Training a Search Engine to Read Like a Scientist

For many bench scientists who study infection and disease, looking for antibodies to use in an experiment can be like looking for the needle in a haystack. Antibodies are proteins that identify and remove bacteria or viruses from the body. They are critical diagnostic tools, so researchers can spend hours reading through hundreds of scientific papers just to find that one specific antibody they require.

BenchSci, a U of T-based startup, is making this time consuming process easier. It is a machine learning search engine that can “read” more than four million open access research papers and help researchers find one of the 250,000 antibodies that are in the database. The goal is to take a process that once could take several days and reduce it to a matter of minutes.

Maurice Shen, a pharmacology and toxicology grad, is on the BenchSci team. He spoke with Faculty of Medicine writer Dan Haves about this project.

Tell me a little bit about the history behind BenchSci.

As former bench scientists, we know the frustration caused by using the wrong antibodies: the loss of time, resources, and research funding. In order to mitigate this risk, we always turned to peer reviewed papers. We would spend several hours and sometimes even days reviewing papers and figures just to find that one antibody that has been proven to work by other scientists. We asked ourselves, why is it that in 2014 – when BenchSci was conceptualized – we still could not find antibody usage data efficiently? We then assembled a team of bench scientists, computational biologists, machine learning experts, bioinformaticists and experienced entrepreneurs to work on a solution. As a result, BenchSci was born.

How are you using machine learning to identify these antibodies?

We use machine learning to analyze sentences and figures in publications. We train the algorithm to read research papers like a scientist. This enabled us to automatically extract relevant figures from papers and decode the specific experimental contexts associated with each antibody use-cases.  

If I’m a researcher looking for a particular antibody, how do I use BenchSci to help me find it?

BenchSci helps research scientists find published figures that contain antibody usage data. The figures are indexed, making them filterable by different variables such as applications, tissues, cell lines, species, diseases, and more. In addition, we also provide scientists an overview of how often an antibody was used for each type of application in the literature. With BenchSci, scientists can find antibodies that have been proven to work in publications in minutes. 

What sort of responses have you heard so far from those who’ve used the platform?

We’ve received very positive feedback from research scientists and antibody vendors. To this day, we have signed up over 5,000 scientists to use our platform. We’ve also partnered with more than 100 antibody vendors, including leading companies like Cell Signaling Technologies, Bio-Rad, EMD Millipore, Novus, R&D Systems, and more. 

Are there any plans to expand BenchSci’s capabilities in the future?

Antibodies are just the first step in our journey to help research scientists conduct more successful experiments. We plan to expand to other reagents such as siRNAs, chemical inhibitors, viral vectors, and more.