Q&A with Dewpoint Therapeutics
AMA style interview with VP, Data Science and Engineering, Jesse Johnson
This is reposted from a Q&A we did in March 2022. Check out our Slack for more.
Bits in Bio interviewed Jesse Johnson, VP of Data Science and Data Engineering, from Dewpoint Therapeutics. We discuss facilitating communication between cross-funcitonal teams, , and much more below!
Nicholas: Welcome to our q&a with @jesse! please feel free to post questions in this channel — let’s try to group related ideas and follow-up questions in threads to make it easier to follow along. i’ll get us started with a few questions, but please join in asynchronously.
Question #1:
Vega: What are some struggles with introducing data science as a concept in a biotech company?
Jesse: It's hard to know where to start answering that, but i think a lot of the friction has to do with different mental models that wet lab scientists tend to employ compared to data folks.
Question #2:
Yohann: Most biotech face the same issues, again and again, despite scientific nuances. what's the barrier to a successful data environment?
Jesse: I think a lot of the people who know what a successful data environment looks like tend to focus on technical solutions, when there are a lot of organizational changes, and shifts in mindset that need to happen to create the conditions for the data environment to me successful.
Nicholas: What are some examples of things you've put in place at dewpoint to build this type of successful data environment?
Jesse: One thing that i think has been very successful is defining a cross-functional team to define how data and metadata flow from the lab to the cloud. there were a number of people who had very good ideas about how this could be improved, but no one felt empowered to make a decision. by explicitly giving the right individuals the responsibility, we got to decisions fairly quickly.
Question #3:
Nicholas: Jesse, i want to start out by asking about your role at dewpoint. can you tell us a little bit more about what you all do and specifically what your role as head of data entails?
Jesse: Dewpoint is a small research-stage drug development biotech using recent (and ongoing) research about biomolecular condensates, sometimes called membraneless organelles - these little balls, formed by phase separation, that cells use to control processes at a faster time scale than the dna -rna -protein cycle can manage.
I'm the head of the data science and data engineering teams, so i spend a lot of time debugging communication issues between the digital and wet lab teams.
Nicholas: Are you designing condensates? designing things that modulate condensates? things that interact with them?
Jesse: We're designing (or discovering) things that modulate condensates, currently mostly small molecules.
Question #4:
Nitan: What are the common deficiencies you see in biotech founders coming from (phd and further) academia?
Jesse: I don't have a lot of experience working with different founders coming straight from academia, but in general, i've seen folks coming from academia struggle with differences in what kinds of management styles are effective in the different contexts. in particular, the expectations that employees coming from the tech side have for how leadership works is very different from expectation on the wet lab side, which is much closer to academia.
Question #5:
Nicholas: You talk a lot about mental models in your blog. can you explain why you think they’re so important?
Jesse: So, my background in math makes me inclined to think about abstractions, which is closely related to mental models. but basically, in the various places i've worked in biotech i noticed that a lot of people on different teams seemed to be spending a lot of time talking to each other without actually communicating. as i dug deeper i started to think that they were often using the same words to mean different things or vice versa, and my hypothesis is that this usually boils down to different mental models. once i figured out the right terms to google, i stumbled on the field of "shared mental models" which studies exactly this phenomenon.
Nicholas: I like this a lot — i feel like i’m always talking to engineers in this field about how there aren’t good abstractions. how do we change this?
Jesse: This is a great question, and one that i don't have a great answer to yet. but i think the place to start is to recognize that there are these differences in mental models, and how those differences can get in the way of communication. once you realize that, i think there are lots of ways to reduce the friction depending on your particular leadership style.
Question #6:
Nicholas: How do you think about using industry specific tools vs general tools? you might phrase it as: is it better to adapt a tool to your mental model or to adapt your mental model to the tool?
Jesse: Great question. i think it's best to adapt your (shared) mental model to the problem at hand, then build or choose tools that support and reinforce that mental model.
Question #7:
Nicholas: What does your data team do? are you mostly working on pipelines? visualizations? ml?
Jesse: A bit of all that. right now a lot of our time is focused on screening, so defining phenotypes in high content imaging that correspond to the underlying physiological change that we want to produce, building a model that can identify them, then running the model to pick compounds large-scale screens.
Question #8:
Vega: Do you use a lims platform to collect and centralize data management? if yes, is it built in-house? and what are some challenges to getting wet lab data into a digital format?
Jesse: Right now, we have a number of internally built tools that perform some of the functions of a lims, but there's also an internal discussion about adopting an actual, commercial lims. i think in the long run, there's no point in reinventing the wheel.
I think the challenge isn't in getting the data - it's the metadata that's a pain. the instruments collect the data, and they usually do it pretty reliably. but it's the wet lab scientists who collect the metadata - the biological context info that you need to analyze the data. and without either solid sops or really good software, they're probably not going to do it in a consistent way.
Question #9:
Yohann: Why do you think companies can do in their org structure to make your (our) life easier and data work more successful?
Jesse: I actually think the org structure is less important than a lot of the less tangible things like norms for how people communicate across teams, how processes are introduced and changed, and expectations for how people will work.
Every org structure is going to end up looking like a tree, so if you rely on that for communication you're going to end up with bottlenecks. for people to actually communicate and coordinate, you have to give them tools to do so outside the org structure.
Question #10:
Nicholas: How did you get into writing a blog? what’s your writing process look like?
Jesse: This is actually the third blog i've written. i started out writing a blog about my math research in 2007 when i was a postdoc working on three-dimensional topology. i then switched to a blog about the topology/geometry of machine learning when i started getting into that. so to some extent it's just habit.
But i find it's really useful for forcing me to collect my thoughts, and reflect on what i'm doing. in the case of scaling biotech, it really helped me shift from thinking about designing software to thinking about designing organizations.
Noah: Thanks for writing the blog @jesse! i've found it very interesting and your ideas around mental models have really got me thinking :slightly_smiling_face:
Jesse: Thanks, @noah!
Question #11:
Nicholas: Where do you see the most fruitful applications of ml in bio?
Jesse: It's hard for me to answer that question because there's often a big difference between what a biotech puts on their public web page and what they're actually doing internally. so it's hard to compare without being inside dozens of companies. but based on the two companies i've been in: building models that use single-cell ngs data to understand the dynamics of cell states, and building models that identify phenotypes in high-content imaging are both have a ton of potential, and we've probably only scratched the surface so far.
Jesse: And i'll also take this opportunity to plug my latest post, which explores why finding these fruitful applications can be so difficult: https://scalingbiotech.com/2022/03/23/ml-vs-wet-lab-the-great-impedance-mismatch/
Question #12:
Yohann: What have you seen recently that got you excited to apply at work?
Jesse: So, i recently read a paper that showed that postmortems/retrospectives can be incredibly effective at creating shared mental models if they 1) ask teams to reflect on how closely they followed specific principles tied to the target model and 2) put equal emphasis on how the team followed the principles, and failed to follow the principles.
Jesse: I've never been good about doing retrospectives, but this has given me some motivation to try to do more of them.
Isaac: Could you share this paper with us?
Jesse: Sure! it's this paper: https://www.researchgate.net/publication/247720345_guided_team_self-correction_impacts_on_team_mental_models_processes_and_effectiveness
Isaac: I agree with you that a lot of problems in tech and biotech companies are related to communication and agreements between teams, departments and business. a lot of ritual/procedures seems simple at first look, but are exactly to ensure a more horizontal communication of goals, tasks and expectations. It's interesting how easy is for people to take this rituals as useless and expendables. Postmortems and retrospectives are those kind of rituals that people usually take for granted. Guided team self-correction: impacts on team mental models, processes, and effectiveness. I've definitely taken postmortems for granted - or rather, i always intend to write them, but usually get distracted with other things instead. Thanks for the paper for your q&a. i really enjoyed your work and your blog. i will follow more closely after that. also, is very good to have a searchable word like shared mental model to deep dive more about this problems inside organizations.
Jesse: Thanks!
Question #13
Boris: @jesse thank you for doing this ama! i don’t have a specific question but need to say hi because i study nuclear bodies using proteomics for my phd work, and it’s rare to find people who have heard about paraspeckles, cajal bodies, stress granules, etc :joy:
in terms of a question, i’m wondering if you could comment on your experience trying to support very cutting-edge research on condensates. i understand you’re trying to find small molecules that modulate condensates, but would you say you do work in a “blinded” way? or are the wet lab folks/biologists giving you specific features to look for in certain conditions (cell type, tagged protein, class of molecule, etc?)
Jesse: Hi, boris. we try to have as much back-and-fourth as possible between the digital and wet lab folks so we can understand the questions they need us to address, and they can understand how the digital tools might help answer questions they hadn't thought of. in some cases, they're suggesting specific features to look for but we try to make it as collaborative as possible.
Nicholas: Thanks @jesse for your thoughtful answers! looking forward to more blog posts
Jesse: Thanks for organizing it! this was fun!
Brandon: Great job @jesse! your blogposts are galvanizing the field!
Jesse: Thanks!