Data availabilityProtein sequences were obtained from UniProtKB/Swiss-Prot and UniProtKB/TrEMBL (https://www.uniprot.org/, release 2023_03). Predicted protein structures were obtained from the AlphaFold Protein Structure Database (https://alphafold.ebi.ac.uk/). Functional annotations and keywords were obtained from InterPro (version 101.0; https://www.ebi.ac.uk/interpro/), and enzyme classification information was obtained from the ENZYME nomenclature database (https://enzyme.expasy.org/). Experimentally determined protein structures used for structural comparisons were obtained from the Protein Data Bank (PDB; https://www.rcsb.org/). UniRef100 (https://www.uniprot.org/help/uniref) was used for sequence-similarity analyses. Fluorescent-protein reference information was obtained from FPbase (https://www.fpbase.org/). The pretrained Pinal weights can be downloaded from https://huggingface.co/westlake-repl/Pinal. All the other data are available within the paper and the Supplementary Information. Source data are provided with this paper.Code availabilityPinal is open-sourced under the Apache 2.0 license. The code repository is available at GitHub (https://github.com/westlake-repl/Denovo-Pinal). The Pinal web server is located at http://www.denovo-pinal.com/. The SaProt-T/O server, corresponding to the second component of Pinal, is linked from the Pinal web server and can also be accessed directly at http://113.45.254.183:9527/. The release of large-scale training resources for generative protein design requires careful consideration of potential dual-use biosecurity risks. In developing Pinal, the large-scale structural component of the training corpus was derived from AFDB, whose released protein-structure collection does not include viral proteomes. Accordingly, Pinal was not trained or intended for viral protein design. Nevertheless, 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. To mitigate this risk, these full-scale training resources were provided only under confidential access for editorial and peer review assessment. Because they will not be publicly released in full, and to support transparency and reproducibility, we instead provide public access to the inference code, evaluation code, model usage instructions, trained model weights and the web server, in line with common release practices for large-scale protein foundation models such as AlphaFold69,74, ESM338 and ESMFold68. Additional information about the composition, sources, filtering strategy, and construction of the training corpus is provided in the Methods and Supplementary Information.ReferencesDauparas, J. et al. Robust deep learning–based protein sequence design using ProteinMPNN. Science 378, 49–56 (2022).Article ADS CAS PubMed PubMed Central Google Scholar Lu, H. et al. Machine learning-aided engineering of hydrolases for PET depolymerization. Nature 604, 662–667 (2022).Article ADS CAS PubMed Google Scholar Watson, J. L. et al. De novo design of protein structure and function with RFdiffusion. Nature 620, 1089–1100 (2023).Article ADS CAS PubMed PubMed Central Google Scholar Arnold, F. H. Directed evolution: bringing new chemistry to life. Angew. Chem. Int. Ed. 57, 4143–4148 (2017).Article Google Scholar Wang, Y. et al. Directed evolution: methodologies and applications. Chem. Rev. 121, 12384–12444 (2021).Article CAS PubMed Google Scholar Yang, J. et al. Active learning-assisted directed evolution. Nat. Commun 16, 714 (2025).Article ADS CAS PubMed PubMed Central Google Scholar Yang, K. K., Wu, Z. & Arnold, F. H. Machine-learning-guided directed evolution for protein engineering. Nat. Methods 16, 687–694 (2019).Article CAS PubMed Google Scholar Zhao, H., Giver, L., Shao, Z., Affholter, J. A. & Arnold, F. H. Molecular evolution by staggered extension process (StEP) in vitro recombination. Nat. Biotechnol. 16, 258–261 (1998).Article ADS CAS PubMed Google Scholar Stemmer, W. P. C. Rapid evolution of a protein in vitro by DNA shuffling. Nature 370, 389–391 (1994).Article ADS CAS PubMed Google Scholar Kuhlman, B. et al. Design of a novel globular protein fold with atomic-level accuracy. Science 302, 1364–1368 (2003).Article ADS CAS PubMed Google Scholar Huang, P.-S., Boyken, S. E. & Baker, D. The coming of age of de novo protein design. Nature 537, 320–327 (2016).Article ADS CAS PubMed Google Scholar Koga, N. et al. Principles for designing ideal protein structures. Nature 491, 222–227 (2012).Article ADS CAS PubMed PubMed Central Google Scholar Fleishman, S. J. et al. Computational design of proteins targeting the conserved stem region of influenza hemagglutinin. Science 332, 816–821 (2011).Article ADS CAS PubMed PubMed Central Google Scholar King, N. P. et al. Computational design of self-assembling protein nanomaterials with atomic level accuracy. Science 336, 1171–1174 (2012).Article ADS CAS PubMed PubMed Central Google Scholar Leman, J. K. et al. Macromolecular modeling and design in Rosetta: recent methods and frameworks. Nat. Methods 17, 665–680 (2020).Article MathSciNet CAS PubMed PubMed Central Google Scholar Kaufmann, K. W., Lemmon, G. H., DeLuca, S. L., Sheehan, J. H. & Meiler, J. Practically useful: what the ROSETTA protein modeling suite can do for you. Biochemistry 49, 2987–2998 (2010).Article CAS PubMed PubMed Central Google Scholar Ferruz, N., Schmidt, S. & Höcker, B. ProtGPT2 is a deep unsupervised language model for protein design. Nat. Commun 13, 4348 (2022).Article ADS CAS PubMed PubMed Central Google Scholar Nijkamp, E., Ruffolo, J. A., Weinstein, E. N., Naik, N. & Madani, A. ProGen2: exploring the boundaries of protein language models. Cell Syst. 14, 968–978.e3 (2023).Article CAS PubMed Google Scholar Notin, P., Rollins, N., Gal, Y., Sander, C. & Marks, D. Machine learning for functional protein design. Nat. Biotechnol 42, 216–228 (2024).Article CAS PubMed PubMed Central Google Scholar Ahern, W. et al. Atom-level enzyme active site scaffolding using RFdiffusion2. Nat. Methods 23, 96–105 (2026).Article CAS PubMed Google Scholar Pacesa, M. et al. One-shot design of functional protein binders with BindCraft. Nature 646, 483–492 (2025).Article ADS CAS PubMed PubMed Central Google Scholar Vázquez Torres, S. et al. De novo designed proteins neutralize lethal snake venom toxins. Nature 639, 225–231 (2025).Article ADS PubMed PubMed Central Google Scholar Liu, Y. et al. De novo protein design with a denoising diffusion network independent of pretrained structure prediction models. Nat. Methods 21, 2107–2116 (2024).Article CAS PubMed Google Scholar Gainza, P. et al. De novo design of protein interactions with learned surface fingerprints. Nature 617, 176–184 (2023).Article ADS CAS PubMed PubMed Central Google Scholar Lin, Y. & AlQuraishi, M. Generating novel, designable, and diverse protein structures by equivariantly diffusing oriented residue clouds. In Proc. 40th Int. Conf. Mach. Learn. 202, 20978–21002 (PMLR, 2023).Huang, B. et al. A backbone-centred energy function of neural networks for protein design. Nature 602, 523–528 (2022).Article ADS CAS PubMed Google Scholar Madani, A. et al. Large language models generate functional protein sequences across diverse families. Nat. Biotechnol 41, 1099–1106 (2023).Article ADS CAS PubMed PubMed Central Google Scholar Ferruz, N. & Höcker, B. Controllable protein design with language models. Nat. Mach. Intell 4, 521–532 (2022).Article Google Scholar Cao, L. et al. Design of protein-binding proteins from the target structure alone. Nature 605, 551–560 (2022).Article ADS CAS PubMed PubMed Central Google Scholar Lauko, A. et al. Computational design of serine hydrolases. Science 388, eadu2454 (2025).Article CAS PubMed PubMed Central Google Scholar Romero, P. A. & Arnold, F. H. Exploring protein fitness landscapes by directed evolution. Nat. Rev. Mol. Cell Biol. 10, 866–876 (2009).Article CAS PubMed PubMed Central Google Scholar Pak, M. A. et al. Using AlphaFold to predict the impact of single mutations on protein stability and function. PLoS ONE 18, e0282689 (2023).Article CAS PubMed PubMed Central Google Scholar Truong, T. Jr. & Bepler, T. Understanding protein function with a multimodal retrieval-augmented foundation model. Adv. Neural Inf. Process. Syst. 38, 53167–53222 (2025).Google Scholar Illergård, K., Ardell, D. H. & Elofsson, A. Structure is three to ten times more conserved than sequence-a study of structural response in protein cores. Proteins 77, 499–508 (2009).Article PubMed Google Scholar van Kempen, M. et al. Fast and accurate protein structure search with Foldseek. Nat. Biotechnol. 42, 243–246 (2024).Su, J. et al. SaProt: protein language modeling with structure-aware vocabulary. In The 12th International Conference on Learning Representations (ICLR, 2024).Heinzinger, M. et al. Bilingual language model for protein sequence and structure. NAR Genom. Bioinform 6, lqae150 (2024).Article CAS PubMed PubMed Central Google Scholar Hayes, T. et al. Simulating 500 million years of evolution with a language model. Science 387, 850–858 (2025).Article ADS CAS PubMed Google Scholar Raffel, C. et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. 21, 1-67 (2020).MathSciNet Google Scholar Gu, Y. et al. Domain-specific language model pretraining for biomedical natural language processing. ACM Trans. Comput. Healthc 3, 1–23 (2021).Article CAS Google Scholar Kaplan, J. et al. Scaling laws for neural language models. Preprint at https://doi.org/10.48550/arXiv.2001.08361 (2020).Su, J. et al. A trimodal protein language model enables advanced protein searches. Nat. Biotechnol. 44, 1310–1316 (2026).Article PubMed Google Scholar Li, F.-Z., Amini, A. P., Yue, Y., Yang, K. K. & Lu, A. X. Feature reuse and scaling: understanding transfer learning with protein language models. In Proc. 41st Int. Conf. Mach. Learn. 235, 27351–27375 (PMLR, 2024).Fournier, Q. et al. Protein language models: is scaling necessary? Preprint at bioRxiv https://doi.org/10.1101/2024.09.23.614603 (2024).Notin, P. et al. Proteingym: Large-scale benchmarks for protein fitness prediction and design. Adv. Neural Inf. Process. Syst. 36, 64331–64379 (2023).Article Google Scholar Liu, S. et al. A text-guided protein design framework. Nat. Mach. Intell. 7, 580–591 (2025).Article Google Scholar Ingraham, J. B. et al. Illuminating protein space with a programmable generative model. Nature 623, 1070–1078 (2023).Article ADS CAS PubMed PubMed Central Google Scholar Kuang, J. et al. PDFBench: a benchmark for de novo protein design from function. In Proc. 43rd Int. Conf. Mach. Learn. 306, 61349–61371 (PMLR, 2026).Yuan, C. et al. Annotation-guided protein design with multi-level domain alignment. In Proc. 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 1855–1866 https://doi.org/10.1145/3690624.3709199 (2025).Sarkisyan, K. S. et al. Local fitness landscape of the green fluorescent protein. Nature 533, 397–401 (2016).Article ADS CAS PubMed PubMed Central Google Scholar Chen, J. et al. Global marine microbial diversity and its potential in bioprospecting. Nature 633, 371–379 (2024).Article ADS CAS PubMed PubMed Central Google Scholar Lu, H. et al. A secretion-based dual fluorescence assay for high-throughput screening of alcohol dehydrogenases. Biotechnol. Bioeng. 118, 1605–1616 (2021).Article Google Scholar Cohen-Addad, C., Pares, S., Sieker, L., Neuburger, M. & Douce, R. The lipoamide arm in the glycine decarboxylase complex is not freely swinging. Nat. Struct. Mol. Biol. 2, 63–68 (1995).Article CAS Google Scholar Kikuchi, G., Motokawa, Y., Yoshida, T. & Hiraga, K. Glycine cleavage system: reaction mechanism, physiological significance, and hyperglycinemia. Proc. Jpn. Acad. Ser. B 84, 246–263 (2008).Article ADS CAS Google Scholar Radford, A. et al. Language models are unsupervised multitask learners. OpenAI blog 1, 9 (2019).Google Scholar Lv, L. et al. ProLLaMA: a protein large language model for multitask protein language processing. IEEE Trans. Artif. Intell 7, 642–653 (2025).Article Google Scholar Guo, X.-Y., Li, Y.-F., Liu, Y., Pan, X. & Shen, H.-B. Ab-initio amino acid sequence design from protein text description with ProtDAT. Nat. Commun 16, 10544 (2025).Article ADS CAS PubMed PubMed Central Google Scholar Liu, N. et al. Protein design with dynamic protein vocabulary. Adv. Neural Inf. Process. Syst. 38, 22638–22669 (2025).Article Google Scholar Ahmad, S. et al. The UniProt website API: facilitating programmatic access to protein knowledge. Nucleic Acids Res. 53, W547–W553 (2025).Article PubMed PubMed Central Google Scholar Bairoch, A. The ENZYME database in 2000. Nucleic Acids Res. 28, 304–305 (2000).Article CAS PubMed PubMed Central Google Scholar Team GLM. ChatGLM: a family of large language models from GLM-130B to GLM-4 all tools. Preprint at https://doi.org/10.48550/arXiv.2406.12793 (2024).Zhou, X. et al. Decoding the Molecular Language of Proteins with Evolla. Preprint at bioRxiv https://doi.org/10.1101/2025.01.05.630192 (2025).Chen, T., Xu, B., Zhang, C. & Guestrin, C. Training deep nets with sublinear memory cost. Preprint at https://doi.org/10.48550/arXiv.1604.06174 (2016).Rajbhandari, S., Rasley, J., Ruwase, O. & He, Y. ZeRO: memory optimizations toward training trillion parameter models. In SC20 International Conference for High Performance Computing, Networking, Storage and Analysis https://doi.org/10.1109/SC41405.2020.00024 (2020).Barrio-Hernandez, I. et al. Clustering predicted structures at the scale of the known protein universe. Nature 622, 637–645 (2023).Article ADS CAS PubMed PubMed Central Google Scholar Yang, G. et al. Tuning large neural networks via zero-shot hyperparameter transfer. in Proc. Adv. Neural Inf. Process. Syst. Vol. 34 (NeurIPS, 2021).Verkuil, R. et al. Language models generalize beyond natural proteins. Preprint at bioRxiv https://doi.org/10.1101/2022.12.21.521521 (2022).Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123–1130 (2023).Article ADS MathSciNet CAS PubMed Google Scholar Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).Article ADS CAS PubMed PubMed Central Google Scholar Rector-Brooks, J. et al. General multimodal protein design enables DNA-encoding of chemistry. Preprint at https://doi.org/10.48550/arXiv.2604.05181 (2026).Lambert, T. J. FPbase: a community-editable fluorescent protein database. Nat. Methods 16, 277–278 (2019).Article CAS PubMed Google Scholar Shcherbakova, D. M. & Verkhusha, V. V. Chromophore chemistry of fluorescent proteins controlled by light. Curr. Opin. Chem. Biol. 20, 60–68 (2014).Article CAS PubMed PubMed Central Google Scholar Pongmuksuwan, P. & Kitisatorn, W. Recycled poly(ethylene terephthalate) for membrane material in membrane bioreactor. Mater. Today Proc. 77, 1154–1159 https://doi.org/10.1016/j.matpr.2023.01.088 (2023).Article CAS Google Scholar Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).Article ADS CAS PubMed PubMed Central Google Scholar Download referencesAcknowledgementsWe thank N. Li and the Westlake HPC Center for providing computing resources and technical support. We also thank A. Gitter, H. Liu and S. Ovchinnikov for their suggestions, which helped improve the manuscript. We acknowledge the support of the Biosciences Central Research Facility (BioCRF (GZ)) and the Bioresearch Lab at the Hong Kong University of Science and Technology (Guangzhou). Their assistance with equipment and facility access, experimental procedures and material characterization was essential to this study. Finally, we thank Y. Mao for assistance with the H-protein-R2 validation experiments.FundingThis work was supported by the National Natural Science Foundation of China (grant nos. 62576286 and 32471547), the National Key Research and Development Program of China (grant no. 2022ZD0115100), the Guangdong Education Department (grant no. 2023ZDZX2073), the Guangzhou-HKUST(GZ) Joint Funding Program (grant no. 2025A03J3791), the Nansha Key Science and Technology Project (no. 2023ZD015), and the Zhejiang Province Leading Geese Plan (grant no. 2025C01094), the Guangdong S&T Program (2024B1111140001) and the Shenzhen Science and Technology Program (RCJC20221008092810021).Author informationAuthor notesThese authors contributed equally: Fengyuan Dai, Shiyang YouAuthors and AffiliationsSchool of Engineering, Westlake University, Hangzhou, ChinaFengyuan Dai, Yudian Zhu, Xibin Zhou, Jin Su, Yuliang Fan, Qingsong Tan, Hui Qian, Zhe Zhang, Chenchen Han, Yajie Wang, Anping Zeng, Chunjun Zhan, Jianming Liu & Fajie YuanZhejiang University, Hangzhou, ChinaFengyuan DaiBioscience and Biomedical Engineering Thrust, Systems Hub, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, ChinaShiyang You, Xiaoxiao Ma, Xianjun Deng, Letong Yu & Hongyuan LuState Key Laboratory of Quantitative Synthetic Biology, Shenzhen Institute of Synthetic Biology, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, ChinaYuan Gao, Lihao Fu & Tong SiSchool of Life Sciences, Westlake University, Hangzhou, ChinaChentong Wang, Yan He, Yitao Ke, Xing Chang & Longxing CaoShanghai Zelixir Biotech, Shanghai, ChinaLiangzhen Zheng & Sheng WangState Key Laboratory of Druggability Evaluation and Systematic Translational Medicine, Tianjin Institute of Pharmaceutical Research, Tianjin, ChinaYuan Ji, Kexin Hao & Jing ZhaoZhejiang Key Laboratory of Low-Carbon Intelligent Synthetic Biology, School of Engineering, Westlake University, Hangzhou, ChinaAnping Zeng, Jianming Liu & Fajie YuanDepartment of Biochemistry, University of Washington, Seattle, WA, USAShunzhi WangAuthorsFengyuan DaiView author publicationsSearch author on:PubMed Google ScholarShiyang YouView author publicationsSearch author on:PubMed Google ScholarYudian ZhuView author publicationsSearch author on:PubMed Google ScholarYuan GaoView author publicationsSearch author on:PubMed Google ScholarLihao FuView author publicationsSearch author on:PubMed Google ScholarXibin ZhouView author publicationsSearch author on:PubMed Google ScholarJin SuView author publicationsSearch author on:PubMed Google ScholarChentong WangView author publicationsSearch author on:PubMed Google ScholarYuliang FanView author publicationsSearch author on:PubMed Google ScholarQingsong TanView author publicationsSearch author on:PubMed Google ScholarXiaoxiao MaView author publicationsSearch author on:PubMed Google ScholarXianjun DengView author publicationsSearch author on:PubMed Google ScholarLetong YuView author publicationsSearch author on:PubMed Google ScholarHui QianView author publicationsSearch author on:PubMed Google ScholarYan HeView author publicationsSearch author on:PubMed Google ScholarYitao KeView author publicationsSearch author on:PubMed Google ScholarZhe ZhangView author publicationsSearch author on:PubMed Google ScholarChenchen HanView author publicationsSearch author on:PubMed Google ScholarXing ChangView author publicationsSearch author on:PubMed Google ScholarLiangzhen ZhengView author publicationsSearch author on:PubMed Google ScholarSheng WangView author publicationsSearch author on:PubMed Google ScholarYuan JiView author publicationsSearch author on:PubMed Google ScholarKexin HaoView author publicationsSearch author on:PubMed Google ScholarLongxing CaoView author publicationsSearch author on:PubMed Google ScholarYajie WangView author publicationsSearch author on:PubMed Google ScholarAnping ZengView author publicationsSearch author on:PubMed Google ScholarShunzhi WangView author publicationsSearch author on:PubMed Google ScholarChunjun ZhanView author publicationsSearch author on:PubMed Google ScholarJing ZhaoView author publicationsSearch author on:PubMed Google ScholarTong SiView author publicationsSearch author on:PubMed Google ScholarJianming LiuView author publicationsSearch author on:PubMed Google ScholarHongyuan LuView author publicationsSearch author on:PubMed Google ScholarFajie YuanView author publicationsSearch author on:PubMed Google ScholarContributionsF.Y., F.D. and H.L. conceived the study. F.Y. led the overall project design and computational framework development. H.L., J.L. and T.S. led and supervised the biological experiments. F.D. and S.Y. carried out the main research. F.Y. and F.D. designed the Pinal framework and data-scaling strategy. F.D. implemented and trained the Pinal model. Y.F. and Q.T. performed some model evaluation. Shunzhi Wang, X.C. and Y.W. contributed to discussions and provided insightful feedback. Y.Z., A.Z. and J.L. contributed to the H-protein validation experiments. Y.G., L.F. and T.S. contributed to the GFP validation experiments. X.M., L.Y., S.Y., Z.Z., C.Z. and H.L. contributed to the PETase validation experiments. X.D., S.Y., Y.J., K.H., J.Z. and H.L. contributed to the ADH validation experiments. S.Y., H.Q., Y.K., C.W., Y.W. and L.C. performed the soluble protein expression experiments. Y.H., L.Z. and Sheng Wang contributed to early-stage exploratory experiments. X.Z., J.S. and C.H. collected and curated the datasets used in this study. F.Y., F.D., H.L., S.Y., J.L. and T.S. wrote the paper, and all authors read the paper.Corresponding authorsCorrespondence to Tong Si, Jianming Liu, Hongyuan Lu or Fajie Yuan.Ethics declarationsCompeting interestsThe authors declare no competing interests.Peer reviewPeer review informationNature thanks Yuanbin Wu and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.Additional informationPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Extended data figures and tablesExtended Data Fig. 1 Detailed architectures of the Pinal generative modules.a: T2struct translates natural language descriptions into structural (3Di) token sequences. A text encoder first generates the text representation according to the function prompt. This representation is then used to guide the T2struct decoder block, which autoregressively predicts the sequence of 3Di structural tokens. The detailed view (right) illustrates the internal composition of each decoder block (repeated N times), consisting of three sequential sub-modules: a structural context module (implemented via multi-head self-attention) to process the generated structural sequence, a text-conditioning module (employing multi-head cross-attention) to integrate guidance from the text representation, and a feature refinement network (a pointwise feed-forward layer) for further feature processing. Layer normalization is applied within each sub-module. b: The SaProt-T module, which designs an amino acid sequence based on both the structural input and the text prompt. A semantic projector maps the text representation into the same embedding space as the structural tokens. This projected text vector is prepended to the structural token embeddings to form a text-protein joint representation. This joint representation is then processed by the SaProt-T block, repeated N times, to predict the final amino acid sequence.Extended Data Fig. 2 Text conditioning is crucial for the sequence design module.a: Ablation study comparing the performance of SaProt-T and SaProt-O (760M parameters) against the original SaProt (650M), the ProstT5 baseline, and the scaled SaProt-T-T5xxl_encoder (~5B). Performance is evaluated based on textual alignment (x-axis) and predicted foldability (y-axis). The results show that our text-conditioned models (SaProt-T/O) outperform the original SaProt and the ProstT5 baseline, particularly in maintaining structural foldability. The size of each bubble corresponds to the number of model parameters, as indicated in the legend. b: Composition of the omni-modal training strategy for the SaProt-O model. This strategy involves a varied combination of masking applied to the amino acid (AA) sequence, the structural tokens, and the text prompt to train a flexible, multi-purpose design model. The left panel details the distribution of structural token masking (no, partial, or full masking) under two different AA masking conditions. The right panel shows that the text prompt is masked in 40% of the training examples.Extended Data Fig. 3 Determining the optimal number of sampling candidates (K).Performance of the top-ranked Pinal designs on the “hard” test set as a function of the number of sampled candidates (K). For each value of K, we generate K candidates, rank them by their joint log-probability, and then evaluate the top-ranked design. Left panel: textual alignment, measured by Max. and Mean Seq. ProTrek score. Right panel: predicted structural plausibility, measured by ESMFold pLDDT. Both textual alignment and foldability improve as K increases from 5 to 50. Further increasing K to 100 yields only marginal gains (or a slight decrease in ProTrek score) while substantially increasing computational cost. Based on this analysis, K = 50 is used as the standard setting for all experiments, offering a robust balance between design performance and inference efficiency.Extended Data Fig. 4 ProTrek score is a robust and discriminative metric for both text-sequence and text-structure alignment.The figure validates the ProTrek score’s performance using a test set of 1,000 protein-text pairs from Swiss-Prot. The top row (a-c) evaluates the text-to-sequence score (Seq. ProTrek), while the bottom row (d-f) evaluates the text-to-structure score (Struc. ProTrek). (a, d) Box plots show the distribution of ProTrek scores for proteins paired with their original descriptions, LLM-paraphrased descriptions, and randomly assigned (mismatched) descriptions. For both sequence and structure, the scores for original and paraphrased text are high and similar, while scores for random text are lower, demonstrating the metric’s discriminative power. Box plots show the median (centre line), 25th-75th percentiles (box), the most extreme values within 1.5 × the interquartile range (whiskers), and outliers (points). (b, e) Scatter plots comparing the scores of original versus paraphrased descriptions for each protein. The strong linear relationship, confirmed by a Spearman’s rank correlation coefficient (ρ) of > 0.98 for both sequence and structure, demonstrates that the metric is robust to variations in wording and captures core semantic meaning. (c, f) Scatter plots comparing the scores of original versus randomly assigned descriptions. The near-zero correlations (ρ ≈ 0) further confirm that the ProTrek score is a valid metric that reliably distinguishes between relevant and irrelevant protein-text pairings. The original ProTrek study also demonstrated that its score selectively responds to biologically relevant keywords while ignoring non-functional terms. Building on this capability, a recent benchmark paper48 adopted ProTrek as a key metric for evaluating the alignment between protein sequences and their textual descriptions.Extended Data Fig. 5 The three-phase training improves model performance and mitigates sequence repetition.a, Performance of the 15.5B T2struct model on the “hard” test set across three distinct training phases. Phase 1 (320k steps) represents initial training; Phase 2 (625k steps) incorporates data filtering by ProTrek score; Phase 3 (740k steps) is a final fine-tuning step on higher-quality data filtered to exclude low-pLDDT scores and high token repetition. The plots demonstrate that the third phase improves protein foldability, while incurring only a minor, acceptable trade-off in textual alignment (ProTrek score). b, Analysis of the sequence repetition phenomenon that motivated the third training phase. The left panel shows that the model after Phase 2 exhibited a high rate of amino acid repetition (≥5) (e.g.,“AAAAA”) at 60.2%, far exceeding that of natural proteins (UniProt, 4.0%). The final model after Phase 3 substantially reduces this rate to 35.3%. The right panel illustrates the effectiveness of the Phase 3 training data-filtering strategy. The high rate of Foldseek token repetition in the Phase 2 model’s outputs (91%) correlates with the high rate found in the AFDB training data (90%). The refined training strategy after Phase 3 successfully reduces the rate of token repetition greater than 10 to just 4%, thereby substantially resolving the Foldseek structure token repetition issue.Extended Data Fig. 6 Detailed performance analysis across diverse protein function descriptions.The figure evaluates Pinal’s design performance across 17 distinct functional categories curated from Swiss-Prot. Left panel: Analysis of textual alignment, measured by the maximum sequence ProTrek score (n = 50 functional prompts per subsection). The performance of Pinal’s top-ranked design (blue) is compared against a random baseline (light grey) and the ground-truth protein (dark grey) for each category. Pinal consistently outperforms the random baseline across all categories and approaches ground-truth performance, demonstrating robust language understanding. Right panel: Predicted structural plausibility of the corresponding Pinal-designed proteins, measured by ESMFold pLDDT (n = 250 Pinal-designed proteins per subsection). The results indicate that Pinal generates proteins with high predicted foldability (median pLDDT is typically > 80) across all functional description types, from concrete categories like “Protein names” to more abstract ones like “Disruption phenotype”. Box plots show the median (centre line), 25th-75th percentiles (box), the most extreme values within 1.5 × the interquartile range (whiskers), and outliers (points).Extended Data Fig. 7 Computational design, screening, and characterization of pinalGFP.a, Flowchart illustrating the multi-tiered computational filtering strategy applied to the initial Pinal-generated sequences. This pipeline culminated in the selection of 29 candidates for experimental validation. b, Multiple sequence alignment of pinalGFP, avGFP, and esmGFP-b8. The alignment highlights the substantial sequence divergence of pinalGFP from both the natural protein and the other synthetic variant. c, Normalized fluorescence excitation (dashed lines, λem = 510 nm) and emission (solid lines, λex = 460 nm) spectra of pinalGFP (blue) and avGFP (green). d, Size-exclusion chromatography profiles of purified GFP variants on a Superdex 200 column. Elution volumes reflected their oligomeric states in solution: avGFP was predominantly dimeric, whereas pinalGFP exhibited a more complex profile dominated by higher-order multimers.Source dataExtended Data Fig. 8 Computational screening pipeline and pNP standard curve for pinalPETase identification.a, Flowchart illustrating the multi-tiered screening strategy used to identify the active PETase candidate from the initial pool of Pinal-generated sequences. Candidates were first filtered on the basis of predicted structural quality (ESMFold pLDDT > 80, PAE < 7) and text–protein alignment (ProTrek score > 15), yielding 2,416 sequences. The top 94 candidates, ranked by ProTrek score, were then subjected to esterase activity screening using the p-nitrophenyl butyrate (pNPB) assay. This workflow led to the identification of a single standout variant, pinalPETase36, which was subsequently advanced to PET hydrolysis validation. The inset shows the predicted active-site architecture of pinalPETase36. b, Standard curve for p-nitrophenol (pNP) used for quantification of esterase activity in the pNPB assay. Absorbance at 410 nm was plotted against pNP concentration, showing a strong linear relationship (R2 = 0.9947). Data are shown as mean ± s.d. (n = 3 independent measurements).Source dataExtended Data Fig. 9 Computational design pipeline and analytical method for Pinal-designed ADHs.a, Flowchart of the computational screening and selection process for ADH candidates. From an initial library of 2,500 generated sequences, 1,639 candidates passing the feasibility filters based on structural confidence (ESMFold pLDDT > 80, PAE < 7) were identified, from which 1,000 sequences were randomly retained for downstream prioritization. From these, eight diverse candidates representing different ADH subfamilies were selected, leading to five successfully validated enzymes. b, UMAP of alcohol dehydrogenase (ADH) sequences. A total of 1,000 Pinal-designed ADH sequences (blue) were analysed together with natural ADH sequences from Swiss-Prot (gray; EC 1.1.1.1). Sequence embeddings were generated using ESM2-650M and projected into two dimensions by UMAP. The Pinal-designed sequences span a substantial region of the natural ADH sequence space. c, Representative HPLC chromatograms of derivatized acetaldehyde standards, corresponding to the product of the ADH-catalyzed reaction. Acetaldehyde standards at 0.1, 0.2, and 0.5 mM showed a consistent retention time of approximately 7.3 min.Source dataExtended Data Fig. 10 Computational design, screening, and mechanistic analysis of Pinal-designed H-proteins.a, Flowchart illustrating the computational design and screening pipeline for H-protein candidates. The initial generation was filtered to yield approximately 1,000 sequences that simultaneously satisfied the protein-text alignment criterion (ProTrek score > 15) and the sequence-divergence criterion (