Not a Discord Bot: The Real Dyno Therapeutics
A Google search for "Dyno" still surfaces a Discord automation bot first, a tool that manages server roles, welcomes members, and logs messages. The name collision obscures the biotech company: Dyno Therapeutics, co-founded by Eric Kelsic, which builds machine learning models to redesign the viral vectors that carry gene therapies into human cells.
Kelsic, Dyno's co-founder and CEO, did his PhD at Harvard under Martin Nowak and George Church, where he spent the final years focusing on model-guided AAV capsid design. Through Church, he met the team already applying machine learning to the same problem. That lineage — Church's lab, Nowak's evolutionary dynamics, high-throughput synthesis — shapes the company's approach.
Dyno focuses on engineering adeno-associated virus (AAV) vectors, the delivery vehicles that shuttle genetic payloads into target tissues. The field's bottleneck has never been the therapeutic gene; it's the capsule. Natural AAV variants offer limited tropism, high immunogenicity, and unpredictable manufacturing yields. Traditional approaches (directed evolution or rational design) are slow and low-throughput. Dyno bets that a trained model can navigate capsid sequence space more efficiently than wet-lab screening alone, predicting variants that target specific tissues, evade neutralizing antibodies, and scale in bioreactors.
The platform iterates between in silico prediction and high-throughput in vivo validation. Each cycle feeds the next: the model proposes variants, the lab tests them in animals, the data retrains the model. The loop compresses years into months. It reshapes the talent profile. A pure computational scientist lacks the biological intuition to constrain the search space; a pure virologist lacks the modeling fluency to exploit the platform. Dyno needs people who speak both languages.
The Platform: Machine Learning Meets Viral Vector Engineering
In a 2022 primer, Kelsic laid out the architecture: synthesize 100,000-plus variant sequences per round on Agilent or Twist chips; run high-capacity ML models that learn sequence-function relationships without heavy feature engineering; and treat the model as a "fake oracle" — query it cheaply before committing to expensive ground-truth assays.
That loop creates overlapping competency zones. First, the synthesis-and-assay engine: Kelsic said they "directly synthesize variation we're interested in at high throughput" and then "assay them for the functions we care about," blending molecular biology, virology, and automation. Second, the modeling layer: researchers fluent in regularized generative models (c-bas and d-bas from Berkeley), feedback-game frameworks (Stanford), and deep exploration networks (Celix) — measured by efficiency, scalability to batches of 100,000 to one million, reproducibility, consistency, independence from oracle bias, and adaptivity. The primer frames these as the evaluation criteria for any deployed algorithm. Third, the translation bridge: scientists who specify the right objective functions — tissue tropism, immune evasion, packaging capacity — and spot when an assay misses a downstream liability. As Kelsic put it, "we are not measuring toxicity, and that might disqualify solutions we find later," so diversity becomes a deliberate portfolio strategy.
Hybrid Skills Over Pure Credentials
Kelsic's own trajectory signals the profile: theoretical evolutionary biology PhD, pivot to model-guided AAV capsid design after meeting his co-founders through Church. That path (theoretical evolution to computational biology to applied ML for viral vectors) is the template.
Three markers distinguish this profile. First, hands-on experience with high-throughput synthesis and screening pipelines. Kelsic noted Dyno runs "10 to the five or more sequences" on Agilent or Twist chips, produced and assayed in parallel. Candidates who've only run simulations on public datasets haven't faced the noise, dropout rates, and library-design constraints of physical oligo pools. Second, a working mental model of AAV biology's failure modes. Kelsic framed it bluntly: "this is a viral capsid... the leading candidate for gene therapy... but in its natural form has a lot of drawbacks we hope to remove." Model-guided design, he said, "attempts to preempt nature's anger by simulating the process in silico before committing to the expensive synthesis step." That phrase, "nature's anger," signals the test: can the candidate anticipate which in-silico wins collapse in vivo? Third, project evidence over publication counts. A candidate who has built a generative model, synthesized the top-k sequences, measured transduction efficiency in a relevant cell type, and fed results back into the next cycle has already done the job.
A Broader Shift in Gene Therapy
Gene therapy has faced the same bottleneck for a quarter century. David Schaefer, professor at UC Berkeley and co-founder of 4D Molecular Therapeutics, calls delivery the field's "Achilles heel." For 25 years the dominant strategy was empirical: screen natural AAV serotypes, pick the best, hope it works in humans. Six AAV-based therapies have reached FDA approval that way. But the ceiling is visible. High-dose systemic delivery (65 milligrams of virus per patient) has triggered immune responses, off-target liver toxicity, and deaths in trials. Natural capsids evolved for infection, not medicine. The field now confronts the limits of what nature provided.
4D Molecular Therapeutics, co-founded by Schaefer in 2013, has built over 40 libraries, each with 10 to 100 million variants, totaling over a billion engineered capsids. Their AV2 retro variant achieves 50-fold better retrograde transport than any natural AAV tested. In the eye, an engineered capsid could enable intravitreal injection covering the entire retina, replacing subretinal surgery that reaches only one-tenth of the surface. In neurodegenerative models, the same platform delivered the first CRISPR efficacy in ALS and the first CRISPR-mediated lifespan extension in Huntington's disease. Schaefer acknowledges: "Manufacturing at these scales still needs progress, but it's good enough to reach the clinic."
Dyno pursues a parallel path: model-guided design rather than directed evolution. Both approaches treat vector design as an engineering problem with a search space large enough to require computational navigation. That shift demands a workforce standard academic pipelines don't produce. Universities still train virologists who code in R and computer scientists who've never seen a biosafety cabinet. The companies winning the hiring race write job descriptions for a hybrid that doesn't yet have a textbook.
The Discord bot still owns the search result. But the biotech Dyno now builds a platform where the loop between model and capsid closes: the visible tip of a workforce rearchitecture the entire gene therapy sector is undergoing.
Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.