Robotics and technology started to play a huge role in domains like medicine or the pharmaceutical industry, as they are able to complete difficult tasks faster and cheaper. With numerous diseases being discovered or spread throughout the world, the need for efficient treatments sometimes surpasses our capability of finding the best solution fast enough, leading to countless hard to treat disorders and a huge demand for better drugs.
To be able to reduce the cost and time of the production of highly potent drugs, mechanical systems with the capacity to learn and adapt are needed. Any drug testing procedure can take up to several years because thousands of experiments have to be done to analyze all the possible effects on any type of genetic variation. These countless experiments need a lot of funding and manpower and researchers from Carnegie Mellon University have developed a learning-driven experimentation active machine that is capable of reducing the cost of drug testing.
As one of the lead author Armaghan Naik has said: “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?” Hence, the need to focus on mechanical machines with the ability to predict the results of similar experiments based on small amount of tests that have already been made.
Under the lead of Professor Robert F. Murphy, head of the Computational Biology Department, a team of researchers was able to create an “active learning” machine, as they call it, capable of choosing the right experiments to do in order to assess a pattern. The robotically driven system developed uses liquid handling robots and automated microscopes and was capable of reducing the number of necessary experiments up to 70%.
The active learning machine was given the task to check the interactions between 96 drugs variants and 96 cell clones each tagged with a different fluorescent protein, meaning a total of 9,216 possible experiments. The “learning” algorithm of the robot had to learn how all proteins get affected in each case, without actually doing all the experiments. The machine started by completing 96 experiments in which he tested one variant of the drug on all 96 cell clone. From this step, the robot assessed where the protein is located in the cell.
All results were then passed through a quality control to analyze phenotypes that might have been related to a previously demonstrated drug effect. Thus, the machine was able to identify new potential phenotypes as well as similar responses, which helped the robot identify a group of proteins that respond similar to drug effects, helping to guess future responses. The active learning robot went on for 30 more rounds and discovered all the patterns needed, giving a 92% accurate response after conducting 2,697 out of 9,216 experiments needed. With only 30% of the experiments done, the active learning machine managed to give an accurate response on how drugs affect various proteins. This process didn’t only reduce the cost of the testing, but also the time in which the calculations have been made.
This research is a step forward for the pharmaceutical industry as it helps develop better drugs, with fewer effects in less time, filling faster the need of proper treatments. Also, it can have a huge implication on multi-site projects like the Cancer Genome Atlas, which tries to understand the molecular basis of cancer development by analyzing possible changes in the genome.