CSHL Single Cell Analysis 2017 - Bioinformatics

July 13, 2017    teaching bioinformatics reflection python jupyter scikit-learn software carpentry

Reflections on teaching math to biologists

Table of Contents

The Cold Spring Harbor Laboratory (CSHL) Single Cell Analysis Course (not to be confused with the “Single Cell Analyses” Meeting) for 2017 just finished a few weeks ago. I was a co-Instructor and was in charge of the bioinformatics curriculum. In some ways, teaching at CSHL is a perfect teaching setting because there are no distractions - the students are there exactly to just work on the course. However, many of the students had minimal programming and command line exposure so while they can completely focus on learning, they also have a lot to catch up on.

This post goes over some lessons learned from this experience. If you want all the teaching materials, I made a github repo containing all of the notebooks we used in the course.

Expectation vs Reality

As always, what you planned isn’t what actually happened.

What I planned to do in the course

To help prepare students for any possible single cell question they run into, I wanted to expose them to several papers’ different methods of analysis. I wanted paoers covering four common questions in single cells

  1. Dissociate a tissue –> ??? –> Celltypes!
  2. Treatment vs control, in a singlecell context
  3. Molecular transformations over time (“pseudotime”)
  4. Perturbations + single-cell RNA-seq

Therefore in an ideal setting, we would have done: - Students read 4 papers ahead of the course - We recreate one figure from each of the papers - We’d cover all possible questions that could be answered using single-cell RNA-seq

What actually happened

  • Students read 4 papers
  • We recreated one figure from one paper
  • We spent a lot of time talking about Python basics, much more than I had anticipated
  • We didn’t get to answering every possible single-cell question
  • We spent far less time on the “interesting” part of bioinformatics - interpretation of results

I often let “perfect be the enemy of good” and would not fully reuse a lesson from last year because it was from last year’s papers, even if the concepts or programming commands were the same.

The way I designed the notebooks was there were two notebooks for each concept: 1. Explain a mathematical concept using images, animations, and interactive Jupyter widgets. 1. Assessment: Class discussion, calling on random_student() to explain the question 2. Show how the plot from the previous figure was created with complete step-by-step breakdowns of each function and parameter 1. Assessment: the students have to apply the function in a slightly new way or read documentation and add a new parameter

Bugs I ran into

  • Jake VDP’s Scikit-learn tutorial, while excellent, has not been updated to the most recent version of Jupyter/IPython widgets. Specifically, the syntax for specifying

Lessons learned: What I would do differently next time

  • Start with simple dataframe manipulation such as with a metadata file
    • show plotting and groupby
  • Focus on one paper, and one analysis
    • Take an existing paper and recreate its figures, but also show how to do the analysis the “right” way, if it differs from what they did in the paper
    • Start from the end - what commands and concepts do they need to know by the end of the analysis? How can we build towards that?
  • Using a single dataset and analyzing it all the way through would help the students get deeper into a single problem, rather than exposing them to a lot of random stuff
  • Don’t expect that people have any experience when we start. Start out the course with an overview of Python and programming

One of the hardest things about designing a curriculum like this is finding a good dataset to showcase all of these analyses on. It’s hard to find a single paper that covers all possible questions, but maybe that’s alright. I always have to remind myself that “good enough” but shipped is still better than perfect and unrelease.

The good thing is that these mistakes made me write out my learning goals for the different sections of the course. I remember now from Software Carpentry teacher training ( Lessons and Objectives) that I should have done this in the beginning, but hey, you live and you learn.

comments powered by Disqus