Why Biology Is Not Software

The metaphor of programming biology is attractive because both computers and cells process information. DNA stores instructions. RNA carries information. Proteins execute functions. Regulatory networks determine which instructions are active. But the analogy breaks down quickly. Software executes on hardware engineered to be deterministic. Biology evolved through selection, redundancy, noise and adaptation. The same genetic change can behave differently in different cell types, organisms, environments or developmental states.

This matters because AI systems are excellent at finding patterns in high-dimensional data. They can learn statistical relationships between sequence and structure, structure and function, or experimental condition and measured outcome. But biological causality is often distributed. A mutation that improves one property may degrade another. A protein that becomes more stable may lose catalytic activity. A metabolic pathway optimized for yield may stress the host cell. A therapeutic candidate that binds strongly in vitro may fail because it is rapidly cleared, triggers immunity or cannot reach the right tissue.

The programmable-biology thesis is therefore most credible when it means biology is becoming more designable, testable and automatable—not when it implies perfect predictability. The engineering objective is to reduce uncertainty, not abolish it.

From Protein Structure Prediction to Biological Infrastructure

AlphaFold changed the scientific baseline because protein structure is deeply connected to function. Before modern AI-based structure prediction, determining a high-resolution protein structure often required expensive experimental methods such as X-ray crystallography, cryo-electron microscopy or nuclear magnetic resonance. Those techniques remain essential, but AlphaFold dramatically expanded the number of proteins for which researchers can begin with a credible structural hypothesis.

The AlphaFold Protein Structure Database, developed by Google DeepMind and EMBL-EBI, now provides open access to more than 200 million predicted protein structures. In 2026 the database expanded further with community datasets and millions of predicted protein complexes, including homodimers and heterodimers. That turns structure prediction from a one-off model capability into infrastructure that researchers can query before designing experiments.

AlphaFold 3 extends the problem from single-chain structure toward interactions among proteins, nucleic acids, small molecules and other biomolecular components. Even here, uncertainty remains important. A 2026 Nature Biotechnology paper showed that experimental measurements can improve inference of protein conformational ensembles, highlighting a major limitation of static structure prediction: proteins are not rigid objects. They move through multiple states, and function can depend on those dynamics.

The deeper lesson is that AI-generated structure is a starting point for biological reasoning, not a substitute for measurement. The more consequential the biological claim becomes, the more important it is to reconnect prediction to physical evidence rather than allowing a model output to become its own proof.

From Predicting Proteins to Designing Them

Prediction asks: given this sequence, what structure or interaction is likely? Design asks the inverse question: given a desired structure or function, what sequence should exist? That inverse problem is much closer to engineering because the system must propose something that biology has not necessarily produced through natural evolution.

ProteinMPNN helped establish a powerful form of inverse folding by generating amino-acid sequences expected to adopt a specified backbone. RFdiffusion introduced diffusion-based generative modeling for protein backbones, allowing researchers to design novel scaffolds, assemblies and binders. These methods can be chained: one model proposes a backbone, another generates sequences predicted to fold into it, and structure-prediction tools check whether the designed sequence is likely to produce the intended geometry.

A 2026 Nature review described de novo protein design as entering a new phase in which several structural design problems are close to being solved, while more demanding functions—especially catalysis, switches and complex conformational machines—remain difficult. That distinction is essential. Designing a stable shape is not the same as designing a molecule that performs a sophisticated chemical reaction efficiently under real biological conditions.

Google DeepMind’s AlphaProteo is another example of the transition from prediction to design. The system was built to generate new protein binders for specified targets. Such systems can reduce the search space dramatically, but the designed proteins still require experimental screening and characterization. The practical value comes from increasing the fraction of tested candidates that are worth pursuing.

The AI Model Stack in Biology

There is no single “biological AI model.” The emerging stack includes structure predictors, sequence language models, generative backbone models, inverse-folding models, property predictors, molecular docking systems, experiment-planning models and laboratory-control systems. Some learn from protein sequence. Others learn from 3D coordinates, molecular graphs, assay data, microscopy images or text. Increasingly, researchers combine several models in one workflow rather than expecting one foundation model to solve the entire biological problem.

The analogy to the broader AI stack is useful. A language model may reason over text while external tools search the web or execute code. In biology, a generative model may propose a sequence, a structure model may evaluate it, a simulation may estimate interactions, and an automated laboratory may test the candidate physically. The model is therefore only one layer in a tool-using scientific system.

This also means benchmark scores can be deceptive. A model that predicts a structural metric accurately may not improve the final experimental success rate. A sequence model that produces realistic-looking proteins may not produce useful proteins. The correct benchmark is downstream: does the system increase the number of successful experiments, reduce the cost per validated candidate or accelerate a real scientific objective?

Why Laboratory Validation Still Dominates Reality

Biology punishes overconfidence. A computer model can generate a candidate in seconds, but a laboratory must answer whether the candidate exists as a stable molecule, whether it folds, whether it expresses efficiently, whether it performs the intended function, whether it interacts with unintended targets and whether it behaves consistently across conditions. Each question can invalidate an otherwise impressive computational result.

This is why experimentally grounded models are becoming important. The 2026 Nature Biotechnology work on experiment-guided AlphaFold 3 is one example: instead of treating structural prediction as a self-contained computational truth, the framework uses experimental measurements to constrain the generated structural ensemble. That is likely to become a broader pattern across AI biology. Models will increasingly incorporate physical measurements during inference or retraining rather than merely being evaluated after the fact.

The same logic applies to generative protein design. A model can increase the prior probability that a candidate will work. It cannot remove the need to test candidates. The economic value of AI comes from reducing the number of failures required to reach a success.

The Rise of the Biofoundry

A biofoundry is the physical infrastructure that connects computational design to repeatable laboratory execution. It combines robotics, automated liquid handling, laboratory information systems, biological synthesis, high-throughput assays and data analysis. Instead of a scientist manually pipetting every experiment, the laboratory becomes partially software-defined. Protocols can be executed repeatedly, measurements can be captured consistently and models can select the next round of candidates.

The Department of Energy’s July 2026 highlight on iBioFAB shows why this matters. Researchers combined AI-guided design with automated protein construction and characterization. The platform tested enzyme variants, measured performance and fed the results back into another model. This is the biological equivalent of closing the control loop.

Automation also creates new failure modes. A robotic system can reproduce a bad protocol very efficiently. Instrument calibration errors can contaminate large batches of data. Model biases can cause the same region of design space to be explored repeatedly. Human expertise remains essential for interpreting anomalies, redesigning assays and recognizing when the automated system is confidently optimizing the wrong objective.

The Design-Build-Test-Learn Loop

Traditional biological engineering often spends most of its time between ideas. Researchers wait for synthesis, culture growth, purification, sequencing, analysis and manual interpretation. Automation compresses those delays. AI can prioritize the next experiment immediately after results arrive. Robotic platforms can execute standardized protocols overnight. Statistical models can update continuously instead of waiting for a human-authored research cycle.

This creates the possibility of self-driving laboratories, but the phrase should be interpreted carefully. A laboratory can automate experiment selection and execution without possessing general scientific understanding. Closed-loop optimization is strongest when the objective is measurable and the experimental space is bounded. Improving enzyme activity is more tractable than asking a machine to discover a completely new theory of cellular regulation.

The advantage is therefore incremental intelligence at machine cadence. Each round generates data. The data changes the model. The model changes the next experiment. If the assay is trustworthy and the search space is well designed, hundreds of iterations can be explored more systematically than through manual trial-and-error.

Enzyme Engineering as a Current Proof Point

Enzymes are strong candidates for AI-enabled engineering because their function can often be measured through well-defined assays and because industry already relies on enzymes for chemical production, food processing, pharmaceuticals, agriculture and bioenergy. Improving stability, specificity or catalytic efficiency can translate directly into lower manufacturing cost or better process yield.

The DOE-supported iBioFAB work is notable because it integrated AI and automation rather than treating protein design as a purely computational problem. The research platform improved the activity of two industrially relevant enzymes by large multiples, according to DOE’s summary of the published work, while reducing manual intervention. The result illustrates a broader industrial thesis: the value may come less from a single foundation model and more from the productivity of the entire experiment loop.

Recent research also shows AI being used to redesign starting proteins before directed evolution. A 2026 Nature paper demonstrated that AI-based protein redesign can create more stable starting points for subsequent laboratory evolution. This suggests AI and traditional experimental evolution are complementary. Models propose better starting regions of design space, while experiments continue exploring the unpredictable details that the model does not capture.

Biomanufacturing Is the Real Scale Test

A designed biological molecule can succeed scientifically and fail industrially. Manufacturing requires organisms or cell systems that grow predictably, raw materials that remain affordable, purification steps that remove unwanted compounds, quality-control systems that detect variation and facilities capable of producing at reliable scale. In therapeutics, the regulatory burden is even greater because consistency, contamination control and clinical safety become central.

Biomanufacturing therefore has its own maturity ladder. A sequence generated in silico is the earliest stage. Experimental expression proves the molecule can be produced. Functional assays establish activity. Pilot-scale fermentation or production establishes a process. Scale-up tests whether the process remains stable in larger vessels. Commercial production requires repeatability, supply chains, quality systems and economics.

This is where the software metaphor fails most decisively. Scaling software often means allocating more compute. Scaling biology can change the biology itself. Oxygen transfer, nutrient gradients, temperature, shear forces and cellular stress behave differently in a large bioreactor than in a laboratory flask. A strain optimized at one scale may produce lower yield or unexpected byproducts at another. AI may help model and control these systems, but it does not repeal chemical engineering.

Biological Data Becomes Industrial Infrastructure

AI biology depends on databases assembled over decades: protein sequences, structures, functional annotations, genetic variation, assays, expression measurements, molecular interactions and clinical data. AlphaFold itself became much more useful when its predictions were released as a searchable global database. In 2026 EMBL-EBI expanded the database with community-contributed structural datasets and millions of predicted protein complexes, increasing the amount of machine-readable biological context available to researchers.

This creates a strategic issue analogous to AI training data in other domains. Biological data quality affects model quality. Experimental protocols differ. Negative results are often underreported. Rare organisms and under-studied populations are poorly represented. Some commercial datasets are proprietary. Clinical and genomic data raise privacy and sovereignty questions. The future advantage may therefore belong partly to institutions capable of generating high-quality experimental data rather than merely training larger models.

Closed-loop laboratories strengthen that advantage. A company or laboratory that operates its own automated experimental platform can generate proprietary datasets directly optimized for its models. The feedback loop becomes a competitive moat: model proposes experiment, laboratory produces data, data improves model, improved model proposes better experiment.

Dual-Use Risk and Synthetic Nucleic Acid Screening

The same tools that make legitimate biological design easier can also lower barriers to misuse. That does not mean protein-design systems automatically enable dangerous biological construction, but it does mean governments are paying closer attention to where digital biological information becomes physical material. One of the most important control points is synthetic nucleic-acid procurement.

The U.S. government created a nucleic-acid synthesis screening framework requiring federally funded researchers to use providers that screen customers and sequence orders. A May 2025 executive order directed the government to revise or replace that framework and strengthen verification and enforcement. HHS and ASPR currently indicate that the 2024 framework is under revision, which means the policy environment remains active rather than settled.

The underlying logic is straightforward: model outputs remain digital until someone synthesizes DNA or RNA, cultures cells or otherwise turns information into biological material. Sequence screening therefore acts as a chokepoint between computational design and physical execution. Customer screening adds another layer by asking whether the purchaser is legitimate and whether the requested material fits a credible research or commercial purpose.

The system is imperfect. Screening databases need frequent updates. Novel sequences can be difficult to classify. Providers operate across jurisdictions with different regulations. Benchtop synthesis may eventually reduce dependence on centralized vendors. These challenges argue for stronger screening infrastructure, privacy-preserving methods and better international coordination rather than abandoning screening.

Why Governance Has to Sit Inside the Stack

AI biology cannot be governed only at the model layer because risk emerges across the full workflow. A language model may provide biological information. A protein model may generate candidate sequences. A synthesis provider turns sequence information into material. A laboratory amplifies, expresses or tests that material. A biofoundry can automate large numbers of experiments. Each layer creates different safeguards and different failure modes.

The strongest governance architecture therefore looks more like defense in depth than a single content filter. Model providers can restrict dangerous assistance. Synthesis companies can screen customers and sequence orders. Laboratories can maintain biosafety and access controls. Funding agencies can impose compliance conditions. Institutions can require review for high-risk work. Audits and logging can create accountability. None of those measures is sufficient alone.

This also protects legitimate science. A poorly designed safety system that blocks broad categories of harmless research can slow therapeutics, agriculture and industrial biotechnology without meaningfully reducing risk. The objective should be targeted friction at points where harmful capability becomes materially easier to realize.

What Is Actually Mature?

AI-enabled biology now spans technologies at very different maturity levels. Protein structure prediction is already research infrastructure. Generative protein design is experimentally productive but still uneven across functions. Automated biofoundries are real and useful in bounded workflows. Fully autonomous discovery across open-ended biology remains developmental. Large-scale industrial biomanufacturing remains governed as much by process engineering and economics as by AI.

The SURVXCOM Programmable Biology Test

The useful measure of AI biology is not how impressive a generated sequence looks on a screen. It is how many independent biological and industrial gates the design survives. SURVXCOM therefore evaluates an AI-enabled biological system across twelve layers.

What to Watch Next

Experimental hit rates in de novo protein design. Watch whether generative systems continue increasing the fraction of physically tested candidates that fold and function as intended. The key metric is not generation volume but validated success per experiment.

Complex function rather than simple binding. Binder design is advancing rapidly. Catalysis, conformational switches and molecular machines are harder. Progress there would indicate that generative biology is moving from shape design toward functional molecular engineering.

AlphaFold Database evolution. The 2026 addition of community datasets and millions of protein complexes suggests that biological AI infrastructure is becoming more collaborative and interaction-aware. Watch whether experimentally validated ensembles and dynamic structures become easier to integrate.

Self-driving biofoundries. DOE-supported work already demonstrates AI-plus-robotics enzyme optimization. Watch for broader closed-loop systems that can select experiments, execute them, analyze results and improve designs with limited manual intervention.

Biomanufacturing scale-up. Watch whether AI-designed molecules move beyond laboratory validation into pilot and commercial manufacturing without losing yield, stability or economics.

Data ownership. Proprietary high-quality assay data may become more strategically valuable than model architecture. Watch biotechnology companies build closed experimental data loops around their own automated labs.

Synthesis screening policy. The U.S. government is revising the nucleic-acid screening framework. Watch whether requirements become more enforceable, internationally compatible and capable of handling novel AI-designed sequences without blocking legitimate research.

Benchtop synthesis controls. As synthesis equipment becomes more distributed, the security architecture may need to move from provider-level screening toward device-level controls, auditability and verified access.

AI safety evaluations for biology. Watch model providers and government agencies improve biological capability evaluations so that risk assessments measure realistic uplift rather than simplistic knowledge tests.

Convergence with Article 024. The most consequential trajectory may be the merger of generative biology with autonomous laboratories. Once models, robotics and instrumentation operate in one closed loop, the cadence of biological research could accelerate substantially.

The New Biological Engineering Stack

The deepest change in biotechnology is not that computers have learned enough biology to replace biologists. It is that the boundary between digital design and physical experiment is becoming thinner. A protein can move from a computational hypothesis to a synthesized sequence to an automated assay more quickly than before, and the result of that assay can immediately shape the next computational design. The scientific method remains intact, but the cycle is becoming faster, more instrumented and more machine-readable.

This creates a new engineering stack. At the top are foundation models trained on sequences, structures and experimental data. Beneath them are design systems that generate molecules or prioritize candidates. Beneath those are DNA synthesis, cell engineering and protein-expression technologies. Biofoundries connect physical experimentation to robotics and laboratory information systems. Bioreactors and purification systems determine whether a successful laboratory result can become a product. Safety, regulation and screening surround the entire pipeline.

The economic consequences could be large because biology already manufactures things society values: medicines, enzymes, fuels, food ingredients, chemicals and materials. AI does not need to invent synthetic life to matter. If it reduces the time required to develop an industrial enzyme, improves a bioreactor strain, identifies a more stable therapeutic protein or helps scientists explore a larger design space with fewer failed experiments, the technology can create substantial value long before “programmable biology” becomes anything like software engineering.

The strategic consequences are equally important. Biological capability depends on data, synthesis capacity, laboratory automation, fermentation infrastructure, specialized talent and regulatory systems. Countries that control those layers will possess more than a biotechnology industry. They will control a platform for converting biological information into products. That makes synthetic biology part of the same sovereignty discussion as semiconductors, AI compute, energy and communications.

The guardrail is that biology remains physical. A model can generate a million candidates overnight. It cannot wish a protein into existence. A biofoundry can automate experiments. It cannot guarantee the assay measures the right thing. A laboratory can validate a promising molecule. It cannot guarantee a manufacturing process will scale. A synthesis company can screen an order. It cannot govern every biological risk in the world.

That is precisely why this technology belongs in the Critical Technology Stack. AI-enabled synthetic biology is not one breakthrough. It is a new systems architecture linking models, biological data, automated laboratories, synthesis, manufacturing and governance.

The future of programmable biology will be determined by how reliably that entire stack can turn digital biological ideas into validated, safe and manufacturable physical outcomes. The decisive advantage will belong not to the organization that generates the most candidates, but to the one that closes the loop between computation, experiment, manufacturing and responsible control most effectively.