This is reposted from a Q&A we did in June 2023. Check out our Slack for more.
We had the privelege to chat with co-founder & CEO, Wyatt McDonnell, from the new biotech startup Infinimmune! Infinimmune is advancing whole human antibodies for human disease.
Read more to learn about:
Antibody drug discovery
Software infrastructure considerings for early startups
Journey of the founders
and more!
Nicholas: Welcome to @wyatt! very excited to have you here to talk about infinimmune and antibody drug discovery. i’ll get us started with a few questions and hopefully the community will jump in as well!
for those of you for whom this is your first q&a — please feel free to ask questions. good topics here are: starting a biotech, single cell sequencing, antibody discovery, computational biology, and more!
Question 1:
Nicholas: Let’s start with a bit of background. how did you get interested in science and the application of software to the research domain?
Wyatt: I grew up in a pretty rural area of michigan with a high-quality library and great public schools -- early exposure to some cool chemistry redox reactions, very basic microscopy, and a lot of ecology were all things i loved as a kid. my folks have an early drawing of mine that shows viruses in wasps (which maybe was why i considered rotating in a wolbachia lab later in life).
Wyatt: The software component was more from utility in grad school -- a lot of people i went to grad school with would've benefited heavily from being able to program a little, and the labs i was mentored in both had software experts who were super helpful and generally brilliant! it was hard to not want to build and apply software to biology after seeing how much it could speed up research and, more importantly, force you to codify a problem.
Question 2:
Nicholas: You spent some time at 10x before starting infinimmune, what do you think their secret sauce was and why have they been so successful?
Wyatt: I joined 10x in 2019, just before their ipo, and i think you can largely read about their "secret sauce" and distill it down to a few things:
• focus on hiring and retaining very brilliant people who do (rather than manage)
• focus on executing extremely difficult goals (3 months vs. 12 months for something you think would normally take 36 months)
• focus on building tools that are push-button-go for both hardware and software
the fact that 10x's tools and pipelines largely are apple-like and "just work" is a thing that they've done since day 1, it's really hard to do, and it's also hard to find people who share that vision. when i joined, 10x was a global talent hub for a lot of smart people in interdisciplinary science, which was also very empowering.
Wyatt: Oh, another thing: 10x built an organization that's very good at relationship building between wet lab and dry lab. it's extremely difficult to get hired into either of those "divisions" if you can't show that you know how to ask good questions and be a good team member "from the other side of the aisle", which is so important and underappreciated.
Question 3:
Nicholas: Tell us a bit about infinimmune – what was the motivation behind it and what is the problem you’re solving?
Wyatt: Infinimmune is predicated on a few very simple beliefs shared by our founding team:
• the tools to understand the immune repertoire are underutilized and underadopted
• we don't actually know the intricacies and nuances of the human adaptive immune repertoire
• the best model organism of a human is a human
• the best targets for antibody drugs haven't been discovered yet
we're using the human immune system as a starting point for discovery for all of the above reasons. our team is extremely experienced when it comes to tool building, and we're focused on building hyperscalable solutions to two problems:
1. finding antibodies in humans that you can't find elsewhere and which can't be found in public data (there are severe biases!)
2. matching antibodies to their targets using antigen-specific (target known) and deorphaning (what's the human proteome target of my antibody) approaches.
Nicholas: I want to follow up on `the best model organism of a human is a human` — this is likely true, but often comes with scalability issues (hence why many tests are done in vitro or in silico). how do you think about this issue?
Wyatt: If you have a better model organism of a human than a human we'd love to know what it is!
we think about this in a few ways:
• the existing in silico data for antibodies are totally inadequate for anything except maybe under extremely limited circumstances predicting binding to a limited extent.
• if you want a molecule that's been "preclinically tested" before it goes into a person, then a molecule coming out of a person is the only sensible starting point when it comes to safety.
• a lot of properties that are a pain to deal with from an antibody drug development setting (solubility, affinity, off-target specificity) are largely presolved when it comes to antibodies developed by human b cells in your body.
• most of the in silico and in vitro tests for antibodies are of partial but not complete utility. there aren't a ton of generalizable and extremely well-proven tests in either category here.
Nicholas: It makes sense that human antibodies have a lot of desirable properties for a drug you would eventually put into humans. what are the existing challenges with this approach/why hasn’t this been done before?
my understanding is that people try to get around this with humanized mice (would love to hear as well why that’s not sufficient)
Wyatt: The main challenge is that people think this can't be done! reasons this has been considered undo-able historically include:
• "well what are you going to do, patent a natural antibody?" --this is a solved problem, and several natural antibodies are already being used to treat humans very effectively
• "it just seems hard" --this is the most common one
• "i can't immunize a human" --maybe your target isn't as good as you think it is!
and humanized mice are of extremely narrow genetic diversity when it comes to making antibodies. there are human antibodies described by others and which we see routinely that can't possibly be generated by a humanized mouse/stuart little unless the entire mouse genome has been humanized. not impossible, but not remotely feasible or reasonable otherwise.
Question 4:
Nicholas: We’re seeing more and more drug discovery companies take advantage of sequencing as part of their core platform. how do you think this changes the way a discovery company should be built?
Wyatt: I think it depends a lot on what you're planning to do with the sequencer. a sequencer can be used for a lot of things, but if you aren't intimately familiar with how that happens from many angles (basic flow cell chemistry which is platform-specific, imaging/detection steps, what types of molecules can effectively be detected and how), probably you don't want to try and build a company around hacking a sequencer.
Wyatt: If you're doing library-based discovery, a sequencer can be hugely empowering, and having it in-house (which is what we do at infinimmune) can accelerate your r&d timeline substantially because you're the highest priority customer within your own company.
Wyatt: Like dna synthesis, costs still need to come down -- but illumina has healthy competition now from several companies on several applications :slightly_smiling_face:
Question 5:
Nicholas: Can you walk us through a typical discovery workflow for you? at a high level, i’m imagining something like:
```design library -sc sequencing -analysis```
but would love to dive into more specifics about where you’re using software to help accelerate this process
Wyatt: A more accurate discovery workflow for us is something like:
```select human samples --sc sequencing --sequence analysis and candidate selection```
can't say a ton about specifics on the last piece, which is where the software comes into play, but there are a ton of general opportunities for software to be useful here including expression optimization, prediction of various drug properties, selection of candidates which are optimal in other "non-assayable" ways, etc.
Wyatt: Library design is great for some things, but like most human theories of "intelligent design" makes a lot of assumptions that are just wrong and can produce not-so-ideal molecules out the other end.
Nicholas: In a previous thread you said:
```most of the in silico and in vitro tests for antibodies are of partial but not complete utility. there aren't a ton of generalizable and extremely well-proven tests in either category here.```
but here you’re talking about using software for prediction of drug properties, candidate selection, etc.
is this the “partial” utility you’re talking about? i.e. we can’t design antibodies de novo yet, but we can use models to help filter them?
Wyatt: Yeah, i think that's right. i think we can "design" sequences which bind, maybe, but we're going to design a lot of things that aren't safe (immunogenicity) or just aren't very good drugs (solubility, formulating issues, etc.).
Question 6:
Nicholas: Without giving away your secret sauce, can you discuss your software infrastructure at all?
i imagine you’re generating quite a large amount of data — what tools and techniques are you using to manage that?
Wyatt: So right now we're a lean team, but there are a few things we focus on doing/have done which are probably useful to other companies besides us:
• eln and some data management: benchling
• automated software for checking whether a sequencing run is done, migrating files to aws/s3 and running basecalling and several automated pipelines
• storing analysis results in a commonly accessible pipefiles server kept behind sso that allows everybody to reliably access and store data
Wyatt: For a lot of interesting new problems, there's no silver bullet when it comes to software. and i think most people who have built software for biology understand and appreciate that none of us have built a silver bullet :slightly_smiling_face:
Wyatt: We do focus on using and committing to/maintaining open source code. we write mostly in rust and python for safety, maintainability, and performance reasons.
Nicholas: Is your team all computational or do you have team members who don’t write code?
Wyatt: `3` of our team members, myself included, are folks-who-code. the other `4` are dedicated wet lab team members, including `2` of our founders!
Question 7:
Nicholas: Infinimmune is relatively young (~1 year i believe). how do you think about building software for your earliest r&d processes? in particular, how do you balance building “pipelines” vs “one-off” tools
Wyatt: Generally we try to not build "one-off" tools.
Wyatt: If there's a question that we've asked ourselves 2-3 times, we build tooling for it which can be deployed again with little effort.
Wyatt: When we build pipelines, it's all about mvp or maybe "minimum viable answers" -- what are the core questions you ask of every dataset? pick 2-3 and make sure those results are baked into a pipeline, and allow 2-4 weeks to do some "one-off-like" work to understand up from down first.
Nicholas: Oftentimes you might ask that question 2-3 times (or 5-10 times), but then never again. is building the tooling worth it in that case?
Wyatt: Yes, it's still usually worth building because at some point you'll want to know where a piece of company knowledge came from, and having reproducible tooling and code is the best way to futureproof for that!
Question 8:
Tj: Do you think we are reaching the limits of what we can accomplish with small molecule drugs, and will need to turn more and more to biologics like antibodies?
Wyatt: I don't think i'm qualified to give a resounding answer to this question!
Wyatt: At infinimmune, we like antibodies and biologics because the body already uses them.
Tj: Im really here for the hot takes
Tj: No qualifications necessary
Wyatt: Antibodies are the best, in my personal and very limited opinion :slightly_smiling_face:
Wyatt: They have excellent on-target and biodistribution properties, and in the past decade huge strides have successfully been made to move away from infusion/transfusion.
Wyatt: I think the big thing about small molecules is they're currently better able to reach internal targets/signaling molecules -- though maybe that will change as we get better at discovering and advancing agonistic antibodies and biologics that don't just act in "antagonize and block the thing i think is relevant."
all speculation though!
Tj: Cool!
Question 9:
Tj: Should all companies have a chief morale officer? is ringo available rn for questioning? i'd really be interested in their pov
Wyatt: Ringo says hello to everyone!
Wyatt: (and believes that all companies should have a canine morale officer or chief morale officer, yes)
Question 10:
Nicholas: We’ve been touching on the topic of rational design in a few places, but wanted to ask directly (https://bitsinbio.slack.com/archives/c02sav8du2j/p1687893891414509?thread_ts=1687893188.835639&cid=c02sav8du2j and https://bitsinbio.slack.com/archives/c02sav8du2j/p1687893193877099?thread_ts=1687892395.044259&cid=c02sav8du2j)
there’s obviously been a lot of excitement around ai in bio, especially in the protein space. where do you currently think ml tools are useful and where are you most excited for new tools to develop?
Wyatt: First things first, i'm not an ai expert!
Nicholas: This is your chance for a wishlist!
Wyatt: I think ml tools are very useful for a couple of things given the right training data. some recent examples include:
• annotation of new/unseen proteins
• prediction of impact on thermostability to a reasonable degree
• prediction of escape avenues for viral targets (personally a huge fan of the work from jesse bloom's lab on this)
Wyatt: I'm most excited for ai tools to work on problems that are usually addressed much, much further downstream for biologics -- think formulation, opalescence, viscosity, etc.
current state-of-the-art/state-of-what-you-can-write-a-paper-on is datasets of <200 or even <30 antibodies, which isn't particularly inspiring :grimacing:
Wyatt: Immunogenicity is a very difficult problem because you're dealing with the most polymorphic part of the human genome -- even more so than antibodies! being able to deal with immunogenicity predictions that are substantially more reliable for folks that don't come from very specific western european populations would be a great application for better tools.
Nicholas: Why do you think these problems haven’t been addressed yet? is it a data scale issues? a lack of interest from the field?
Wyatt: Definitely a data scale issue and the folks who can solve that problem are very niche experts. the reliable training data involves making a bunch of cell lines with specific hla alleles, eluting the bound peptides from hla, and then doing mass spec -- each of those are rate-limiting bottlenecks on their own! and there's an unbelievable amount of hla diversity.
Nicholas: Forgive the naive question, but why do you need to make specific hla alleles? wouldn’t you want a lot of diversity to train the models?
i guess this also touches on why you’re skeptical of public datasets here
Wyatt: You need specific alleles because the alleles have different peptide-binding fingerprints. so you need diversity on both the "many peptides bound by one allele" and "many alleles not just one" fronts.
Wyatt: And then you also open cans of worms like class i hla and class ii hla -- class ii hla has sliding/less narrowly definable binding registers :woozy_face:
Question 11:
Nicholas: We haven’t talked too much about the details around the single-cell work you’re doing. do you have any resources you can point us to for learning more about state of the art single-cell x antibody work?
Wyatt: Sure! a couple of noteworthy things:
• beam-ab from 10x (product i worked on!)
• how to retrieve stupidly potent human antibodies for covid that are bulletproof years after 2020
Question 12:
Nicholas: You touched on it a bit here: https://bitsinbio.slack.com/archives/c02sav8du2j/p1687894354768659?thread_ts=1687894167.751989&cid=c02sav8du2j, but if you had a magic wand and could solve just 1 problem that plagues ab discovery, what would it be and why?
Wyatt: For us, one very useful magic wand that we'd love to wave would be one that could reveal which antibodies would lose potency in a given manufacturing cell line -- antibodies are proteins which come with post-translational modifications, and being able to quickly figure out which ones are essential for function of an antibody would be great!
Question 13:
Nicholas: Can you tell us a bit about the founding journey? how did the team all meet? how did you decide to actually start a company?
Wyatt: So we all met and/or worked together in various capacities at 10x, but we hadn't all worked closely together before. each of our founding team members left 10x in 2022 for their own reasons, but one somewhat core/shared reason each of us had decided to leave was that we wanted to work on new problems and to not build tools for customers.
Wyatt: We had a beer together mid-july last summer and i pitched our soon-to-be team on "antibodies for humans from humans", and our other 4 founders thought that was a convincing and worthy problem, and we were off to the races :slightly_smiling_face:
Nicholas: Beers for bio would be a great happy hour segment
Wyatt: Totally! :beers:
Question 14:
Garrett: Do you prefer keeping biological data in flat files, databases, data warehouses, or somewhere else?
Wyatt: A sadly accurate and unhelpful answer is "it depends on the type of data."
an infinimmune-specific answer would be "we prefer flat files and databases for data generated in-house, in that order."
Wyatt: We try to keep recently used data on high-disk-speed-and-access cloud environments, and we try to write code that can be run locally (because then it scales really well in the cloud). this doesn't work well once you have a ton of data obviously, but for getting a grip on a problem as you start it's not a bad approach.
Wyatt: And we try to use more human-readable files and/or reports for things that more humans look at :slightly_smiling_face: (i.e. things that all of our team members look at and not just comp bio)
Garrett: Do you find that the performance and memory safety of rust is worth it vs throwing more python at problems?
Wyatt: Absolutely, not to mention rust versioning/portability is superior to python. no putzing around with `conda` or `pip` or any of that.
Wyatt: Rust isn't the answer to every thing, but it's an excellent answer for many things!
Garrett: Haha agreed
Nicholas: We have just a few minutes left. what did we forget to ask? what amazing part of infinimmune or your journey have we not talked about?
Wyatt: I think that was a fairly complete starting point! if you want to read what others have written about us or hear my answers to questions from other folks in podcast form, check out our company news page.
Wyatt: We could talk at length at a different time about how we secured our $12m seed round in 3 months last year, but that's a whole conversation of its own :slightly_smiling_face:
Nicholas: Amazing! thanks so much for your answers — that was a really fun q&a. if people have more questions, feel free to add them to the channel and if wyatt has time he can answer them
Wyatt: Thanks again for the opportunity to talk a bit about https://infinimmune.com|infinimmune! have a great day everybody! :nerd_face:
Tj: This was an incredible q&a thanks @wyatt!
Anonymous: This was super insightful !
thanks to @nicholas for the thought provoking questions and @wyatt for the insightful answers and sharing their journey of infinimmune !