Generate Biomedicines: Designing New Proteins with Generative ML
What do they do?
Generate applies machine learning to create wholly new proteins, which they term “generative biology”.
These proteins are unique because they are completely new.
Algorithms trained on millions of samples of previous protein-protein interactions can understand the rules and patterns of protein interactions. With this systemic knowledge, new proteins can be intelligently designed to fuel specific interactions (e.g. therapeutic intervention in cancer).
They call these proteins biomedicines:
Generating an effective biologic medicine requires control of the relationships between protein sequence, structure, function, and immune activation. Our generative biology platform can recode protein sequences while retaining structure and function and evading immune activation.
Imagine a disease for which a protein is a therapeutic. To date, desgining a protein for this use case has been challenging. Immune response is a challenge. Using generate, we could design the right de novo protein and use it to drug a target.
What’s their story from the beginning to now?
Generate is a fusion of two previous Flagship incubation efforts both focused on protein engineering, machine learning, and large-scale protein data. The founding team consisted of Geoff von Maltzahn, Avak Kahvejian, Molly Gibson, all Flagship team members, and Gevorg Grigoryan, a Dartmouth professor focused on protein mapping. The joint team began incubating the company in stealth and officially launched in September 2020.
They raised $370M in November 2021. This funding went towards supercharging the company’s workforce and lab space in order to tighten the loop between designing compounds in silico and testing them in the lab.
In June 2022, they signed their first major R&D partnership agreement with Amgen, with the potential for ~$2B in payments and subsequent royalties. As part of the deal, the two companies will partner on 5-10 drug candiates.
Why does it matter?
The first reason Generate matters is because they are targeting a uniquely meaningful space. Proteins are the functional layer of biology. Designing them is the holy grail of biotech. Many companies have aspired to apply computation to this challenging, valuable problem. If Generate can develop a scalable approach to this recurrent problem of protein creation, it would be hugely valuable. They are not the first, but their goal is relevant.
The second reason Generate matters is because the generative approach they are taking is fresh. Most discussion in protein science revolves around understanding existing protein structures in nature and human biology. However, the truly unique aspect of machine learning is its ability to elicit underlying patterns and apply them in often unexpected ways. Generate is harnessing this ability by building ML models that can learn from existing protein structures and ultimately design new ones, which would open up a vast new realm in therapeutics.
Finally, Generate matters because it has stacked talent. Apart from the tremendously impressive founding team from Flagship, Generate has attracted names like:
- Nobel winner Frances Arnold,
- machine learning expert Andrew Beam, who teaches at Harvard and has authored numerous interesting ML blog posts and papers
- machine learning expert Jonathan Ingraham, who worked with Tommi Jaakola and Regin Barzilay at MIT
This is just a sampling of the talent, which has a very deep bench of lab experts, computational experts, and biotech strategy experts.
- https://generatebiomedicines.com/biomedicines
- https://generatebiomedicines.com/news/generate-biomedicines-origin-story
- https://www.flagshippioneering.com/companies/generate-biosciences
- https://www.flagshippioneering.com/stories/the-coming-age-of-generative-biology
- https://spectrum.ieee.org/generate-biomedicines-emerges-from-stealth-mode-and-takes-aim-at-covid19
- https://www.businesswire.com/news/home/20211118005426/en/Generate-Biomedicines-Announces-First-External-Equity-Raise-of-370-Million-to-Advance-Its-Drug-Generation-Platform
- https://endpts.com/flagships-machine-learning-startup-generate-bags-370m-in-latest-round-with-plans-for-a-big-hiring-spree/
- https://www.fiercebiotech.com/biotech/amgen-chooses-generate-for-1-9b-biobucks-deal-to-churn-out-up-to-10-multi-modality-drug
- https://pubmed.ncbi.nlm.nih.gov/33910104/
- https://yangkky.github.io/2018/03/26/learning-the-language-of-proteins.html
- https://medium.com/lifesci-beat/molly-gibson-co-founder-of-generate-biomedicines-on-innovating-at-the-intersection-of-ai-ml-and-3338652f72fd
Cellarity: Graph ML and Network Science to Model Disease
What do they do?
Cellarity applies network science to biology to understand the complexity of disease.
To do this, they make use of advances in transcriptomics and single cell mapping to understand how entire systems of cells work.
In practical terms, what this looks like is a combination of wet lab and machine learning work. Scientists in the wet lab perform experiments to modify and disease systems of cells. These systems are then intensively mapped down to the granularity of individual cells. The resulting transcriptomic data, or mRNA readout, is used to inform a network science-based model of the tissue. Labeled data from changes to the network via experiments are fed to machine learning models, which are then used to predict potentially efficacious interventions.
What’s their story from the beginning to now?
Funnily enough, I have a personal story with Cellarity. In the spring of 2019, I was introduced to the early team there (before they came out of stealth). I was terrible at Python so I didn’t get the job, but it was an interesting window into the formation of an early Flagship biotech.
Cellarity was founded by the Flagship team of Nick Plugis and Avak Kavahejian. By the time I met them in March 2019, they had grown to a team of 50 and had added Chad Nussbaum as CTO and Milind Kamkolkar as CDO. They raised $50M in December of 2019.
A year later, they raised $123M and added Fabrice Choraqui as CEO.
They’ve since been a bit of a black box. Unlike Generate, they haven’t announced any major deals. They’ve also seen a fair bit of executive turnover, as much of the digital team has left. They have a number of candidates programs that are advancing.
To date, they’ve been rather quiet about the progress or future of their technology. Of course, they have an incredible amount of runway. They recently moved into new Flagship-focused lab space, indicating a strong commitment to the wet-lab + computation model of drug design they were created around.
Why does it matter?
The first reason they matter is cultural approach. Cellarity invested deeply in the integration between wet lab work and computation. They carefully designed experiments to generate massive volumes of high quality single cell data. They applied software engineering to their lab notebooks to fuel the exchange of knowledge between wet lab and computational professionals. Most importantly, they infused leadership across each of these disciplines at the highest levels. These kinds of investment built a unique cross-cutting culture. No department superseded the other at Cellarity (as they do so often in traditional biotech and pharmacy). All disciplines worked in harmony.
The second reason is talent. Cellarity pulled together a tremendous agglomeration of technical talent. There were and are a lot of really great people that worked at Cellarity. In my own interview cycle, I was deeply impressed at the bench they assembled. Some names that pop out are Alex Wolf (the creator of Scanpy), Umut Eser (Head of Deep Learning), Jesse Johnson (Head of Data Eng), and others.
The third reason they matter is timing. Cellarity was one of the first true powerhouse AI in biotech companies assembled (circa 2018). Their results are a referendum on that approach of that era, which infused tremendous amounts of capital into companies that bet heavily on advanced deep learning as a panacea to drug discovery. As we enter a bear market, particularly for biotechs, I will watch Cellarity’s results closely to see if they can raise again and demonstrate significant scientific results publicly. Has their approach succeeded? We will find out soon.
Valo Health: Re-engineering the entire drug development process
What do they do?
Valo is building an end-to-end drug discovery and development platform that leverages a unified machine learning and data architecture to drive better results in creating new therapeutics.
Put simply, Valo wants to do everything. In their world view, the current drug creation paradigm is artificially broken into individuals steps (e.g. target identification, lead optimization, etc.) that are actually inefficiences in the process of discovering and developing drugs.
The legacy biopharma model is basically this point-to-point system [where up to 15 groups] do some work, and then they basically take the result of it and they throw it over a wall to another group that has its own framework. The model is intrinsically disintegrated… We’ve created a single underlying architecture so that we’re using the same species, we’re using the same decision-making criteria. We’re using the same KPIs throughout the entirety of the R&D cascade.
Valo’s bet is that by completely reimagining this process starting with a really sexy data architecture and with machine learning algorithms baked in at every step, you can power transformatively better results in drug development.
It’s bold. But does it make sense? How does it actually work and look different from the traditional drug development process?
On their website, Valo highlights three examples of steps that they have done differently: target discovery, molecule design, and clinical development. For each process, they try to apply different sources of data and “intelligence” to improve the output of the process. For example, they claim that their active learning machine learning system for molecule design can “go from zero to a molecule in hand in as little as 3 weeks” by “[screening] trillions of molecules computationally and billions of molecules empirically in a matter of days to weeks.”
That’s an impressive claim, and it exemplifies Valo’s approach to the entirety of the drug development stack: tear it up, rebuild it with machine learning, and see how much better it is. You could write a whole article outlining the various ways Valo is trying to do this (which I might do at some point).
The hard part here is actually delivering the results all the way downstream, with real drug outcomes.
What’s their story from the beginning to now?
Valo was founded in 2019 under the leadership of David Berry, a general partner at Flagship. Soon after their founding, they came out of stealth with a massive $100M Series A. They parlayed this large Series A into unusually acquisitive behavior reminiscent of another well-funded biotech, Resilience. Valo bought two companies, Numerate and Forma Therapeutics in 2019 into 2020.
In early 2021, they raised an additional massive $300M Series B. This round was backed by a number of pension funds and larger funds than traditional venture firms, indicating a unique investor composition.
All of this activity culminated in an extremely active end to 2021. In June 2021, Valo sought to take advantage of the SPAC boom and go public in a SPAC with Khosla Ventures. Curiously, the company and Khosla Ventures actually terminated the SPAC deal, which would have contributed up to $750M, due to “adverse market conditions” in November 2021. This was likely due to cratering biotech valuations, which have been hammered in the past year.
Since the termination of their SPAC, Valo has continued to be acquisitive and active on the BD side in 2022. They have advanced two drugs into Phase 2, acquired two additional companies (TARA Biosystems), and signed a unique partnership with CRO Charles River Labs to sell integrated clinical research and computational drug development services. Additionally, they’ve announced a specific focus on drug candidates in cardiovascular, oncoloy, and neurodegen
Why does it matter?
Valo Health is stupid ambitious. Plain and simple. Their ambitious vision is the top reason they matter. Many companies have tackled reinventing parts of the drug development lifecycle with AI. For example, Atomwise specializes mostly in the discovery phase and then partners on promising compounds with larger firms. Valo aims to entirely own and change the way drug development is done, from target to drug to trial. It’s a bold vision that actually makes a lot of sense.
The second reason Valo really matters is their exceptionally deep executive bench. David Berry, their Flagship founder, is one of the top VCs in biotech with an incredibly stellar pedigree—undergrad from MIT, an MD/PhD in the HST program, and tutelage under the Robert Langer of MIT. The wunderkind has helped found 30+ companies while at Flagship and continues to serve as founder-CEO. Few are better positioned to run such an ambitious company. Additionally, they have several unique and experienced executive. Nish Lathia, their Chief Product Officer, is a veteran of the Amazon machine. Hilary Malone, their COO, used to regulatory at Sanofi. Brandon Allgood, their Chief AI Officer, came over from the Numerate acqusition and is an experienced data science and machine learning lead. Their whole executive suite is exceptionally strong and a possibly compelling reason to believe in their success.
Finally, their capital raises and acquisitive behavior make them an interesting case study to watch. Many companies raised tons of money in the biotech craze of 2020 following pandemic hype. Nearly every company had a COVID-19 story and raised far more money than they knew what to do with. Of course, they never expected biotech valuations to fall as far as they have in the past year. This complicates outcome stories for all these companies, especially the clock is ticking and their burn continues. We’ll see how this plays out for Valo, which couldn’t complete its SPAC due to the sudden change in valuations.