Pinal AI Turns Plain Language Into Working Proteins

TL;DR: Researchers reported on 21 October 2026 in Nature that an AI system called Pinal can generate candidate proteins directly from plain-language descriptions of desired function, and lab tests confirmed working examples in all four tasks tried.

Type a function, get a protein candidate

A team led by Professor Fajie Yuan at Westlake University reported a 16-billion-parameter foundation model called Pinal that designs candidate proteins from natural-language descriptions of desired function, according to the Nature article titled Programmable design of functional proteins from natural language.

The paper, published on 21 October 2026, describes Pinal as a link between functional intent and protein sequence and structure. The model was trained on 1.7 billion synthetically annotated protein–text pairs, according to the same Nature article.

In computational evaluations, generated candidates combined high predicted foldability with functional-description alignment and sequence diversity, providing a basis for prioritizing experimentally testable designs, according to the Nature article. The authors then applied the system to four distinct laboratory design tasks and experimentally observed the intended function in each case.

The four tasks were a fluorescent protein, a polyethylene terephthalate hydrolase, an alcohol dehydrogenase and a metabolic H-protein, according to the Nature article. That set included catalytic activity for both designed enzymes. A Westlake University announcement titled Researchers Develop AI Model That Designs Proteins From Plain-Language Descriptions describes the same four tests as a green fluorescent protein, an enzyme that breaks down the plastic known as PET, an alcohol dehydrogenase, and a metabolic H-protein.

Without iterative experimental optimization, a Pinal-designed H-protein increased product titre by 1.7-fold relative to the corresponding native Escherichia coli H-protein in a multi-enzyme CO2 fixation pathway, according to the Nature article. The Westlake University announcement reports the same result as an increase in final product concentration to 1.7 times the level achieved with the natural protein.

Why protein design has needed templates and translations

Proteins perform much of the work that sustains living organisms, according to the Westlake University announcement. They accelerate chemical reactions, transport molecules and help cells build and repair themselves, according to the same announcement. Scientists have long sought to design new proteins for applications ranging from biological research to more sustainable manufacturing, according to the same announcement.

Having a computer design a protein remains a demanding task, according to the Westlake University announcement. Researchers often need to start with an existing protein or translate their goal into detailed instructions about molecular structures and amino acid sequences, according to the same announcement. Amino acid sequences are the chains of building blocks that make up proteins, according to the same announcement.

Turning a desired function into such specifications requires considerable expertise, according to the Westlake University announcement. Programming biological function — designing bespoke proteins to perform specified molecular tasks — is described in the Nature article as a foundational goal of molecular engineering.

Current design models remain fundamentally limited, according to the Nature article. They typically require either natural proteins as starting points for optimization or manual reformulation of functional goals as geometric and sequence-level constraints to guide candidate generation, according to the same Nature article.

Pinal aims to narrow the gap by allowing researchers to begin with a description of the function they want, according to the Westlake University announcement. The model generates candidate protein sequences for screening and experimental testing, reducing reliance on manually defined structural templates or detailed sequence constraints, according to the same announcement. The Nature article states the same conclusion in research terms: the findings support natural language as a high-level interface for candidate generation in protein design, enabling programmable exploration with reduced reliance on manually specified structural or sequence constraints.

How Pinal turns sentences into sequences

Pinal uses a two-stage design process and large-scale training data to learn the text-to-protein link, according to the figure description in the Nature article. The underlying approach first generates structures based on given language instructions, followed by designing sequences conditioned on both the structure and the language, according to the earlier bioRxiv preprint record summarized by Exa.

That ordering matters because it facilitates searching within the smaller structure space rather than the vast sequence space, according to the same preprint summary. The production Nature paper describes the trained outcome as a system that links functional intent to protein sequence and structure.

The resulting 16-billion-parameter model generated diverse protein sequences, according to the Westlake University announcement. Computational evaluations suggested these were consistent with the requested functions and likely to fold into well-defined structures, according to the same announcement. The Nature article frames that computational result as high predicted foldability with functional-description alignment and sequence diversity.

The model generates candidate protein sequences for screening and experimental testing, according to the Westlake University announcement. This gives researchers a more direct way to link a biological goal to possible molecular designs, according to the same announcement. Experimental screening and validation remain essential, according to the same announcement.

The listed authors begin with Fengyuan Dai, Shiyang You, Yudian Zhu, Yuan Gao and Lihao Fu, with additional contributors including Xibin Zhou, Jin Su, Chentong Wang, Yuliang Fan and Qingsong Tan, according to the Nature article. The corresponding authors include Tong Si, Jianming Liu, Hongyuan Lu and Fajie Yuan, according to supplementary author information for the Nature article. Affiliations include Westlake University, The Hong Kong University of Science and Technology (Guangzhou), the Shenzhen Institutes of Advanced Technology of the Chinese Academy of Sciences, Tianjin Institute of Pharmaceutical Research, Shanghai Zelixir Biotech Co., Ltd., and the University of Washington, according to the same supplementary information.

The 1.7-billion-pair training corpus

To build this capability, the team assembled data from multiple sources covering approximately 135 million distinct protein sequences, according to the Westlake University announcement. A smaller portion contained expert-reviewed functional annotations, while much of the collection used functional descriptions generated with models the team had previously developed, according to the same announcement.

By pairing individual proteins with multiple language descriptions, the researchers expanded the collection to approximately 1.7 billion pairs of proteins and text descriptions, according to the Westlake University announcement. These data helped Pinal learn relationships between functional descriptions and protein sequences and structures, according to the same announcement. The Nature article likewise reports training on 1.7 billion synthetically annotated protein–text pairs.

Protein sequences were obtained from UniProtKB/Swiss-Prot and UniProtKB/TrEMBL release 2023_03, according to the Data availability section of the Nature article. Predicted protein structures were obtained from the AlphaFold Protein Structure Database, according to the same section. Functional annotations and keywords were obtained from InterPro version 101.0, and enzyme classification information was obtained from the ENZYME nomenclature database, according to the same section.

Experimentally determined protein structures used for structural comparisons were obtained from the Protein Data Bank, according to the same Nature Data availability section. UniRef100 was used for sequence-similarity analyses, according to the same section. Fluorescent-protein reference information was obtained from FPbase, according to the same section.

Together, these results support natural language as a practical starting point for protein design, according to the Westlake University announcement. Researchers can describe a functional goal and explore candidate proteins without prescribing every molecular detail in advance, according to the same announcement.

What the computer predicted before any test tube

In computational evaluations, generated candidates combined high predicted foldability with functional-description alignment and sequence diversity, according to the Nature article. That combination provided a basis for prioritizing experimentally testable designs, according to the same article.

The Westlake University announcement describes the same stage as computational evaluations that suggested the generated diverse protein sequences were consistent with the requested functions and likely to fold into well-defined structures. The emphasis on foldability reflects a central requirement in protein work: candidate chains must be likely to adopt defined three-dimensional forms before they can be treated as serious test candidates.

The Nature paper presents scaling and benchmarking of design performance, generation of a diverse landscape of new proteins, and translation of diverse functional prompts into high-quality protein candidates in its figure set, according to the figure titles in the Nature article. Those computational steps came before laboratory work on four distinct functional classes.

The researchers then tested its designs in the laboratory across four distinct tasks, according to the Westlake University announcement. They observed the intended function in each case, including catalytic activity in both designed enzymes, according to the Nature article.

Four lab tests: glow, plastic, alcohol chemistry and carbon handling

The researchers tested Pinal designs in the laboratory across four distinct tasks: a green fluorescent protein, an enzyme that breaks down the plastic known as PET, an alcohol dehydrogenase, and a metabolic H-protein, according to the Westlake University announcement. From breaking down plastic to emitting green fluorescence, Pinal-designed proteins showed their intended functions across four experimentally tested tasks, according to the same announcement.

The Nature article identifies the same four as a fluorescent protein, a polyethylene terephthalate hydrolase, an alcohol dehydrogenase and a metabolic H-protein. The team experimentally observed the intended function in each case, including catalytic activity for both designed enzymes, according to the same Nature article.

A highlight was the H-protein, which the team tested in a pathway where several enzymes work together to convert carbon dioxide into useful compounds, according to the Westlake University announcement. That description corresponds to the multi-enzyme CO2 fixation pathway named in the Nature article.

Replacing the corresponding natural H-protein from Escherichia coli with a Pinal-designed version increased the final product concentration to 1.7 times the level achieved with the natural protein, according to the Westlake University announcement. The improvement required no repeated rounds of experimental optimization, according to the same announcement. The Nature article states the same outcome as a 1.7-fold increase in product titre relative to the corresponding native Escherichia coli H-protein without iterative experimental optimization.

The result demonstrates that a language-guided design can outperform its natural counterpart in a specific biological system, according to the Westlake University announcement. One Pinal-designed protein even outperformed its natural counterpart in a carbon dioxide fixation pathway, according to the same announcement’s summary. The finding is presented as evidence that experimental screening of language-generated candidates can identify designs with useful properties, rather than as proof that designed proteins are generally superior.

Open code, open weights, withheld training pipeline

Pinal is open-sourced under the Apache 2.0 license, according to the Code availability section of the Nature article. The code repository is available at GitHub, according to the same section. The Pinal web server is located at denovo-pinal.com, according to the same section.

The pretrained Pinal weights can be downloaded from Hugging Face, according to the Data availability section of the Nature article. The authors also provide public access to inference code, evaluation code, model usage instructions, trained model weights and the web server, according to the Code availability section.

The release approach distinguishes public tools from full-scale training resources, according to the same Code availability section. The release of large-scale training resources for generative protein design requires careful consideration of potential dual-use biosecurity risks, according to the same section.

In developing Pinal, the large-scale structural component of the training corpus was derived from the AlphaFold database, whose released protein-structure collection does not include viral proteomes, according to the same section. Accordingly, Pinal was not trained or intended for viral protein design, according to the same section.

Unrestricted public release of the full pretraining corpus together with the complete training pipeline could lower the barrier for adapting similar systems to sensitive biological datasets, according to the same section. To mitigate this risk, these full-scale training resources were provided only under confidential access for editorial and peer review assessment, according to the same section.

Because they will not be publicly released in full, the authors instead provide public access to inference code, evaluation code, model usage instructions, trained model weights and the web server, according to the same section. That approach is described as in line with common release practices for large-scale protein foundation models such as AlphaFold, ESM3 and ESMFold, according to the same section.

Why this changes the starting point

What if researchers could design a protein simply by describing what they need it to do, the Westlake University announcement asks in introducing the work. Pinal could make it easier to investigate proteins for different biological tasks and identify designs with useful properties, according to the same announcement.

Experimental screening and validation remain essential, according to the same announcement. But Pinal could make it easier to investigate proteins for different biological tasks and identify designs with useful properties by reducing reliance on manually defined structural templates or detailed sequence constraints, according to the same announcement.

These findings support natural language as a high-level interface for candidate generation in protein design, according to the Nature article. The Westlake University announcement expresses the same point for non-specialists: researchers can describe a functional goal and explore candidate proteins without prescribing every molecular detail in advance.