Skip to main content

Thank you for visiting You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

  • ADVERTISEMENT FEATURE Advertiser retains sole responsibility for the content of this article

Software powered by artificial intelligence accelerates drug discovery

Despite continued increases in R&D spending, the five-year average number of new medicines approved by the US Food and Drug Administration (FDA) sits stubbornly at around 50 drugs per year1. As the cost of bringing a new drug to market increases, with estimates suggesting it has nearly doubled in the past decade2, it has become vital to improve R&D productivity to ensure the future viability of the pharmaceutical industry and advances in healthcare.

Many drug candidates fail before they enter clinical trials. Preclinical studies are designed to identify and validate a lead candidate. It is at this crucial stage of the process that 50% of spending is wasted, amounting to $28.2 billion in the United States alone each year3.

There are three main reasons for this:

Redundancy. Scientists often have no or poor visibility of experiments other scientists have performed in the past, even within their own organizations.

Experimental failure. Many experiments fail to produce usable data because of faulty reagents and protocols.

Irreproducibility. Even when an experiment seemingly succeeds, it often cannot be reproduced, perhaps because of problems with reagents, documentation and data gathering.

As a PhD researcher, Tom Leung, BenchSci’s CSO experienced first-hand the waste of resources in preclinical studies. When working as a cancer researcher, he lost rare patient samples because of an inappropriate antibody. Problems with antibody selectivity and specificity are widely acknowledged as one of the main causes of incorrect, inconsistent and irreproducible results in life sciences.

Keen to tackle the problem, Leung teamed up with David Q. Chen, Elvis Wianda and Liran Belenzon at the University of Toronto, Canada, to apply AI to antibody selection. In 2015, they founded BenchSci and started building AI-powered applications that analyse the full text of scholarly articles to extract relevant information for choosing the most suitable reagent for a particular experiment.

In 2017, BenchSci launched its first application: AI-Assisted Antibody Selection. “Scientists used to have to spend hours or even days reviewing publications to find the antibody that would be most likely to work in their experiments. With BenchSci they can get antibody validation information within minutes,” said Leung.

Since then, BenchSci’s platform has expanded to encompass more types of reagents as well as model systems, including recombinant proteins, RNAi, cell lines, CRISPR, animal models and PCR. Today, more than 49,000 scientists in 16 of the top 20 pharmaceutical companies and 4,500 academic institutions use BenchSci’s AI-Assisted Antibody Selection or Reagent Selection applications to plan more successful experiments, with proven savings of millions of dollars per year.

How the BenchSci system works. Accelerating R&D, the system above enables faster reagent selection, more reproducible experiments and increased research capacity.

Running fewer but better experiments

By helping researchers select the most-appropriate reagents in seconds, BenchSci aims to halve the time it takes for new drugs to reach the market by 2025.

“Roughly 80% of experiments in the preclinical phase of the drug development process are unnecessary to progress a novel medicine to clinical trials,” said Leung “Performing only the 20% that are necessary would mean carrying out an average of 1,400 versus 7,000 experiments—so new drugs would get to patients much faster and be much cheaper.”

BenchSci’s approach involves:

• The collection of terabytes of data from sources including scientific publications, reagent catalogs, and customer databases. These include closed-access publications, thanks to partnerships with publishers.

• The use of more than 100 proprietary machine-learning models that decode millions of unstructured scientific documents and extract insights from images, as well as text.

• A user-friendly interface that organizes the resulting information using a proprietary ontology. This allows researchers to understand the biomedical significance of the extracted data and establish relationships between biological entities. Users can easily search by target and reagent- or model system-type to obtain experiment-specific reagents and compare their specifications in minutes.

Meeting customer needs

Building on the success of current applications, BenchSci is evolving its technology to provide even more comprehensive solutions to help leading pharmaceutical companies solve their biggest R&D challenges.


  1. Mullard, A. Nature Rev. Drug Discov. 21, 83–88 (2022).

    Article  PubMed  Google Scholar 

  2. Statista

  3. Freedman, L. P., Cockburn, I. M. & Simcoe, T. S. PLoS Biol. 13, e1002165 (2015).

    Article  PubMed  Google Scholar 

Download references


Quick links