was able to supply extra help to train a deep learning model.
The model our team trained performed with accuracy 100x better than initially targeted.
Our client is now poised to design and generate its first in silico antibodies.
THE COMPANY
Our client is one of the oldest and largest pharmaceutical companies in the world, focused on oncology, neuroscience, gastroenterology, immunology, and plasma products. Their focus on improving the patient experience and advancing a new frontier of treatment positions them as pioneers at the cutting edge of technology, leveraging the latest developments to advance their impact.
THE CHALLENGE
Our client aimed to employ AI to design antibody protein sequences. Conventional lab-based sequencing was expensive and slow, and the company’s executives saw great potential for AI to not only accelerate the process and save money, but design sequences that were more suitable for use as drugs than those coming from the lab. At the time they engaged Hop, this project was already underway, spanning multiple contracting teams and several technical internal contributors. However, key expertise was missing – multi-objective optimization, which would implement the algorithms that traverse the vast universe of possible antibody sequences. Hop had the experience to tackle this piece of the puzzle.
“I see Hop as a company that basically helps de-risk research. If there are big questions like, ‘Should we invest in this?’, ‘Is the technology mature enough?’, or ‘What are the capabilities?’, I think Hop is a good resource.”
THE APPROACH
This optimization problem was a technically challenging one: high-dimensional, multi-modal, multi-objective, and with noisy objective function estimates. The Hop team quickly reviewed the literature on algorithms for solving optimization problems with these properties and implemented a solution that incorporated the most common algorithms.
As we focused on optimization, our team also flagged a major risk to overall project progress, stemming from challenges arising in training a deep learning model. As it became clear that additional help was needed in order to train the model within the expected timeframe, Hop brought in another team member to pitch in. An expert in the deep learning field with more than a decade of experience, within a few short weeks he implemented a training pipeline for the model and kept the project moving forward.
“Your team provides a service and skill set that I think is hard to find in other vendors, and I really enjoyed working with the people that you provided.”
THE RESULTS
The performance of the model our team trained far exceeded expectations, with accuracy 100x better than initially targeted on a general dataset. This model artifact supports a broad, flexible antibody design process. Additionally, because of its performance and generality, the model can support the work of other teams around the organization’s research division.
With the optimization algorithm implemented and a model trained, this project primed our client to design and generate its first in silico antibodies, leveraging the latest technology for not only significant savings in time and money, but acceleration in drug discovery that could save lives.
Looking to accelerate your org’s antibody discovery process? Contact us to learn how Hop can help.