ESM3 Protein Generation Model: How Does AI "Program" Proteins?
Published on November 5, 2026
Fig. 1 | ESM3 models sequence, structure and function together, programming proteins like a language model
If AlphaFold let AI learn to "read" protein structure, ESM3 lets AI start to "compose" proteins. As the first generative model for biology that reasons jointly over protein sequence, structure and function, ESM3 turns biology into a language that can be programmed — you tell it the function, structure or sequence clue you want, and it generates proteins that have never existed in nature. In the most representative experiment, a chain-of-thought prompt built around just the chromophore residues led ESM3 to generate esmGFP, a new fluorescent protein sharing only ~58% identity with its closest natural counterpart — the equivalent of simulating roughly 500 million years of evolution in the lab — and it was verified to fluoresce in a real wet-lab experiment.
Background: From "Predicting Proteins" to "Generating Proteins"
Proteins are the executors of nearly every life function, and engineering them usually means a hard trade-off between "preserving function" and "changing properties." Traditional directed evolution mutates and screens round after round in the lab — slow and costly. AI models like AlphaFold excel at "prediction": given a sequence, predict its structure. What truly excites engineers is "generation" — not just "how does this sequence fold," but "give me a brand-new protein that performs a desired function."
ESM3 was created for exactly this. Released by EvolutionaryScale (a team largely from Meta/FAIR) in June 2024, it was trained with enormous GPU compute (the large model has ~98B parameters and ~1×10^24 FLOPS of training compute) on billions of proteins from diverse environments on Earth, spanning roughly 3.8 billion years of evolutionary history. It models sequence, structure and function as three "languages" simultaneously — the key difference from earlier protein models that specialized in a single dimension.
How ESM3 Works: Joint Reasoning over Sequence, Structure and Function
The core idea of ESM3 is to split a protein into three "programmable" modalities that share one representation:
• Sequence: the amino-acid letter string — a protein's "text";
• Structure: atomic coordinates and fold — a protein's "shape";
• Function: function tokens, Gene Ontology (GO) and other semantic labels — a protein's "purpose."
The model uses a bidirectional Transformer with geometric attention, encoding all three inputs into a shared latent representation, then uses masked generation to complete/generate the other parts from any modality. So ESM3 can generate a sequence from a structure, a structure from a function, or constrain multiple dimensions in a single prompt. The figure below illustrates this multi-modal mechanism.
Fig. 2 | ESM3's multi-modal mechanism: sequence, structure and function inputs flow into a shared representation to generate new proteins
The most tangible outcome of this architecture is esmGFP. The team prompted ESM3 with a chain of thought: keep the residues that form and catalyze the chromophore, plus the structure of part of the central alpha helix from a natural GFP. In two experiments, ESM3 found bright GFP candidates distant from any known GFP; the protein was synthesized and measured to fluoresce in *E. coli* lysate. It shares only ~58% identity with the closest natural fluorescent protein (from the sea anemone *Entacmaea quadricolor*) — a distance estimated at ~500 million years of natural evolution.
Capability Boundaries and Evaluation: How to Judge a Generative Model
ESM3 is powerful, but to use it well, you must first see its bounds:
• Generation is not correctness. The model outputs "candidates"; their foldability, activity and stability must be confirmed experimentally. Common evaluation metrics include similarity/novelty to known proteins, foldability scores, and wet-lab success rates.
• Prediction is not measurement. In any comparison, distinguish "model prediction" from "experimental measurement." esmGFP's fluorescence is a measured result; the "500-million-year span" is an estimate/inference — the two must not be conflated.
• Single-task vs. multi-modal. If you only need "sequence → structure," predictive models such as AlphaFold remain more mature. ESM3's value is taking function or structure constraints as prompts to do multi-constrained design, e.g. new fluorescent sensors, tailored enzymes, or even de novo scaffolds.
• Safety and control. As a frontier foundation model, ESM3 released a small, open-weight ESM3-open plus safety measures; the stronger the generation, the more prompt constraints and a wet-lab loop are needed to manage risk.
Conditions and Workflow: From Prompt to a "Design–Synthesize–Test–Learn" Loop
For real R&D, the model is only the starting point. A workable pipeline is a dry-wet closed loop: the AI generates candidate designs (Design) → wet-lab synthesis and expression (Synthesize) → measure the functional phenotype, e.g. fluorescence (Test/Assay) → feed results back to the model for preference optimization (Learn) → next round. Iterating this way turns "generation" into a usable product.
Fig. 3 | ESM3 plugged into a "design–synthesize–test–learn" dry-wet closed loop
Take esmGFP as the example: prompt with chromophore residues → ESM3 generates design candidates → synthesize and express in *E. coli* → measure fluorescence. It is precisely the real "measured" fluorescence feedback that confirmed the design as valid.
How MatwingsVenus® Empowers You: Connecting Generative Models to the Dry-Wet Loop
A generative model is only "paper candidates" without reliable wet-lab and engineering validation. What MatwingsVenus® focuses on is connecting frontier AI generative models with real experimental capability into one loop:
• [Wet Lab]: executing and optimizing sequence synthesis, protein expression, purification and activity assays, so ESM3-generated candidates can reach real measurement quickly.
• [Expert Center]: molecular and AI experts help deconstruct prompts, design evaluation metrics and interpret experimental data, cutting the trial-and-error of "not knowing how to use it or where its limits are."
• [Dry-Wet Closed-Loop Platform]: unified management of the design–synthesize–test–learn data flow, so every round of experiments flows back into the model and forms a continuous improvement flywheel.
For many teams, models like ESM3 provide a powerful "ideation engine," while MatwingsVenus® turns those ideas into verifiable, reproducible and scalable outcomes. Pairing the two is a more pragmatic way to work in the "computational design + experimental validation" era.
FAQ
1. What is ESM3, and how is it different from AlphaFold?
ESM3 is the first generative model to simultaneously reason over protein sequence, structure and function. AlphaFold is "predictive" — given a sequence, predict the structure. ESM3 is "generative" — from function/structure/sequence constraints, it can create proteins that do not exist in nature.
2. Was esmGFP really made, or is it just a simulation?
It really was synthesized and measured. The team synthesized the esmGFP designed by ESM3 and measured fluorescence in *E. coli* lysate. It shares ~58% identity with the closest natural protein; the "~500-million-year span" is an estimate/inference.
3. Can ESM3 be used directly for drug R&D or in the clinic?
Not directly. ESM3 produces candidate sequences/structures that still need folding prediction, wet-lab validation, safety assessment and much more follow-up. It is most valuable in early stages such as hit discovery, probe design and enzyme engineering.
4. Is ESM3 free to use?
ESM3 is offered in small/medium/large sizes via EvolutionaryScale's API; a small open-weight ESM3-open is released. Access terms and licensing follow the official documentation.
5. Are ESM3's results always correct?
No. Generated outputs are "candidates" and require foldability assessment and wet-lab validation. Evaluation should keep "model prediction" and "experimental measurement" as separate classes of evidence.
6. How does ESM3 relate to directed evolution?
They are complementary. Directed evolution screens progressively within a real biological system; ESM3 can generate sufficiently novel, multi-constrained candidates at once, greatly widening the "starting point," then directed evolution optimizes locally — a combination of "AI generation + experimental screening."
7. How can an ordinary research team start using such a model?
Start with prompts that specify clear function/structure constraints, paired with folding prediction and wet-lab validation. If you lack wet-lab or engineering capacity, mature dry-wet closed-loop services (e.g. MatwingsVenus®·Wet Lab) can turn generated results into measured data quickly.
Conclusion
ESM3 moves AI's understanding of proteins from "prediction" to "generation." It reads sequence, structure and function simultaneously, lets scientists program proteins with prompts, and proves the potential to "simulate 500 million years of evolution" with a new fluorescent protein, esmGFP. But the value of a generative model is ultimately cashed out by wet-lab experiments. Connecting an engine like ESM3 to a "design–synthesize–test–learn" dry-wet closed loop is the key to turning a frontier model into usable results — exactly the direction MatwingsVenus® and research teams are refining together.
One-Sentence Takeaway ESM3 moves protein engineering from "prediction" to "generation": it reads sequence, structure and function together and uses prompts to make proteins that do not exist in nature, but results must be validated by wet-lab experiments before they can be used in real applications.
Related Products & Services
• Dry-Wet Closed-Loop Platform
