Canadian start-up BenchSci was founded with the mission of helping scientists to rapidly find the appropriate antibody for their drug discovery experiments. Using AI, the company can examine millions of research papers and tell scientists which antibodies are likely to provide useful results in the experiments that they’re planning.
BenchSci is the brainchild of Chief Scientific Officer Tom Leung, a University of Toronto alumnus who is acutely familiar with the headache of failed experiments owing to the use of antibodies that were unable to detect levels of target proteins.
“It becomes frustrating because, after weeks of preparing and growing cells and collecting samples, the experiment would fail – not because I did something wrong in the procedure, but because the antibody wasn’t good at detecting the protein I was looking for,” said Leung to UofT News.
“That really prompted me to think that there has to be a better way for scientists to assess the quality of an antibody product before buying it.”
BenchSci’s AI can reportedly deliver a result in as little as 30 seconds with over 93% accuracy. This could knock months off the time involved in getting a drug to market, which could make all the difference for patients.
In the space of five years, BenchSci has grown to a 120-member team and has achieved $56 million in funding from sources including Google’s Gradient Ventures. Their clients include over 4,300 academic labs and 15 of the world’s top 20 pharmaceutical companies.
Antibodies and the selection process
Antibodies are produced by the immune system in response to foreign substances like bacteria or viruses, and they are an invaluable part of the drug discovery process. Their biological function is to recognize and bind to specific parts of a foreign substance, called antigens, and that specificity is commonly exploited in the lab to target specific biomarkers to monitor their concentrations or locations.
They help scientists probe for information about disease markers and their response to new drug candidates. But for any given biomarker, there may be thousands of potential antibodies and hundreds of commercial vendors, and each may come with benefits and drawbacks for different applications.
Quality testing and validation by vendors is limited, so the task of figuring out what antibody to use falls to the researchers, who must manually sift through mounds of literature to identify potential candidates. Data is often scattered across many sources and many variables need to be taken into account, like the experimental context.
“Scientists know that every antibody is not going to work in every experimental context,” said VP of Science, Casandra Mangroo, to Nature. “Even if the vendor has done some sort of testing, they don’t have the capacity to test every antibody in every single experimental context.”
Making things even more complicated, the same product can have different names since vendors frequently relabel them when trading with one another, and search engines like PubMed may not be able to account for protein synonyms.
Besides the time and effort lost to this activity, it can also be inaccurate: 36% of all failed experiments in preclinical R&D involve inappropriate reagents, which include antibodies. This leads to frustration, wasted funding, and delayed treatments for people in need.
Furthermore, around 42% of drug development spending goes into preclinical R&D, and some $28 billion is lost every year due to irreproducible research.
The BenchSci founders realized that machine learning may be the solution as it could enhance the searchability, accuracy, and completeness of available reagent data which would, in turn, speed up research timeframes by months.
Scanning millions of papers in seconds
BenchSci’s Antibody Selection platform automates the process of scanning scientific literature and can analyze both images and text to identify antibodies that may support a certain experiment. The available database contains 10.2 million antibodies from more than 256 commercial vendors and coupled data drawn from 11.1 million papers. BenchSci continues to work on expanding the dataset by establishing relationships with more journals and increasing the system’s computational capacity so it can keep up.
“We do monthly runs to update data on the platform and train our algorithms. We basically had to create a whole wall of computers to do that,” said Mangroo to Nature.
From the scientist’s perspective, they can search the database for antibodies against a particular protein and the AI will put together a set of figures depicting the use of different antibodies in various experimental contexts. They can look up certain antibodies that have been used in experiments in a particular subfield and then compare the performance of different antibodies across related studies.
Besides product selection, the tech can also expedite experiments by identifying a published track record for a certain antibody in a specific validation test, and that means researchers don’t have to waste time repeating the same experiment.
There were two important factors in making this technology possible. Firstly, capitalizing on the cutting-edge of AI as developments occurred was vital: “if this idea was hatched two years earlier, it probably wasn’t going to work because deep learning and machine learning were not as mature yet,” said Leung to Nature.
Secondly, the large repository available on open-access publishing domains like PubMed Central gave the fledgling company a foothold to begin training their machine learning algorithm. The team’s quick success gave them purchasing power when developing connections with prestigious closed-access journals like Nature, ultimately building up a formidable database of biomedical experimental data.
COVID-19 and expansion plans
When the pandemic hit home around March of 2020, BenchSci announced their strategy to help out researchers by adding more reagents beyond antibodies to their database for COVID-19 studies. They achieved this by using their AI to identify novel reagents related to the virus, and also by uncovering relevant experimental design insights. The resulting data were made available for free to the scientific community.
This is part of a wider trend for BenchSci, who are working to expand into other types of reagents like recombinant proteins for general purposes to create a more comprehensive and powerful product. These plans were made possible by the latest round of funding, which was announced in early 2020.
“By linking these other reagents together, we can provide a much more complete picture of what has transpired in different publications,” said Leung. “Then we’ll be able to help scientists in a much fuller regard to planning their experiments.”
