The Biological Imagination

In partnership with

Dear reader,

Biodesign has long been limited by what biology already evolved. Materials like mycelium, algae, and bacterial cellulose were mostly explored through craft, observation, and slow experimentation, even as more rational bioengineering methods emerged in parallel.

Over the past five or so years, that landscape has shifted: not because experimentalists suddenly became better, but because modern AI systems now expose structure and pattern at scales that were previously inaccessible.

Below are three grounded ways AI is expanding what designers consider possible, anchored in real research rather than hype.

1. Making hidden structure legible

Biology is hard to design with because much of its structure is invisible at the scale of intuition. AlphaFold2, introduced in 2021, changed that by delivering highly accurate protein structure predictions for a vast range of sequences, and by making those predictions accessible through the AlphaFold Protein Structure Database. This did not replace experiments, but it did give scientists and designers a way to inspect predicted folds, motifs, and interfaces before stepping into the lab.​

AlphaFold DB now hosts well over 200 million predicted protein structures derived from UniProt-scale sequence sets, taking structural coverage from hundreds of thousands of experimental models to hundreds of millions of inspectable 3D forms.

By 2023–2024, many labs working on enzymes and binders routinely used AlphaFold2 predictions and confidence metrics (such as pLDDT and PAE maps) to decide which designs or constructs were worth testing experimentally. For designers, this is analogous to the way CAD made mechanical components legible: biological “shape” and potential interaction surfaces became something you can browse, filter, and reason about at unprecedented scale.​

A parallel at the material scale comes from mycelium composites. Studies have shown that mechanical and physical properties (such as compressive stiffness, density, thermal conductivity, and water absorption) depend systematically on substrate type, fibre condition, and processing.

These are precisely the kinds of structure–property relationships that AI and statistical models can help explore, not by replacing experimentation, but by narrowing which combinations are worth making. What was once opaque is becoming legible, and legibility expands imagination.​

2. Exploring “what if?” beyond natural precedent

AI now enables systematic exploration of biological variations that would have been slow or impractical to probe experimentally at scale. In protein design, this shift is especially visible in de novo generative methods.

ProteinMPNN provides deep-learning-based sequence design on specified protein backbones, while diffusion-based models such as RFdiffusion generate entirely new backbone geometries and scaffolds. Together with newer frameworks like Chroma, these tools allow researchers to propose new shapes, topologies, and functional surfaces that do not occur in natural proteins, and to evaluate them in silico before committing to experimental work.​

At the same time, updated versions of AlphaFold (such as AlphaFold3) are extending prediction beyond single proteins to complexes with DNA, RNA, and small molecules, enabling in silico exploration of many candidate interactions that would be difficult to screen solely by hand or traditional docking workflows.

For biodesigners, this means questions like “What if this protein folded differently?”, “What if this interface bound a different partner?”, or “What if this scaffold reinforced a living material?” can now be explored far more systematically and at much greater scale than was realistic five years ago. The important nuance is that de novo design existed before, but current AI systems make these explorations broader, faster, and more accessible.​

3. Bringing speculation closer to feasibility

Historically, speculative biodesign concepts often sat far from biological constraints; many ideas could not be checked without extensive, slow experimentation. AI now helps close that gap, not by guaranteeing outcomes, but by acting as an early plausibility filter.

In mycelium composites, for example, properties such as stiffness, density, and water absorption have been shown to correlate with measurable parameters like substrate type, fibre size, fungal species, and incubation or processing conditions.

Across materials science more broadly, machine-learning models (including random forests, neural networks, and convolutional architectures) are already used to learn structure–property relationships and to pre-filter candidate formulations or process windows before fabrication. For bio-based composites like mycelium, this suggests a near-term path where AI helps select promising regions of parameter space, while the actual materials are still validated through growth and testing.​

Bacterial cellulose (BC) offers a similar story. Experimental and review papers show that BC thickness, porosity, and mechanical performance respond in systematic ways to culture conditions such as oxygen availability, vessel geometry, medium composition, and strain engineering. While designer-facing AI tools specific to BC are still emerging, general modeling and ML approaches can, in principle, help explore how boundary conditions and geometries might influence growth and properties, guiding which “what if” samples to prioritize.​

Most strikingly, AI-designed binders and mini-proteins generated by workflows based on models like RFdiffusion have been expressed in the lab and shown to bind their targets with high affinity, including recent examples where AI-designed binders modulate biological pathways or improve genome editing efficiency. These pipelines do not eliminate experimental work, but they demonstrate a clear path from computational imagination to physical, functional molecules.​

For designers, this changes the nature of speculation. The goal is not perfect prediction, but triage: identifying which ideas are impossible, which are merely interesting, and which are plausibly worth a growth cycle or a set of assays. AI provides early signals that used to require full experimental loops to obtain.​

Why this shifts biodesign’s horizon

When new structures become visible, new ideas become plausible. When new variations can be explored at scale, new design spaces open up. When feasibility can be tested computationally before committing to lab work, imagination can expand without losing touch with constraints.

AI does not replace biodesign practice; it reconfigures it. The field is moving from a practice largely bounded by natural precedent and hand-tuned craft toward one animated by computational imagination, in which designers navigate vast spaces of possible forms, behaviours, and interactions, then bring a carefully chosen subset into the lab to see what biology will actually do.​

Until next time,

Raphael

Biodesign Academy

Stay up-to-date with AI

The Rundown is the most trusted AI newsletter in the world, with 1,000,000+ readers and exclusive interviews with AI leaders like Mark Zuckerberg, Demis Hassibis, Mustafa Suleyman, and more.

Their expert research team spends all day learning what’s new in AI and talking with industry experts, then distills the most important developments into one free email every morning.

Plus, complete the quiz after signing up and they’ll recommend the best AI tools, guides, and courses – tailored to your needs.

Sign up to start learning.