Researchers develop automated cell-screening system
With the speed and accuracy that they provide, machines are rapidly replacing humans in almost all fields today. Humans, though slow and prone to errors, still hold the supreme position in jobs that require thinking.
The field of life sciences is one such field in which tasks like analyzing microscopic images have been done by humans and not machines because of the tasks’ complexity. However, this scenario is about to change.
Researchers at Carnegie Mellon University’s Ray and Stephanie Lane Center for Computational Biology have developed an automated system that can analyze images of cells much faster and far more accurately than the human eye.
The technique, developed by Geoffrey Gordon, a robotics and machine learning professor; Robert Murphy, a computational biology professor; and Shann-Ching Chen, a recent Ph.D. graduate from Carnegie Mellon, was published in the Journal of Machine Learning Research.
As stated in a Carnegie Mellon press release, the technique will be capable of analyzing more than 100,000 cells and will also be able to establish relationships between those cells.
Automation has been used in the field of biology for some time now. Machines are being used today to screen biological samples. High-Throughput Screening, or HTS, is a well-known technique that relies on data processing and robotics and is capable of conducting simple tests on biological samples.
Tests, like counting cells that express a particular protein, can be done easily using HTS.
However, machines employing HTS are still not able to beat humans at understanding complicated patterns between cells in an image.
This is why the newly developed technique is special.
According to Gordon, “Whenever there is a human in the loop slowing things down, it’s a great opportunity to speed things up by introducing machine learning.”
Particularly, the group was interested in classifying cells and figuring out the relationships between neighboring cells in an image.
Gordon explained that it is much easier to classify cells by studying a group of cells rather than just a single cell. According to Gordon, the major drawback of analyzing a group of cells is that it requires the computer to take into account a very large amount of information.
Manipulating such large amounts of information is very difficult and also time-consuming.
“What we did is that we came up with a new reasoning methodology that is potentially far faster at doing the reasoning of entire groups of cells,” Gordon said.
Murphy explained the overall working of the system, saying, “The system describes the distribution of a given protein in each cell using numerical features, and then learns which features are associated with which subcellular pattern — the same way you might use measures of color, smoothness, and sphericity to distinguish different fruits.” Thus, the system relies on protein distribution in the cells to classify them into different groups.
Initially, the data regarding protein distribution is gathered. Based on this information, the probability of a particular cell belonging to a specific group can be estimated. The information is then put into a factor graph.
Gordon explained that factor graphs are essentially used to encode probabilities.
After the factor graph is created, an algorithm called the “belief propagation algorithm” is used to analyze the data. The belief propagation algorithm forms a key element of the system.
“The belief propagation algorithm is a general machine learning algorithm to make inferences on each node in an interconnected network,” Chen said. “In our case, each cell is represented as a node.”
Chen explained that the belief propagation algorithm passes the information obtained from each cell to its neighboring cell and compares the information between the cells.
Thus, the system is capable of classifying groups of cells and understanding peculiar interactions between the cells.
The system is also capable of recognizing subtle differences between cells. “Our system has recognized subcellular patterns of two Golgi proteins, Giantin and GPP130, with very high accuracies,” Chen said.
The system will bring about a remarkable change in the way biological research is done today. Biologists today still analyze images manually. This is not only time-consuming, but also compromises on accuracy.
According to Chen, the system has the capability of revolutionizing the way research is done. “Our system, [which] can produce objective, quantitative, and repeatable analysis, can really push this field of research forward. “The way research is done now will not change very fast, but eventually it will.”