Robotically driven experimentation system created to reduce cost of drug discovery

Researchers from Carnegie Mellon University (PA, USA) have created the first robotics-driven system to determine the effects of drugs on proteins, drastically reducing the number of necessary experiments.

Like Comment

A team from Carnegie Mellon University has created a robotically driven experimentation model, research that could lead to rapid and accurate predictions of the interactions between novel drugs and their targets, thus significantly reducing the costs associated with drug discovery.

"Biomedical scientists have invested a lot of effort in making it easier to perform numerous experiments quickly and cheaply," explained lead author Armaghan Naik. "However, we simply cannot perform an experiment for every possible combination of biological conditions, such as genetic mutation and cell type. Researchers have therefore had to choose a few conditions or targets to test exhaustively, or pick experiments themselves. The question is which experiments do you pick?"

The balance between performing experiments with confidently predicted outcomes and those without is a significant challenge in drug discovery, as it requires the researcher to consider a huge range of hypothetical outcomes simultaneously.

The researchers have previously described the application of “active learning”, a machine learning approach involving a computer repeatedly choosing which experiments to undertake, in order to learn from patterns it observes in the data. This approach had previously only been tested on synthetic or previously extant data, but the team’s current model builds on this work by permitting the computer to choose which experiments to perform. These experiments were then carried out using an automated microscope and liquid-handling robots.

The system studied potential interactions between 96 drugs and 96 cultured mammalian cell clones, tagged with different fluorescent proteins. 9216 experiments were possible, where each would consist of acquiring imagines of a cell close in the presence of a single drug. The algorithm was challenged to learn how proteins were affected in each experiment, without performing every one.

The first round of experiments ran each cell clone against a single drug, totalling 96 experiments, with images represented by numerical features that captured the protein’s location in the cell. At the end of each round, all suitable experiments were used to identify phenotypes that may or may not have related to a previously characterized drug effect.

The learner identified potentially new phenotypes independently as part of the learning process, by clustering the images into phenotypes, then forming a predictive model to allow the model to guess the outcomes of unmeasured experiments. The learner repeated the process for 30 rounds, completing 2697 of a possible 9216 experiments, progressively identifying more phenotypes and patterns.

The team determined that the algorithm was able to produce a 92% accurate predictive model, from only 29% of experiments conducted. Senior author Robert F. Murphy elaborated: "Our work has shown that doing a series of experiments under the control of a machine learner is feasible even when the set of outcomes is unknown. We also demonstrated the possibility of active learning when the robot is unable to follow a decision tree.”

Murphy concluded: "The immediate challenge will be to use these methods to reduce the cost of achieving the goals of major, multi-site projects, such as The Cancer Genome Atlas, which aims to accelerate understanding of the molecular basis of cancer with genome analysis technologies."


Naik AW, Kangas JD, Sullivan DP, & Murphy RF. Active machine learning-driven experimentation to determine compound effects on protein patterns. eLife, 5, e10047. doi:10.7554/eLife.10047 (2016);

Go to the profile of Stella Bennett

Stella Bennett

Contributor, Future Science Group

No comments yet.