HRMAn – Host Response to Microbe Analysis
HRMAn enables you to automatically analyse parameters of host-pathogen interaction derived from immunofluorescent experiments. HRMAn is a custom-built open-source analysis solution based in KNIME. HRMAn uses artificial intelligence to assess host protein recruitement to pathogens.
All downloads and tutorials can be found here: www.hrman.org
Public Release: September 5th, 2018. Now on Biorxiv!
doi: https://doi.org/10.1101/408450An Artificial Intelligence Workflow for Defining Host-Pathogen Interactions.
An interview with Daniel and Artur
Why did you create HRMAn?
DF: Honestly, when I started work in the lab and heard of the painful analysis involved in studying host-pathogen interaction, meaning manually counting and quantifying infection processes while being bent over a microscope, I decided that I did not want to do that. So, I explored different approaches and teamed up with Mike Howell to first establish automated image acquisition. But then at one point I found myself drowning in imaging data that wasn’t analysed yet, so I started building the analysis pipeline, that now, after roughly one and a half years, has turned into HRMAn.
AY: In the past 5-8 years computer vision has made an amazing technological leap thanks to the advances in machine learning and AI. We wanted to place these advances at the fingertips of the biomedical researchers assisting their discovery in the field of infectious biology.
What do you like best about HRMAn?
DF: The speed and the amount of data he can process. It enables me to do experiments on a whole different scale than compared to classical manual counting. With every experiment I am able to analyse 100 times more pathogens and cells than I would be able to do manually. I’m basically only limited by the computing power available, which luckily is constantly improving! So, I can do almost unlimited amounts of conditions in my experiments, meaning as many cell types/ genotypes/ pathogens as I can think of, and I would still be able to analyse everything within a day or so.
AY: Two things actually. First is that HRMAn is user-friendly, since it operates fully through graphic interface. It consists of boxes representing modules, which are connected by strings depicting the flow of your experimental data. This is the simple basic idea, making the analysis accessible and intuitive. No “black magic”. Second, this very simplicity makes it easy for wetlab experts to adapt HRMAn to their needs and build upon it.
Why is the HRMAn recruitment analysis better than what I get out of a commercial software package that comes with the microscope? Why bother with the AI?
DF: Because commercial software does not incorporate artificial intelligence (yet). The analysis is just way more robust to inhomogeneous phenotypes as classical image analysis. And while it isn’t perfect yet, with every new model our accuracy increases! That I think is amazing, considering that my computer is basically teaching himself! Another great thing about this automated analysis is the lack of bias and the lack variability in between experiments. As I said it’s not perfect yet, but at least the tiny error introduced during analysis is the same across all my experiments, which makes them comparable. Probably the error is even smaller as compared to manual counting, but unfortunately, I cannot quantify this (meaning my own mistakes).
AY: Most commercial software is based on “black boxes”, i.e. companies need to protect their trade secrets. While understandable, this are bad news for science, which relies on full disclosure to aspire complete reproducibility. Not only is HRMAn open source, it also is built based on an open source platform. Furthermore, built for AI, HRMAn will allow biologists to take a full advantage of the recent advances in data analysis. Biology like no other field relies on the opinions of highly trained experts. So why bother with AI? Imagine that every time you are analyzing your data you have a consolidated opinion of world’s best biologists as a result. This is what we are trying to build with HRMAn.
Why is HRMAn based in KNIME and not some other open-source program?
DF: Initially, I tried a lot of different programs to set up something like HRMAn, e.g. FIJI or Matlab. But while they may offer a greater range of tools, for instance FIJI, would not have been able to combine image analysis and the following data analysis and statistics work as KNIME does it and they would have probably suffered from the amount of image data we are now able to process with KNIME. Moreover, the integrated user interface of KNIME allows even new users to easily adjust or extent HRMAn to their needs.
AY: KNIME is an open source platform originating from an academic, supported by a company and widely adopted by a large user community. It is leveraging its cross-platform applicability from a very famous java development environment – Eclipse. All this will ensure that HRMAn is running across a variety of hardware and has a long-term support.
What are HRMAn’s flaws and weaknesses at the moment?
DF: Well, it’s a work in progress, so of course we sometimes find general or user specific errors. However, most of the time we can easily address them and fix it. Like this HRMAn is constantly being updated and improved.
AY: I see the main weakness of HRMAn in the fact that the AI field is progressing at astonishing pace. Better algorithms emerge every few months. However, that was the main design decision behind building it as an open workflow. This way HRMAn can progress hand in hand with AI.
How easy is it for a lab to install and use HRMAn? What are the biggest stumbling stones?
DF: Super easy! You just install KNIME and load the workflow. We even have a tutorial video on how to do that. The biggest stumbling stone may be the computer itself that different labs use, e.g. having different operating systems. But KNIME already takes a lot of trouble away there and so far, HRMAn is running on Windows, MAC and Linux without any bigger issues that I know of.
AY: The basic level of the HRMAn analysis is a click-through: get images, tweak some settings and run it. If the lab is working on a specific pathogen we didn’t look at yet, AI part would require more work in the sense of re-training the neural network. However, the beauty of the methodology we have used lays in the fact that one could use our models as a training primer (so called pre-trained model), which is far less work than starting from scratch. This way adaptation of the AI model becomes easier.
What’s next for HRMAn? What is its greatest unused potential?
DF: Next on the agenda is definitely 3D and time-resolved analysis, features we haven’t used so far because we didn’t need it in the beginning, but that we are already developing! Also, we’ve been trying to engage citizen scientists for image annotation for a while now, but never managed to do so because other things were more important to address. I think this would be a great thing to do in the near future. The greatest unused potential at the moment is that we don’t look a lot at the host cell yet. Most of the analysis is focus on the pathogens. We’ve just started to explore this, but there are hundreds of additional parameters that we could extract from the images and maybe uncover even more interesting phenotypes that have so far been hidden to the eye of researchers!
AY: Any research software exists through the user community it creates. In this sense it outlives and transcends its original limitations. I see HRMAn as call-out to the community: “let’s talk bio-AI, let’s accelerate research!” With that in mind, the greatest unused potential of HRMAn is the vast community adaptation and a creative network of bio-AI models shared by the community. This is what we would like to build here, this why the only way is open source.
Thank you both for your time!