Biology Is Becoming an Engineering Discipline

For most of its history, biology was a science of patient observation. You collected specimens, ran gels, waited for colonies to grow, and hoped the results told you something useful. The pace of discovery was governed by the pace of nature. The biggest breakthroughs -- penicillin, CRISPR, monoclonal antibodies -- came from researchers who noticed something unexpected and spent years figuring out what it meant.

That model is ending. Not because biology has run out of surprises, but because the tools have changed enough to flip the discipline's fundamental orientation -- from finding what exists to building what doesn't. The biggest biotech companies of the next decade won't discover biology. They'll engineer it.

The Compression Problem

Every experimental science lives and dies by its design-build-test cycle. In software, that cycle can run in minutes. A developer writes code, compiles, runs, and iterates before lunch. In traditional wet lab biology, the same loop could take months -- design a protein candidate, synthesize it, express it in cells, purify it, test it for activity, fail, and begin again. The sheer friction of that process meant that even well-funded labs could only test a small corner of the possible molecular space.

Three forces are now compressing that cycle in biology the way compilers and version control compressed it in software.

The first is AI-driven molecular design. Large language models trained on protein sequences -- and the physical structures those sequences fold into -- can now propose novel proteins, antibodies, and gene editors with specified properties before a single experiment runs. The design phase, once measured in researcher-hours, collapses to compute hours.

The second is high-throughput automation. Robotic platforms can run hundreds of parallel experiments, dispensing nanoliter volumes with precision no human pipette can match. What used to require a team of postdocs working nights now runs unattended on a liquid-handling robot while the scientists sleep.

The third, and least appreciated, is the feedback loop between the two. AI generates hypotheses at scale; automation tests them at scale; the results train better AI. The cycle becomes self-reinforcing. Biology starts to behave less like a field and more like a codebase -- one that accepts pull requests from machines.

The Companies Writing the New Playbook

EvolutionaryScale is perhaps the clearest expression of this shift. Founded by Alexander Rives and other researchers who built Meta's ESM protein language models, the company is training on evolutionary data -- billions of years of biological trial and error encoded in the sequences of every known organism. Their ESM3 model treats proteins as a kind of language, with sequence, structure, and function as three parallel vocabularies that can be read and written simultaneously.

The implications are significant. Evolution has already explored an enormous but still bounded region of protein space. ESM3 can interpolate between what evolution found and generate candidates in the gaps -- proteins that have never existed in any organism, optimized for properties evolution never had reason to select. This is not protein discovery. It is protein authorship.

Profluent took the same logic and applied it to gene editing. In 2024, the Berkeley-based company published OpenCRISPR-1, the first fully AI-designed gene editor -- a functional protein conceived entirely by a generative model with no evolutionary ancestor. It worked. The protein edited human cells with precision comparable to natural CRISPR systems, and Profluent open-sourced it, the biological equivalent of releasing software to the commons.

The proof-of-concept here is more important than the specific protein. It establishes that the generative approach works not just for passive structural proteins but for complex molecular machines with precise functional requirements. If you can design a gene editor from scratch, you can design enzymes, receptors, and therapeutic payloads with similar methods. The tool is general.

Generate Biomedicines is applying generative AI to one of the most commercially important categories in medicine: antibody therapeutics. Antibodies are the backbone of the modern biologics industry -- drugs like Humira, Keytruda, and Dupixent represent hundreds of billions in annual revenue. But developing them traditionally requires years of animal immunization, screening, and optimization. Generate's platform treats antibody design as a generative problem, producing candidates computationally and screening them in parallel. The company has partnered with Amgen, which suggests the established industry is not betting against this approach.

Latent Labs, founded by veterans of Alphabet and DeepMind's biology programs, is working at the level of molecular generation itself -- using diffusion models, the same architecture behind image generators like Stable Diffusion, to propose molecular structures from scratch. Where image diffusion models learn to reconstruct photos from noise, molecular diffusion models learn to reconstruct viable chemical structures. The generative intuition is identical; the substrate is biology instead of pixels. Latent is building in a quieter register than some peers, but the technical lineage -- people who worked on AlphaFold-adjacent problems -- gives them a credible claim on the hard parts of the problem.

Parallel Bio is solving a different but equally important bottleneck: testing. Even if you design a perfect protein, you still need to know how it will behave in a human immune system before you can give it to a patient. The current answer is animal models -- expensive, slow, and often poorly predictive of human response. Parallel Bio is building synthetic human immune systems in vitro, culturing human immune cells into functional organoid-like systems that respond to drugs the way a real patient would. The goal is to replace animal testing with something that is simultaneously faster, cheaper, and more accurate. If they succeed, the design-build-test loop tightens yet again, this time at the validation stage rather than the design stage.

Biology as a Programmable Medium

What unites these companies is not a single technology but a shared assumption: that biology is a programmable medium, and that the right abstractions can make it as manipulable as software.

This is a genuinely radical claim. Software runs on silicon built to precise specifications. Biology runs on cells that evolved under selection pressures humans didn't set and can't fully model. The assumption that you can engineer reliably on top of that substrate is not guaranteed to hold -- and the history of biotech is littered with platforms that worked beautifully in the lab and failed catastrophically in humans.

But the bet is that better models and faster iteration can manage that complexity the way good engineering manages complexity in any domain: not by eliminating it, but by developing tools robust enough to work despite it. AlphaFold's structure predictions are not perfect, but they are useful enough that the field reorganized around them. The same pragmatic standard applies here.

What engineering disciplines create that observational sciences do not is compounding leverage. Software engineers don't rewrite from scratch each generation -- they build on libraries, frameworks, and platforms accumulated by thousands of prior contributors. The field improves faster than any individual can by hand. Biology, once treated as an engineering discipline with shared tooling and reusable components, could accumulate leverage the same way. A protein scaffold designed by EvolutionaryScale becomes a building block for a therapeutic at Generate. A gene editor published by Profluent becomes infrastructure for a next-generation therapy no one has yet imagined.

The companies that understand they are building infrastructure -- not just products -- are the ones worth watching. The design-build-test cycle is compressing. What gets built inside it is only beginning to come into view.