This is a hackathon on multiview learning - part of the DAMVL workshop @ ECML 2019 - with a dataset from the field of developmental biology. Developmental biology is concerned with the study of how an embryo develops from a single fertilized cell into a complex and organized multicellular system. This process involves dynamics at multiple scales which are recorded using numerous acquisition techniques, from live movies using fluorescent reporter proteins to fixed samples in in situ hybridization and immunocytochemistry techniques. To study how cell fates are established by gene regulatory networks in Drosophila melanogaster embryogenesis, it has recently been proposed that a first necessary step is to integrate multiple views from heterogeneous image datasets.
We focus on the dorso-ventral patterning in Drosophila melanogaster early development. We provide a dataset that gather together multiple snapshots (images of 128x128 pixels in single channel) acquired during the development of few subjects. Nuclei is referred to as view 0, it corresponds to the morphology ; protein expression of doubly phosophorylated ERK (dpERK is view 1) ; Twist (view 2) and Dorsal (view 4) ; and mRNA expression of ind (view 3) and rho (view 5). In total, the dataset consists in 255 instances of drosophila embryo during its development, featured into 6 views. There are missing data such that none of the instances has the entire 6 views. Examples of data are shown below. Once completing missing views, combining all together results in a colored sequence showing the five molecular components (views 1 to 5) during embryo development:
One goal is to develop methods to predict such missing data, a first step in this direction is to predict one view from the others. We then formalize the given task as predicting view 1 given views 0, 2, 3, 4 and 5. This is a multiview regression task where one might learn a function f : R128x128x5 → R128x128, in a supervised way but with missing views. Read the Participate/Get Data section for more information about data and submission.
Details about the task and data can be found in Villoutreix et al., 2017
Or contact organizers:
We compute the average mean square error MSE metric over all predictions in the test set: we compare pixelwise predicted values with the groundtruth. The Leaderboard shows best submission per participant, the lower is the better.
Participants must submit results as a zip archive (without folder). Predictions must be as a numpy data file matching the same structure as target data described in the Participate/Get Data section.
Start: March 14, 2016, 11:59 p.m.
Description: Hackathon phase
You must be logged in to participate in competitions.Sign In