Thanks Dr. Albert! The dataset we’re working with is still highly sensitive since it involves a new drug in clinical trials, so we unfortunately could only talk about our method and not specific results. We are working on analyzing other less sensitive datasets now in preparation for a manuscript, though. We wish we could have talked more about what we found too! Soon, hopefully. Thanks again for your comment.

Hello Dr. Shin. Thank you very much for your question. First, the presented method can be used for experiments on any time scale. Second, in the specific data set analyzed here the time points 2h, 6h, 12h, 24h and 48h were determined based on prior experiments and chosen to capture both the early and late responses of the cell lines to the drug. Also, the drug under investigation interferes with the mechanisms involved in cell cycle (cell division), a process that generally occurs on average every 24 hours. The progression of time points is chosen such that short term processes are sampled on a more fine grained scale than long term responses. While technically it is possible to measure at many more time points, the number of time points is limited by the the cost of each microarray (on the order of hundreds of dollars) used to measure gene expression. In this data, 150 microarrays were used.

Dr Buneo you are correct, there are already software packages available to analyze time series data (e.g. BETR, EDGE, SAMR, STEM, LIMMA). Each tool approaches the problem differently in how it formulates the time series statistical model and how it makes the results available to the user. Early in the project, we explored many of these tools and found them either to be not functional, difficult to use, or not exactly what we were looking for. We focused on making our tool as general purpose as possible without making it difficult to use, and tried to make exploring the results intuitive to non-technical users as well as those comfortable in computational environments. Our statistical model is not particularly novel, but we’re unaware of any other software that visualizes time series data in this way so as a whole package we feel the tool might be valuable to other investigators.

Thank you for your comments, you make very good points.

Addressing your question about direct vs feedback responses, we can’t know for sure with our data, though we do know some specifics about the drug’s biochemical targets. It is true that this analysis method, and microarray data in general, cannot easily distinguish between direct and indirect responses to external stimuli. One might speculate that changes at the earliest time points after treatment are more likely to be direct responses, while later differential expression corresponds to reactive processes, but there would need to be experimental validation to confirm this. Indeed, one aim of the analysis is to identify testable hypotheses that can then be subject to targeted studies later. However, direct or indirect, the changes we observe in gene expression are all effects of the drug treatment in human cells by experimental design, which is what we seek to better understand at the genome-wide level. Simply understanding which genes change may have direct translational implications even without a detailed characterization of specific mechanisms.

We agree that it is unclear how good expression is as a proxy for protein levels, though validation experiments often confirm protein activity predicted from gene expression analysis. Including other phenotypic data (e.g. viability), or high-throughput proteomic data in our analysis could lead to greater biological insight, but unfortunately those data are not available for this study.

The purpose of our tool is to help biologists quickly generate specific hypotheses that can then be tested in the lab to directly investigate real biochemical relationships. The overall project has both of the goals you mention: develop a tool to better understand time course data and then apply it to our specific time series dataset. These two goals are synergistic in the sense that seeing how our tool captures (or misses) meaningful biology in our dataset will inform iterative development of the tool. Due to our data being highly sensitive, we can’t say much about what we’ve learned about DLBCL from this analysis yet, unfortunately. Watch for a publication within the next year. :)

Excellent question, Dr. Pfromm. One translational application of our findings might be to better understand the reasons underlying the heterogeneity of DLBCL treatment efficacy. For example, if we discover that the drug hits a particular pathway in one of our cell lines and not the others, we can look for differences between those cell lines to potentially develop tests that predict whether a patient will respond to the drug. The treatment targets aspects of cell cycle, which is involved in many diseases besides cancer, so better understanding what happens when cell cycle processes are perturbed could be valuable not only in other clinical applications but to basic biology as well.

Thanks Stephanie! Yes, we are currently working on an R package that performs the analysis and produces the visualization. We hope to get it out sometime this summer.

That would be great! My lab currently uses R for our microarray normalization. If you could remember, please let me know when it comes out, I’m very interested in it.

I’ll put a reminder in the TODO file of our software repository. :)

Thanks, glad you liked it.