Home
 /
Papers
 /
Bioengineering
 /
Toward De Novo Protein Design from Natural Language
Preprint
Open access
Top 1% · Validated by QED

Toward De Novo Protein Design from Natural Language

Fengyuan Dai, Shiyang You, Yudian Zhu +23 more
September 2025
 ·
bioRxiv
 ·
DOI 
10.1101/2024.08.01.606258
99
%
QED Score
QED Impressions

By demonstrating that natural language alone programs novel, functionally active proteins without structural templates or domain expertise, this work dismantles protein design's foundational dependency on existing molecules, making biological function directly addressable from conceptual description.

Run full QED analysis
Read paper
Save

QED claims analysis

Summary

A 16-billion-parameter generative foundation model trained on 1.7 billion synthetic protein-text pairs can translate natural language descriptions directly into diverse, structurally novel, and functionally active proteins—including catalytically active enzymes—by decomposing the language-to-sequence problem through an intermediate structural representation, with experimental validation demonstrating that all four designed proteins spanning distinct functional classes (fluorescent protein, PETase, alcohol dehydrogenase, and H-protein) exhibited activity and one designed enzyme surpassed its natural counterpart by 1.7-fold.

See full analysis
Main Claim 1 –

Structural intermediate bridges language-to-sequence protein design

Supporting evidence

  • Data and parameter scaling drive language-guided design performance
  • Larger models improve loss, foldability, and language alignment
Main Claim 2 –

Pinal outperforms prior methods in foldability and text alignment

Main Claim 3 –

Pinal designs novel protein folds that express as soluble proteins

Supporting evidence

  • Pinal expands sequence diversity beyond natural evolutionary clusters
Main Claim 4 –

Natural language prompts yield diverse, functional proteins via Pinal

Supporting evidence

  • Pinal designs catalytically active enzymes matching or exceeding natural performance
  • Single prompt yields functional ADHs across multiple subfamilies
  • Pinal-designed H-protein outperforms native within glycine cleavage system

Abstract

Programming biological function—designing bespoke proteins that execute specific tasks on demand—is a foundational goal of molecular engineering. Yet, current protein design paradigms remain fundamentally limited, typically requiring either an existing protein to evolve from, or deep, family-specific expertise to guide the design process. Here we introduce Pinal, a generative model that overcomes this barrier by translating functional descriptions in natural language directly into diverse and active proteins. This capability is built upon a 16-billion-parameter foundation model trained on an unprecedented synthetic corpus of 1.7 billion protein-text pairs, enabling it to ground functional language in the biophysical principles of protein structure. To provide definitive experimental validation, we tasked Pinal with designing four proteins from distinct functional classes: a fluorescent protein, a polyethylene terephthalate hydrolase, an alcohol dehydrogenase, and a metabolic H-protein. Remarkably, all four designs were functionally active and the two Pinal-designed enzymes achieved catalytic turnover for their respective reactions. Notably, the Pinal-designed H-protein even surpassed its natural counterpart, exhibiting 1.7-fold higher performance. Our results establish that natural language can serve as a programmable instruction set for biology, democratizing protein design and shifting the paradigm from the incremental modification of existing molecules to the direct creation of function from a conceptual description.

Related papers

How the QED Score works

QED Score is an AI-based quality metric for life science research. It rates each paper on two dimensions: originality (is the finding new and important?) and validity (does the evidence support the claims?). Author names and affiliations are removed before scoring. Each manuscript is broken down into its claims and the experiments behind them, and specialized AI agents then examine those claims for statistical rigor, consistency across figures, and contradictions with the existing literature. A verification layer re-checks every issue they flag. Scores are percentile ranks against all papers in the corpus. In a blinded study of papers where QED Score and journal rank disagreed, domain experts sided with QED Score in 75% of cases.
Read the methodology

Free access for academic researchers

Create your free QED account to validate your research, strengthen grant proposals, and uncover scientific insights.
We've sent you an access link.
Please check your inbox.

Didn't get your email? Check your spam folder or reach out to info@qedscience.com

Oops! Something went wrong while submitting the form.
Looking for QED for pharma, biotech or life science organizations?