CSHL Automated Imaging & Phenotyping meeting Day 1

Going to try a live blog for day 1 to see how it goes.

Arrived earlier this afternoon.  So beautiful here on Long Island.  Took the train to Syosset, and then shuttle to CSHL.  Great, great campus.

First speaker is opening up with a description of what sorts of problems this nascent meeting will seek to address.  E.g How do you make images available publicly, and for publication?

First session chaired by Bob Robertson.  Five speakers tonight, including Erin Styles (yeah!).  First up though is Norbert Perrimon from Harvard to talk about HCS using RNAi in Drosophila

  • Lab is interested in signal transduction networks to identify players in network topology.  Pipeline uses RNAi screens, MS, RNAseq.  (Slides going too fast for me to keep up.)  Examples of the techniques used in his lab via some different projects.
  • One is investigating Rac and Rho networks.  RNAi experiments yield 145 phenotypic features (what types?).  Poor clustering due to correlations between features.  Used NN with z-scores to learn combinations of features that lead to more sensible clustering.
  • Second examples is RNAi screen in primary neuronal tissue.  Which RNAi treatments will show similar phenotypes?  Used a segmentation-based algorithm to extract axon / neurons, and then classified the degree of connectivity of the axons identified in the image (Q: are these features robust to translation / rotation?)  Same methodology repeated in muscle tissue.
  • Last example is nucleolus size screen, which is a good marker for cell size (they claim).  The sample images he presents are *great*.  They use a confocal stacking w 6 layers (Q: do they fix the cells?)  After max intensity projection, they determine cell area, nucleus area, nucleolus size.

Second is E. Styles.  I won’t blog this one, since I talk about the particulars in other posts.

Third talk by Arvind Rao from CMU (joint work with Rob Murphy!)

  • Talk focuses on mining immunological assays for human biomarker discovery.  They want to investigate finding BMs using both abundance and location changes between healthy & diseased cells.
  • Need two elements to find new BMs: Human Protein Atlas for prior info on differences between location of human proteins in healthy and diseased tissues. (Q: What sort of natural location variation do you see within both classes for a typical protein?)
  • Data quality issues: how to deal with them?  Use some spectral unmixing to decompose the images (Newberg & Murphy 2008).  Extract field level features as well as multiresolution texture features (moment calculations from the joint DF, haralick features).
  • Field level feature extraction: spatial proximity features from binarized images.  They use 3 types, but are focusing on Commute time distribution.  On a point set of protein points & DNA points connected by NN paths, they build a dist’n of mean commute time, where commute time between nodes A&B is the number of edges separating the points.  Should be very informative of spatial proximity of protein to DNA & vice versa.
  • Too expensive computationally, so they sample 100×100 windows in each image (Q: how does this vary as you choose relatively dense windows?)
  • This dude speaks my language.  I should look into working in this lab 🙂
  • A parallel approach is to train classifiers to identify the localization of proteins in 11 compartments (previously identified).
  • Q: in any image sample there must be a mixture of cancerous and normal cells.  Do you have any way of unmixing the two in an image? A: Sadly no, not without any more information (genotype, fine grained pathological info).

Break now, what a frenetic pace!

Next up, Ronald Eils.  Talking automating liquid handling & slide production to produce screens.  Biophysics & molecular biology combo talk

  • Want to ask questions about high-throughput studies for phenotyping of biomechanical properties.
  • BM important to cells with motility, adherence, integrity, etc.  Want to compress BM state of cells into a 4 tuple: stiffness, compressibility, elasticity, viscosity.
  • Conventional methods: microstretching, micropipetting, optical tweezer.  Extensive human-cell interaction required (1 person / 1 cell ratio!)
  • They have developed image & model based parameter estimation in a contactless (unconventional) framework of PDEs modeling the reaction of cells to the three descriptors and a method to estimate the unknown parameters in this framework.
  • Also in collaboration with another lab they have developed a laser based contactless method to process & estimate the parameters of 15 cells / min, a huge improvement in throughput.
  • Example of relationship between genotype & mechanical phenotype using drug treatment to perturb cells.  Measured ~ 5000 cells that pass QC, roughly 100x more than conventional method studies.

Running low on memory, may not be able to blog the last talk by Anne Carpenter in situ, but will try to reproduce from memory tomorrow.  Her talk is an overview of work to image quantifications in physiological contexts.

  • In general studies to screen cells with perturbations, any microscope, identify outliers from the WT population.
  • Cell based assays: try to kill bacteria in individual wells of multi-well plates
  • New idea: study bacteria in the presence of human cells, chemicals that can inhibit the infection process.  Looking for an anti-infective profile rather than simpel anti-biotic profile.  Nothing fancy in terms of image analysis, just separation from
  • Remainder is an overview of lots of different projects and collaborations, but the slides pass so quickly that I can’t keep up.  So much exciting work going on at the Broad institute.

Well that was day 1.  Really enjoyed the whole experience, and excited to see what day 2 brings.