Effectively Communicating with Computational Folks — #CBSymp16
I had the chance to hear a number of thoughtful perspectives broadly on the field of computational biology when I attended the @PLOSCompBiol symposium #CBSymp16, held last month at the @NIH. One of the highlights of the symposium was the set of talks by experimental/clinical researchers about challenges in their fields that could be addressed by computational approaches.
Here, I want to use one such ‘external’ talk* — the one by cancer research pioneer Bert Vogelstein — to briefly reflect on how to effectively communicate with computational biologists. Vogelstein’s presentation on “The War Against Cancers and the Critical Role of Computational Biology” captured my attention not just because of its fantastic scientific content, but also because of how well it was crafted to simultaneously solve two conundrums computational folks constantly face when trying to forge a collaboration with experimentalists: 1) framing a computable problem, and 2) framing a generalizable problem.
Framing a computable problem
Computing can complement experimental work in a powerful way. However, the first difficulty is in teasing out the parts of an experimental enquiry can actually be framed as clear quantitative problems that can be approached using computational methods. Figuring this out is an essential early requisite, achieved by patient conversations between the experimental and computational collaborators to gauge both expectations and practicalities. A productive conversation can begin with a proposition of the following kind: “Here’s data that we generated to ask this specific question; here’s what we want to find out, and here’s what we can test experimentally if you can provide candidates/predictions.”† Starting here, it is vastly helpful to the computational biologist to work out what is useful, measurable, and testable by the experimentalist (e.g., individual genes, interactions, activity levels, clusters, mechanism, pathways, sample metadata, outliers). Defining these parameters aid in making the problem concrete and goes a long way in framing the right problem.
Framing a generalizable problem
To understand this second aspect, it helps to appreciate that, on the outset, computational scientists often significantly differ from experimental/clinical researchers in how they think about problem-solving. When presented with a specific biological problem, it is appealing for computational biologists to take a step back and see if that problem is generalizable and if the methods/approaches to be developed can be general-purpose and widely-usable (in addition to being applicable to the specific problem at hand). Seasoned computational scientists are able to spot such opportunities for new general quantitative approaches from a stated specific challenge. Young trainees, on the other hand, reap huge benefits from interactions with domain-experts who can explicitly state gaps in their knowledge and practice, guide in scoping out the general problem, and inform if that general problem is regarded in the community as relevant and needing a solution.
Adequately resolving these two quandaries is a vital to the computational scientist to build the conceptual scheme (s)he needs to dive into the problem. In his talk on the war on cancer, Vogelstein struck both these chords perfectly. He presented four well-formed high-level questions: 1) “How many driver genes?”, 2) “How many drivers required?”, 3) “What causes these mutations?”, and 4) “How heterogeneous are they?” Then, within each question, he weaved in historical context, recent experimental/findings findings, and limits of current knowledge. Additionally, in each case, he explicitly stated specific open questions such as “Which among the observed mutations in a tumor are drivers and which are passengers?”, or “Can we predict the antigenicity of mutations observed in tumors?”. But, what about existing computational methods to address these questions? In a not-so-often-seen discussion from the perspective of an expert-user, he clearly pointed out how current methods perform, what is still lacking, and where the opportunities are for newer/better methods.
“You don’t really understand something unless you can mathematically model it” [J̶o̶h̶n̶ ̶v̶o̶n̶ ̶N̶e̶u̶m̶a̶n̶n̶ Bert Vogelstein]
Beneficial talks by experimental researchers need not just be one that is expository. I remember feeling similarly thrilled to listen to Ellen Rothenberg on “Genome-wide Transcriptional Machinery of a Cell-fate Choice: The Early T-cell Pathway” at a RECOMB/ISCB Regulatory Genomics meeting in 2014. After a bunch of talks about methods for computationally inferring molecular regulatory networks, Rothenberg — a leading experimental biologist — concluded the meeting with a very ‘computational-biologist-friendly’ talk that painted a beautiful, realistic, and dazzlingly complex picture of transcriptional regulation in T-cells, implicitly showcasing how unsatisfactory current methods are in comparison to ground truth and unveiling the higher standard future methods should aim for.
Ergo, in more ways than one, there is plenty of value to be reaped in having experimental/clinical investigators address and interact with computational scientists — esp. budding researchers — in symposia and conferences, and my hope is that this practice becomes a permanent feature of all future meetings.
Footnotes
\* Ginger Hunter wrote about another excellent ‘external’ talk at the symposium by Jennifer Lippincott-Schwartz on super-resolution microscopy in cell biology.
\† The range of propositions brought to a computational/informatics person go from (the rare) lets-start-at-the-drawing-board-to-co-design-experiments to (the surprisingly-frequent) can-you-analyze-and-make-sense-of-my-samples/data! Through the years, I have also encountered different interesting questions of the kind “I have an experimental finding. Is there ‘computational’ support for this finding?” Here, ‘computational support’ means external data or predictive methods corroborating the ‘pattern’ observed in the experiment. Conversations beginning here have lead to many nice ideas and useful follow-ups that have eventually blossomed into fruitful collaborations.
Arjun Krishnan is currently a senior researcher at Princeton University and soon to be an Asst. Professor at Michigan State University
www.arjun-krishnan.net | @compbiologist
Any views expressed in the blog post are his own and not necessarily those of PLOS.
