In 2003, David Baker's lab at the University of Washington published the computational design of Top7 — a novel 93-amino acid protein with a fold not found in nature. It took two years of computation and experimental validation. When the crystal structure was solved, it matched the design model within 1.2 angstroms root-mean-square deviation.

That was the proof of concept. It took another 20 years to turn it into a pipeline.

Today, the same process — design a protein on a computer, synthesize the gene, express it in bacteria, solve its structure — takes weeks, not years. The 2024 Nobel Prize in Chemistry, awarded to Baker alongside AlphaFold's Hassabis and Jumper, recognized that this transition from proof to pipeline is one of the most important advances in modern biology.

The Design Toolkit

The protein design revolution rests on three distinct breakthroughs, each addressing a different part of the design cycle.

AlphaFold 3 (2024) predicts the structure of any protein sequence, including complexes with DNA, RNA, and small molecules. It's a prediction tool, not a design tool — it tells you what a sequence will look like, but not what sequence to make.

RFdiffusion (2022) is the generative breakthrough. Based on diffusion models (the same technology behind DALL-E and Stable Diffusion), it generates novel protein backbones conditioned on target specifications — binding to a specific receptor, fitting into a specific pocket, forming a specific symmetric assembly. It doesn't generate sequences, only structures.

ProteinMPNN (2022) solves the inverse problem: given a protein backbone structure, find the amino acid sequence that will fold into it. It's the equivalent of finding the right notes for a melody you already composed.

Together, RFdiffusion + ProteinMPNN form a design pipeline: RFdiffusion generates a novel backbone geometry, ProteinMPNN finds a sequence that encodes it, and AlphaFold predicts whether the design will fold as intended. The cycle takes days.

ESM3 (2024) from EvolutionaryScale collapses this pipeline into a single model. At 98 billion parameters, it jointly models sequence, structure, and function — the largest biological AI model ever built. It can design functional proteins with no evolutionary precedent, as it demonstrated by creating a novel fluorescent protein with no sequence homology to known GFP variants.

What Has Been Achieved

The most concrete demonstration is in binder design — creating proteins that bind to specific therapeutic targets.

The number of AI-designed protein binders grew from two in 2019 to 350 in 2025. More importantly, the experimental success rate — the proportion of designs that actually bind their target when tested — went from 8% to 72%. This is the critical metric. When designs fail 92% of the time, the approach is a research tool. When they succeed 72% of the time, it's a product pipeline.

What Remains Unsolved

For all the progress, three fundamental problems remain.

Enzyme design. Designing a protein that catalyzes a specific chemical reaction — an enzyme — is qualitatively harder than designing a binder. Binders just need to fit. Enzymes need to fit, bind, orient, stabilize a transition state, and release a product. Success rates for computational enzyme design are below 10%, and the designed enzymes are typically millions of times less active than natural enzymes.

De novo function prediction. ESM3 can design a fluorescent protein, but it required a "thematic" prompt — the model was guided toward a known functional class. Designing a protein with a completely new function — a sensor for a molecule that no natural protein senses, a catalyst for a reaction that no natural enzyme performs — remains beyond current capabilities.

Dynamics and allostery. Current methods design static structures. But proteins are dynamic — they move, breathe, and change conformation. Allosteric regulation (a signal at one site affecting function at another) is critical for many therapeutic applications and is essentially not addressed by current design tools.

What This Means

The protein design revolution is real. It has already changed how academic labs and biotech companies approach binder discovery, antibody engineering, and vaccine design. Within 3-5 years, it will also change how enzyme engineering, biosensor design, and therapeutic protein optimization are done.

The limitation is no longer computational — it's experimental. The number of designs that can be tested is limited by DNA synthesis capacity, protein expression throughput, and binding assay bandwidth. The companies that will win in this space are the ones that close the design-build-test loop faster, not the ones with better AI models.

---

Data Sources: Baker lab publications (RFdiffusion, ProteinMPNN — Science 2022-2023), EvolutionaryScale ESM3 preprint (bioRxiv 2024), AlphaFold 3 paper (Nature 2024), Nobel Prize Committee press release (Chemistry 2024). Experimental success rates compiled from published studies by the Baker lab, Generate Biomedicines, and Arzeda.

About the Author: Martin DAVILA is a bioeconomy analyst and the founder of Bioinfometrics.