KCAIL Partners with Kansas State University in the Fight Against Malaria
Malaria kills. You might not know it but according to the World Health Organization half of the world’s population is at risk for Malaria and there were 445,000 deaths in 2016.
Research for the treatment and cure of Malaria is an active area of research and a large contribution to this is coming from the Open Source Malaria Project (OSM). OSM uses open source methods to achieve a bold vision: a cure for malaria.
Recently, KCAIL set off to join this fight with the best weapon we have in our arsenal: machines. We partnered with a research group at Kansas State University who was already involved in the pursuit of a cure for malaria and used artificial intelligence to suggest compounds to investigate further for potency against malaria. We started with a data-set of 47 chemical compounds that have already been tested against Malaria and 23 compounds that have not yet been tested. These weren’t large data sets, but because each test for potency is slow and expensive, there is incredible value to identifying potentially potent compounds. The decision of which compounds to test is not taken lightly!
Our project goals were to identify an untested compound that is expected to be potent and to identify which chemical descriptors describe potency. The first goal would help researchers to narrow down their tests among untested compounds while the second goal would help researchers modify tested compounds to increase their potency.
This area of research is called Quantitative Structure-Activity Relationship analysis or QSAR. We brought machine learning expertise to an entirely new area of research, but we could only be successful if we relied on the expertise of the malaria researchers. Our first step was to learn how they would use our outputs. This informed both what our predictive target would look like and how we would want to present the findings.
We also performed background research of our own to better understand the data that was given to us. We received over 1,400 columns of numeric chemical descriptors that initially made little sense to us. Before we were comfortable building predictive models, we turned to OSM’s extensive research and were able to read through other’s work to get ideas for how we wanted to tackle this problem.
One of our greatest assets during this process was playing our own devil’s advocate. If we could be more skeptical about our results than any stakeholder, then there would be fewer surprises when it came to presenting our findings. High levels of predictive accuracy can at times be obviously unrealistic. If you don’t have the best data, the largest data-set, or the best understanding of the problem at hand then you probably wouldn’t expect a high level of accuracy.
Our healthy skepticism led to a slightly different approach than what we started with and it improved our final outcomes. Any stakeholder would want more evidence before they would be willing to accept something as reliable. “Robust” became the name of the game. Instead of producing one “accurate” model, we decided to produce 80 different models. These models allowed us to develop confidence intervals that gave us more flexibility for any one model to be wrong.
We knew the predictions weren’t perfect and that wasn’t a problem. As George Box said, “All models are wrong but some are useful”. Our goal was to produce evidence for which untested compound was the most likely to be potent. Within our final report we described untested compounds that we believed to have potent properties and chemical descriptors that appear to have a relationship with potency against Malaria.
To see our findings and the code we produced, here is the GitHub repository for you to read, clone, and build from. We hope that you read through our technical report and find value in our research!
If you want to become a data superhero, send us a message:
