In 2024, David Baker of the University of Washington shared the Nobel Prize in Chemistry for computational protein design, building proteins that don't exist in nature, from scratch, to do a specific job. His lab has produced thousands of them: new enzymes, vaccine components, cancer-targeting molecules. The tools that made this possible, RFdiffusion for generating protein backbones and ProteinMPNN for designing the sequences to fold into them, are open-source and used by labs worldwide.

This is a different capability than the structure prediction covered in AI Doomsday's look at biologists, where AlphaFold predicts the shape of an existing protein. Biotech engineering is about designing biological systems that don't exist yet: proteins, metabolic pathways, engineered microbes, industrial bioprocesses. Biotech engineers are rated Moderate risk, with a horizon toward 2044, and the reason the horizon is that long comes down to a single stubborn fact about this field.

An AI can design a protein in minutes. Whether that protein actually folds, expresses, binds its target, and does something useful inside a living cell is a question only a wet lab can answer, and the answer, far more often than the headlines suggest, is no.

Key Points

  • Biotech engineers are rated Moderate risk with a horizon toward 2044, as generative AI accelerates the design phase of biological engineering while experimental validation remains a hard, slow, unautomated bottleneck.
  • Generative design tools including RFdiffusion and ProteinMPNN, central to David Baker's 2024 Nobel Prize, now design novel enzymes and binders, and genomic models like the Arc Institute's Evo 2, trained on 9.3 trillion nucleotides, can generate genome-scale sequences.
  • Published "90% binding" figures are measured in ideal conditions; real production campaigns report experimental hit rates for AI-designed binders in the low single digits, with aggregation and expression failure the largest category.
  • The field has already seen the "science works, business doesn't" failure mode: Zymergen imploded within four months of a $3 billion IPO in 2021, and Amyris filed for bankruptcy in 2023 despite functioning technology.
  • The U.S. Bureau of Labor Statistics projects 8% employment growth for biomedical engineers through 2035, faster than average, with a median salary of $109,370.

What a Biotech Engineer Actually Does

The role spans designing proteins and enzymes for a target function, engineering metabolic pathways so a microbe produces a desired molecule, optimizing fermentation and downstream processing to manufacture that molecule at scale, and running the design-build-test- learn cycles that connect a computational idea to a validated, manufacturable product. It sits at the intersection of molecular biology, chemical engineering, and increasingly, machine learning, and the "engineer" in the title is load-bearing: the job is judged on whether the system works in reality, not whether the design is elegant.

What AI Is Already Doing

On the design side, the acceleration is real. RFdiffusion, now in its third open-source generation, designs DNA-binding proteins and advanced enzymes at success rates that would have been implausible a few years ago, when de novo design hit rates could be one in ten thousand. ProteinMPNN recovers native-like sequences at roughly 52% accuracy, well above the older physics-based methods, and is routinely used to rescue designs that earlier tools failed on. The Arc Institute's Evo 2, a 40-billion-parameter model trained on 9.3 trillion nucleotides across 128,000 genomes, can identify pathogenic mutations at over 90% accuracy and generate synthetic sequences at the scale of a simple bacterial genome.

On the build side, Ginkgo Bioworks runs organism engineering as an industrial process: automated strain construction executing tens of thousands of design-build-test-learn iterations in parallel, serving more than 130 customers including Novo Nordisk, Bayer, and Pfizer, on top of a platform the company says represents around $2 billion of accumulated R&D. Self-driving lab platforms, where an AI agent plans experiments, writes the protocols, runs them on cloud robotics, and iterates on the results, are moving from demonstration to early deployment.

THE SILICON-TO-CELL GAP

The "90% binding" figures that make headlines are measured under ideal conditions on curated targets. Protein engineering consultancies that track real production campaigns report experimental hit rates for raw AI-designed binders in the low single digits, with aggregation and poor expression the largest failure category, followed by affinity too weak for therapeutic use and binding to the wrong site. Multiple studies have found that failed designs sometimes carry better confidence scores than successful ones, which is precisely why experimental testing cannot be skipped, and why the wet lab remains the rate-limiting step.

When the Science Works and the Business Doesn't

This field has its own version of the pattern AI Doomsday keeps finding in professions where working technology still fails to displace people: the technology works, and the company built on it doesn't. Zymergen, a SoftBank-backed synthetic biology company valued above $3 billion after its 2021 IPO, imploded within four months when its flagship product, a biofabricated material, failed to reach market. Amyris, which had real, functioning technology for producing engineered molecules, filed for Chapter 11 in 2023 after spreading itself across fermentation, ingredients, and consumer brands while burning cash it couldn't replace. The engineering was not the problem in either case.

LanzaTech is the counterexample worth holding alongside them: it uses engineered biocatalysts to convert industrial carbon emissions into fuels and chemicals, runs commercial plants, and has supply relationships with companies like Unilever and BASF. The difference between LanzaTech and Zymergen was never the sophistication of the biology. It was whether there was a product a customer wanted badly enough to sustain the infrastructure.

The Labor Market Reality

The U.S. Bureau of Labor Statistics projects 8% employment growth for biomedical engineers, the closest tracked category, from 2025 to 2035, faster than the average occupation, with a median annual wage of $109,370. That projection largely predates the generative-design wave, and industry commentary suggests the effect of AI on this role is more likely to be a shift in the skill mix, toward engineers who can bridge computational design and wet-lab validation, than a contraction in headcount. The bottleneck the field is trying to solve is a shortage of people who can do both halves well.

How to Use AI as a Biotech Engineer Now

For design: generative tools genuinely compress the ideation phase. Build the developability and expression filtering into your workflow before synthesis, since the raw output's real-world hit rate is far below its benchmark numbers.

For build-test-learn cycles: automated platforms and self-driving labs are worth adopting for throughput, but the value you add is in framing the right experiment and interpreting an ambiguous result, not in running the pipettes.

For career positioning: the engineers least exposed are the ones who can move fluently between the computational design layer and the experimental reality that validates or kills it. Pure in-silico design skill without wet-lab judgment is the more automatable half.

What I Think

The 2044 horizon looks right, and the silicon-to-cell gap is the reason. Design has genuinely been accelerated to the point where it is no longer the constraint. Validation has not, and there is no credible near-term path to automating away the need to actually build the thing and see if it works in a living system that doesn't care how confident the model was.

What belongs in an honest accounting of this field is the biosecurity question. Microsoft Research and others have shown that AI-generated proteins can be functionally equivalent to known toxins while sharing almost no sequence similarity, which makes the homology-based screening that synthesis providers rely on effectively blind to them. That is a governance and safety problem for the field and its regulators, running in parallel to its genuine scientific promise, and it is not one the labor-market framing captures.

"The model designed a thousand proteins over lunch. Twenty of them might fold. The job is figuring out which twenty, and there's still only one way to find out."