A generalized protein design ML model enables generation of functional de novo proteins

A generalized protein design ML model enables generation of functional de novo proteins | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Brief Communication A generalized protein design ML model enables generation of functional de novo proteins Kathy Wei, Timothy Riley, Mohammad Parsa, Pourya Kalantari, Ismail Naderi, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6683338/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Despite significant advancements, the creation of functional proteins de novo remains a fundamental challenge. Although deep learning has revolutionized applications such as protein folding, a critical gap persists in integrating design objectives across structure and function. Here, we present MP4, a transformer-based AI model that generates novel sequences from functional text prompts, that enables the design of fully folded, functional proteins from minimal input specifications. Our approach demonstrates the ability to generate entirely novel proteins with high experimental success rates or effectively redesign existing proteins. This transformer-based model highlights the potential of generalist AI to address complex challenges in protein design, offering a versatile alternative to specialized approaches. Biological sciences/Computational biology and bioinformatics/Machine learning Biological sciences/Computational biology and bioinformatics/Protein design de novo protein design molecule programming text-to-protein protein AI model machine learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Protein function is determined by the interplay between sequence and structure, making it essential when designing new proteins to account for both aspects. Traditional methods, such as Rosetta 1 , employ empirical and physics-based approaches to link sequence with structure. More recently, deep learning based approaches, trained on extensive datasets, have demonstrated that large protein language models can learn sufficient information to accurately predict protein structures. Further advancements have shown that these deep-learning approaches can also capture some functional properties, such as protein-protein interactions and antibody complex structures 2 , 3 . Most protein language models are trained on highly curated datasets and are designed to predict relatively narrow functions. For instance, some models can predict protein structures with atomic-level accuracy given a specific sequence 4 . Others, like ProteinMPNN, focus on identifying sequences that will fold into a predefined backbone 5 . These models excel at tasks where the function is well defined, but they often require a large amount of a priori knowledge to generate meaningful results. While such approaches are highly effective for specific design goals, they limit the flexibility of these models in more generalist settings, where predicting novel protein functions or adapting to diverse design challenges is more complex. This restriction underscores the need for models that can handle broader design spaces, enabling de novo design of functional proteins across various applications. Here, we present the molecular programming model version 4 (MP4), which utilizes broad and diverse datasets to generate protein sequences from minimal input. Trained on 138,000 tokens and 3.2 billion unique data points, MP4 incorporates a comprehensive range of protein-related information to learn the complex relationships between sequence, structure, and function. Furthermore, the specific inclusion of text-based datapoints enables the model to interpret plain-language prompts of protein descriptions and design accordingly. To evaluate the models’ capabilities, we randomly generated thousands of unique protein descriptions that specified various functional characteristics, such as binding partners, catalytic activity, and subcellular localization. These descriptions were used to design novel sequences that were evaluated for stable structural folds and functional matches. A subset of these de novo designed proteins was then explored experimentally, with the majority stably expressing and exhibiting favorable thermodynamic properties. Thus, MP4 not only generates novel protein sequences, but also optimizes key functional and structural features, making it a powerful tool for protein design. Results Overview of the MP4 model MP4 is a transformer-based text-to-protein AI model designed to translate natural language prompts into de novo protein sequences that align with specified functions and properties. Unlike traditional methods that often follow a conventional pipeline - first defining a backbone structure and then generating sequences to match, MP4 utilizes a text-to-protein approach. This allows it to generate proteins directly from functional text prompts, making it more flexible and capable of addressing complex design objectives simultaneously. MP4 is designed to tackle some of the primary challenges in protein science, particularly the programmability of proteins - creating proteins that can perform specific functions. One of the key innovations in the MP4 model is the integration of conditional language models, such as the conditional transformer language, which allows the model to generate sequences based on specific annotated functions or properties 6 . Each protein sequence generated by the MP4 model undergoes evaluation for amino acid com-position, structural confidence, and functional similarity to ensure that the proteins are not only theoretically feasible but also practically functional. This method enables a joint sequence-function distribution, making it easier to tailor proteins for desired characteristics. MP4 generates protein sequences with high predicted foldability and functional activity To evaluate MP4’s ability to generate novel sequences from functional descriptions, we created over 1,000 prompts that specified diverse protein characteristics, including enzymatic activities, intracellular localizations, and binding partners. MP4 then generated diverse and unrelated se-quences based on these prompts (Fig 1), which were subsequently analyzed to assess their plausibility and realism. This evaluation focused on key metrics such as amino acid composition, predicted foldability, and alignment with known biochemical principles, providing insights into the model’s capacity to design biologically relevant proteins. The full repository can be explored at https://310.ai/mp/repo. We began by examining the amino acid composition of the de novo sequences generated by MP4, comparing their distributions to verified sequences from the non-redundant protein (NR) databases 7 . All amino acids were represented across the generated sequences, and their frequencies closely matched those observed in native UniProt sequences 8 (Fig 1A). MP4 ensures the natural-like distribution of amino acids in the generated sequences, with amino acid composition (AAcomp) scores ranging from 80 to 100 (Fig 1B). This metric identifies repetitive sequences, flagging potentially biologically implausible proteins. A defining feature of the MP4 model is its ability to generate de novo protein sequences that significantly differ from natural sequences. Sequence novelty is assessed using the seqdif score, which quantifies how distinct a generated protein sequence is from known reference in the NR database. Seqdif scores range from 0 to 100, with higher values indicating greater novelty. According to the observed seqdif score distribution, the majority of the generated sequences cluster in the 50-60 score range, signifying sequences at least 50% different from any natural sequence (Fig 1C). A smaller subset of proteins exhibits seqdif scores approaching 70-80 score, representing sequences that are highly divergent from natural proteins (Fig 1C), highlighting MP4’s capacity to explore novel sequence space while maintaining a balance between sequence novelty and biological feasibility. Next, we evaluated the structural stability of the generated sequences by predicting their folded structures using ESMFold 9 . For each sequence, we calculated the average predicted local distance difference test (pLDDT) as a measure of structural confidence 10 . Similar to the amino acid distribution, most sequences were predicted to fold into stable protein structures (Fig 1D), with an average pLDDT of 82.6, indicating high local confidence in the predicted folds. Structural similarity was further evaluated using FoldSeek and the reported TM-score to compare the generated structures to those in the Protein Data Bank, reported as structdif 11 (Fig 1E). Despite their sequence novelty, most generated proteins adopted folds that are well-established in nature, consistent with the principle that structure is often tightly linked to function. These findings demonstrate that MP4 not only interprets intended functional descriptions but also designs novel sequences that adopt the necessary structural folds to perform those functions. We evaluated how well the generated sequences aligned with their input prompts using ProtNLM, a UniProt-supported method that predicts protein functions from amino acid sequences 12 . Nlmsim, a ChatGPT-based similarity score, compares the input prompt with ProtNLM’s output (OpenAI (2024)). Scores of 80–100 indicate exact or subset matches, while 60–80 suggests similar words, though synonyms or broader categories may score poorly. Many sequences showed keyword matches in ProtNLM outputs (Fig 1F), highlighting MP4’s ability to translate functional descriptions into protein designs. Proteins generated by MP4 have desirable experimental properties To validate the experimental properties of the sequences generated by MP4, we characterized a subset of these designs to assess whether they possessed favorable traits beyond computational predictions. Specifically, we cloned a representative subset of 94 sequences, emphasizing those with stable predicted structural and diverse functional properties. This selected subset maintained sequence diversity (Fig 2A), highlighting that MP4 is not converging onto a single solution, nor replicating natural proteins. Each protein was expressed in a prokaryotic cell-free system using a split-GFP tag, and relative protein levels were quantified through a split-GFP assay 13 . Notably, a significant proportion of the cloned sequences successfully translated into measurable protein yields, with 79 out of 94 sequences (84%) yielding detectable protein levels (Fig 2B). Full results can be explored at https://310.ai/mp/lab/1. Thermostability, a key property of rationally designed proteins, was also assessed. This characteristic is defined by a protein’s ability to maintain structural integrity under increasing temperatures 14 . Material from each expression construct was subjected to differential scanning fluorimetry (DSF) to determine the melting temperature (Tm), representing the temperature at which 50% of the protein remains folded 15 . However, due to small expression volumes and low tryptophan content for fluorescent detection 16 , reliable signals were obtained from only 17 protein samples (Table 1). Despite this limitation (which could be overcome by prioritizing buried tryptophans during design), the average thermostability measurement exceeds 62°C, with the most stable proteins approaching 90°C (Fig 2C). In addition to characterizing these 17 by DSF, we selected an additional 10 samples to quantify by dynamic light scattering (DLS) 17 . These samples, although providing no measurable signal by DSF, resulted in a uniform peak by DLS implying that stable protein characteristics were incorporated throughout the design panel. These findings, although limited to a representative subset of the MP4-designed proteins, suggest that the MP4 model not only generates sequences with intended functional properties but also accounts for additional attributes such as expression efficiency and thermostability. The results highlight the model’s capacity to design proteins with a high likelihood of successful experimental translation and robust structural properties. Property interrogation of MP4 designed proteins We next assessed how well commonly used computational metrics predict protein behavior and expression levels. First, we examined the relationship between predicted secondary structure composition and expression levels. While no strong correlation was observed overall, MP4-designed proteins exhibited a broad range of alpha-helical content (20–90%, Fig 3A). Notably, well-expressing designs were found across this spectrum, including some with minimal alpha-helical content (traditionally considered difficult for computational design) and others composed almost entirely of alpha helices. However, the two designs with the highest predicted alpha-helical content failed to express, likely due to prediction biases from ESMFold. These findings indicate that MP4 does not impose a strong preference for specific protein folds and is capable of generating diverse, viable scaffolds. Hydrophobicity is another commonly used metric for ranking and evaluating protein designs, as it is often linked to increased aggregation, which can negatively impact both expression levels and thermostability. A hydrophobicity prediction model indicated that most MP4-designed proteins exhibited minimal hydrophobic content 18 . However, only weak correlation was observed between predicted hydrophobicity and measured expression levels (Fig 3B), reinforcing the notion that while hydrophobicity plays a role in protein behavior, it is not the sole determinant of successful folding and expression. Given the multifaceted nature of protein developability, we next evaluated a composite ‘developability’ predictor that integrates hydrophobicity, charge, and solubility into a weighted usability score, NetSolP 19 . Unlike hydrophobicity alone, MP4-designed proteins span a broad range of predicted usability, indicating that some sequences may lack optimal characteristics for experimental expression. Despite this variation, the usability score showed only a modest improvement over hydrophobicity in correlating with expression levels (Fig 3C), suggesting that even multi-parameter predictors struggle to fully capture the complexity of factors influencing experimental protein expression. MP4 designs for function Although the proteins described above were designed to maximize diversity across a wide range of protein families, we observed that a substantial subset (n = 32) were predicted to bind and catalyze adenosine triphosphate (ATP). ATP is a fundamental metabolic molecule, serving as a universal energy currency across all different cell types 20 . Due to its central role in cellular processes, ATP-binding proteins have been widely developed for diverse applications, including cancer diagnostics, biosensing, and targeted therapeutics 21,22 . Given the functional and translational relevance of this class, we selected this subset of MP4-designed proteins for deeper investigation and functional validation. The proteins in this ATP-associated subset were classified by ProtNLM as either ATP-binding cassette (ABC) transporters or adenylate kinases. Consistent with our broader findings, many of these sequences were successfully expressed in a cell free expression system, with 22 out of 32 (69%) yielding measurable expression levels. Furthermore, 11 of these proteins produced a detectable nanoDSF signal (with an average Tm of 64℃), which provided the basis for subsequent screening of ATP-binding activity. Full results can be explored at https://310.ai/mp/lab/2. To rapidly screen for ATP-binding activity, we repeated nanoDSF measurements in the presence of 2 mM ATP or AMP-PNP, a non-hydrolyzable analog of ATP. Binding of AMP-PNP is expected to stabilize the ATP-binding pocket, resulting in a measurable thermal shift. Indeed, we observed a 2–5 °C increase in melting temperature in 8 proteins, with responsive candidates identified from both the adenylate kinase (Fig 4A-C) and ABC transporter (Fig 4D-E) groups. Closer examination of the top-performing designs revealed notable divergence from native proteins. On average, MP4-generated sequences differed by 97 amino acids (55.3% identity) from their closest known structural analogs, indicating substantial novelty. Of particular interest were changes observed in the ATP-binding sites. While native adenylate kinases often conserve residues such as F35, present in 100% of aligned natural sequences and the majority of aligned natural structures (Fig. S9). MP4 introduced a rare variant seen in only a handful of kinases, F35L, into designs MS4BB and MBMLF (Fig. 4B,C). This substitution suggests that MP4 can infer viable structural solutions based on underlying biophysical principles rather than strict sequence homology. These findings demonstrate that MP4 is capable of generating functional proteins that meet core biochemical requirements from simple, natural-language prompts—even when departing from canonical evolutionary patterns. Discussion One of the key strengths of MP4 is its ability to generate protein sequences that can be translated into experimentally validated molecules. By optimizing multiple properties simultaneously, MP4 designs proteins that are structurally robust and stable under experimental conditions. This integrated approach highlights the potential of MP4 as a powerful tool to advance protein engineering and overcome practical challenges in the de novo protein design. This study demonstrates the capability of MP4 to generate protein sequences that exhibit desirable experimental properties, such as efficient expression and thermostability, while maintaining a high success rate in translation. The findings underscore the value of generalist protein design models, which consider a range of structural and functional properties simultaneously. By achieving measurable protein expression in 84% of the tested sequences and identifying several proteins with thermostability exceeding 65°C, MP4 highlights its potential as a versatile tool for rational protein design. The thermostability of the proteins designed by MP4 further underscores its utility for applications requiring robust protein performance under extreme conditions. Although only 17 proteins yielded reliable thermal melting curves due to a combination of low tryptophan content and technical constraints, the average melting temperature was 62°C, with the most stable protein nearing 90°C. These findings suggest that MP4 inherently considers stability as part of its design process. This is particularly significant for industrial and therapeutic applications, where proteins must remain functional under harsh environmental conditions. Unlike tools based on structure (AlphaFold-, RFdiffussion-, ProteinMPNN-based models) or protein language models dependent on sequence (like ESM-, ProGen-based models), our generative approach enables the design of novel proteins from a simple, a programmable description. MP4 is deliberately designed to translate natural language into novel and functional protein sequences, allowing for more diverse and unanticipated solutions by avoiding the constraints of predefined structures or starting sequences. Thus, MP4 provides a unique and flexible framework for de novo protein design, capable of generating functional biomolecules from intent-driven instructions. While the current vocabulary understood by MP4 is limited, future iterations will incorporate an expanded, precise, and technically sophisticated lexicon. This advancement will enable true molecular programming, where users can specify target protein properties—function, stability, binding affinity, and more—with fine-grained control. The model will then generate optimized protein sequences in a single inference step, transforming biological design into a deterministic, programmable process. Declarations Acknowledgements Experimental lab work was done at Adaptyv Bio, the cloud lab for proteins, Tierra Biosciences, and Arctoris Ltd. AMD Instinct Team for GPUs. N.B. for amazing graphic design. Competing Interests The authors are employees of 310 AI. Data Availability All data generated or analyzed during this study are included in this published article, its supplementary information files, and in the following url links: https://310.ai/mp/lab/1, https://310.ai/mp/lab/2/ Code Availability The code and model used in this study are proprietary to 310 AI and are not publicly available due to commercial confidentiality. References Leaver-Fay, A. et al. ROSETTA3: an object-oriented software suite for the simulation and design of macromolecules. Methods Enzym. 487 , 545–574 (2011). Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630 , 493–500 (2024). Wohlwend, J. et al. Boltz-1: Democratizing Biomolecular Interaction Modeling. (2024) doi:10.1101/2024.11.19.624167. Tunyasuvunakool, K. et al. Highly accurate protein structure prediction for the human proteome. Nature 596 , 590–596 (2021). Dauparas, J. et al. Robust deep learning–based protein sequence design using ProteinMPNN. Science 378 , 49–56 (2022). Keskar, N. S., McCann, B., Varshney, L., Xiong, C. & Socher, R. CTRL - A Conditional Transformer Language Model for Controllable Generation. ArXiv Prepr. ArXiv190905858 (2019). Boratyn, G. M. et al. BLAST: a more efficient report with usability improvements. Nucleic Acids Res. 41 , W29–W33 (2013). Bateman, A. et al. UniProt: the Universal Protein Knowledgebase in 2025. Nucleic Acids Res. 53 , D609–D617 (2024). Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379 , 1123–1130 (2023). Mariani, V., Biasini, M., Barbato, A. & Schwede, T. lDDT: a local superposition-free score for comparing protein structures and models using distance difference tests. Bioinformatics 29 , 2722–2728 (2013). van Kempen, M. et al. Fast and accurate protein structure search with Foldseek. Nat. Biotechnol. 42 , 243–246 (2023). Gane, A. et al. ProtNLM: Model-based Natural Language Protein Annotation. Preprint (2022). Bignon, C., Gruet, A. & Longhi, S. Split-GFP Reassembly Assay: Strengths and Caveats from a Multiparametric Analysis. Int. J. Mol. Sci. 23 , 13167 (2022). Vihinen, M. Relationship of protein flexibility to thermostability. “Protein Eng. Des. Sel. 1 , 477–480 (1987). Hellman, L. M. et al. Differential scanning fluorimetry based assessments of the thermal and kinetic stability of peptide–MHC complexes. J. Immunol. Methods 432 , 95–101 (2016). Wen, J., Lord, H., Knutson, N. & Wikström, M. Nano differential scanning fluorimetry for comparability studies of therapeutic proteins. Anal. Biochem. 593 , 113581 (2020). Stetefeld, J., McKenna, S. A. & Patel, T. R. Dynamic light scattering: a practical guide and applications in biomedical sciences. Biophys. Rev. 8 , 409–427 (2016). Malleshappa Gowder, S., Chatterjee, J., Chaudhuri, T. & Paul, K. Prediction and Analysis of Surface Hydrophobic Residues in Tertiary Structure of Proteins. Sci. World J. 2014 , 1–7 (2014). Thumuluri, V. et al. NetSolP: predicting protein solubility in Escherichia coli using language models. Bioinformatics 38 , 941–946 (2021). Dunn, J. & Grider, M. H. Physiology, Adenosine Triphosphate. in StatPearls (StatPearls Publishing, Treasure Island (FL), 2025). Wang, T., Ma, F. & Qian, H.-L. Defueling the cancer: ATP synthase as an emerging target in cancer therapy. Mol. Ther. Oncolytics 23 , 82–95 (2021). Yegutkin, G. G. & Boison, D. ATP and Adenosine Metabolism in Cancer: Exploitation for Therapeutic Gain. Pharmacol. Rev. 74 , 797–822 (2022). Table Table 1 is not available with this version. Additional Declarations Yes there is potential Competing Interest. Authors are employees at 310 AI. Supplementary Files nrreportingsummary.pdf reporting summary Editorialpolicychecklist.pdf editorial checklist NatureBiotechSupplemental2505211.docx A generalized protein design ML model enables generation of functional de novo proteins Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6683338","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Brief Communication","associatedPublications":[],"authors":[{"id":459879210,"identity":"fc13adbb-9a15-428b-8104-662724aa5b75","order_by":0,"name":"Kathy Wei","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAmklEQVRIiWNgGAWjYBACCSA+8OGHBA8DA2MD8VoOzuyxkCNNCzMPW4Ux8Q6TnJF88LANj0RiA/thIm2RlkhLOJxjAdTCk0ikFjnpHIPDOSBbGIjXkv/hsAUbUAv/Q2IdJp3DcJiBTcKYQYJYWyTnPzM42NsjIccmQawtEmcOP/7w40cdDz9/+gPitMABG4nqR8EoGAWjYBTgAwDupCjTxqtRZwAAAABJRU5ErkJggg==","orcid":"","institution":"310 AI","correspondingAuthor":true,"prefix":"","firstName":"Kathy","middleName":"","lastName":"Wei","suffix":""},{"id":459879211,"identity":"a66bcf1b-9c6d-4c6c-87a4-3d13daefa1cd","order_by":1,"name":"Timothy Riley","email":"","orcid":"","institution":"310 AI","correspondingAuthor":false,"prefix":"","firstName":"Timothy","middleName":"","lastName":"Riley","suffix":""},{"id":459879212,"identity":"7f66d1fe-d471-4afd-877f-5b89dff7d99d","order_by":2,"name":"Mohammad Parsa","email":"","orcid":"https://orcid.org/0000-0002-3416-739X","institution":"University of California, Berkeley","correspondingAuthor":false,"prefix":"","firstName":"Mohammad","middleName":"","lastName":"Parsa","suffix":""},{"id":459879213,"identity":"ebc637bc-9c33-43b4-9baa-bad7d9f8b317","order_by":3,"name":"Pourya Kalantari","email":"","orcid":"","institution":"310 AI","correspondingAuthor":false,"prefix":"","firstName":"Pourya","middleName":"","lastName":"Kalantari","suffix":""},{"id":459879214,"identity":"22d1d56f-4c2b-4f06-9dcd-8a780e7d59c6","order_by":4,"name":"Ismail Naderi","email":"","orcid":"","institution":"310 AI","correspondingAuthor":false,"prefix":"","firstName":"Ismail","middleName":"","lastName":"Naderi","suffix":""},{"id":459879215,"identity":"bc8df6ed-3a58-4fef-98a0-df303936beaa","order_by":5,"name":"Oleg Matusovsky","email":"","orcid":"","institution":"McGill University, Montreal","correspondingAuthor":false,"prefix":"","firstName":"Oleg","middleName":"","lastName":"Matusovsky","suffix":""},{"id":459879216,"identity":"33749621-f60b-45a7-8548-cadff876e224","order_by":6,"name":"Kooshiar Azimian","email":"","orcid":"","institution":"310 AI","correspondingAuthor":false,"prefix":"","firstName":"Kooshiar","middleName":"","lastName":"Azimian","suffix":""}],"badges":[],"createdAt":"2025-05-16 20:25:24","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6683338/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6683338/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":83532985,"identity":"cb346aff-cf77-44de-a0c3-d6a867494f8f","added_by":"auto","created_at":"2025-05-28 05:36:08","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":139732,"visible":true,"origin":"","legend":"\u003cp\u003eComputational metrics of 1000+ AI designed proteins generated by MP4 model. A) Amino acid composition (aacomp) per sequence, normalized to UniProt database proteins. B) Sequence comparison to NR/NT database proteins. C) Averaged ESMFold confidence pLDDT. D) Structure comparison to Protein Data Bank database proteins. E) Functional similarity based on prompt and predicted sequence function using ProtNLM model. F) Most common functions by ProtNLM prediction.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6683338/v1/a60e8c71b209c701d90732fc.png"},{"id":83532987,"identity":"0e4d9453-24fe-4583-a197-f85c9e41f586","added_by":"auto","created_at":"2025-05-28 05:36:08","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":517194,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental evaluation of 94 selected de novo designed proteins. A) Pairwise sequence similarity (upper right) and structure similarity (lower left) heatmap. B) Expression profile in a cell free expression system. C) Thermostability, measured by DSF, of 4 diverse MP4 designed proteins.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6683338/v1/fce1e90c942b1ddf87938334.png"},{"id":83532990,"identity":"5b4294f6-c531-4e93-879d-6893a5159d47","added_by":"auto","created_at":"2025-05-28 05:36:08","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":154335,"visible":true,"origin":"","legend":"\u003cp\u003eStructural and property analysis of designed proteins. A) Alpha-helical content vs relative expression levels. (B) Predicted hydrophobicity vs relative expression levels. (C) Predicted developability (usability score by NetSolP) vs relative expression levels.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6683338/v1/00f05063b86789e9e39b118b.png"},{"id":83532991,"identity":"c94ca079-6b75-45a2-8377-700a9081ee02","added_by":"auto","created_at":"2025-05-28 05:36:08","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":527173,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental ATP binding evaluation for adenylate kinases (top row) and ABC transporter binding proteins (bottom row). A) M17H6 vs PDB ID 3h86. TM-score = 0.947, mutations = 86. B) MS4BB vs PDB ID 2cdn. TM-score = 0.944, mutations = 90. C) MBMLF vs PDB ID 2oo7. TM-score = 0.973, mutations = 77. D) M2RXT vs PDB ID 2pcj. TM-score = 0.951, mutations = 177. E) MREGP vs PDB ID 4yms. TM-score = 0.979, mutations = 85. F) MB11S vs PDB ID 3c41. TM-score = 0.924, mutations = 127.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6683338/v1/62c7ba6afad91543d6d72b2a.png"},{"id":84858004,"identity":"e2c7a267-f1da-4ce0-bf43-a7fa3447b51a","added_by":"auto","created_at":"2025-06-18 06:28:29","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1651512,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6683338/v1/0ee75969-834d-4c37-8a07-2b08d0eae7f3.pdf"},{"id":83532988,"identity":"f9346e95-9495-4ed4-abc6-316eefa26079","added_by":"auto","created_at":"2025-05-28 05:36:08","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1664714,"visible":true,"origin":"","legend":"\u003cp\u003ereporting summary\u003c/p\u003e","description":"","filename":"nrreportingsummary.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6683338/v1/290cb0ff39f6e238c70efee5.pdf"},{"id":83533472,"identity":"d26d18eb-5485-4c28-adf3-53e693ee8be5","added_by":"auto","created_at":"2025-05-28 05:44:08","extension":"pdf","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":1682213,"visible":true,"origin":"","legend":"\u003cp\u003eeditorial checklist\u003c/p\u003e","description":"","filename":"Editorialpolicychecklist.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6683338/v1/6002f737a436479e547ef86e.pdf"},{"id":83532992,"identity":"4fe097a5-0c21-4310-9952-8753d857eaf8","added_by":"auto","created_at":"2025-05-28 05:36:08","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":2577647,"visible":true,"origin":"","legend":"\u003cp\u003eA generalized protein design ML model enables generation of functional de novo proteins\u003c/p\u003e","description":"","filename":"NatureBiotechSupplemental2505211.docx","url":"https://assets-eu.researchsquare.com/files/rs-6683338/v1/b93329cb2b797fbde9b72976.docx"}],"financialInterests":"\u003cb\u003eYes\u003c/b\u003e there is potential Competing Interest.\nAuthors are employees at 310 AI.","formattedTitle":"A generalized protein design ML model enables generation of functional de novo proteins","fulltext":[{"header":"Introduction","content":"\u003cp\u003eProtein function is determined by the interplay between sequence and structure, making it essential when designing new proteins to account for both aspects. Traditional methods, such as Rosetta\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e, employ empirical and physics-based approaches to link sequence with structure. More recently, deep learning based approaches, trained on extensive datasets, have demonstrated that large protein language models can learn sufficient information to accurately predict protein structures. Further advancements have shown that these deep-learning approaches can also capture some functional properties, such as protein-protein interactions and antibody complex structures\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eMost protein language models are trained on highly curated datasets and are designed to predict relatively narrow functions. For instance, some models can predict protein structures with atomic-level accuracy given a specific sequence\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Others, like ProteinMPNN, focus on identifying sequences that will fold into a predefined backbone\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. These models excel at tasks where the function is well defined, but they often require a large amount of a priori knowledge to generate meaningful results. While such approaches are highly effective for specific design goals, they limit the flexibility of these models in more generalist settings, where predicting novel protein functions or adapting to diverse design challenges is more complex. This restriction underscores the need for models that can handle broader design spaces, enabling de novo design of functional proteins across various applications.\u003c/p\u003e \u003cp\u003eHere, we present the molecular programming model version 4 (MP4), which utilizes broad and diverse datasets to generate protein sequences from minimal input. Trained on 138,000 tokens and 3.2\u0026nbsp;billion unique data points, MP4 incorporates a comprehensive range of protein-related information to learn the complex relationships between sequence, structure, and function. Furthermore, the specific inclusion of text-based datapoints enables the model to interpret plain-language prompts of protein descriptions and design accordingly. To evaluate the models\u0026rsquo; capabilities, we randomly generated thousands of unique protein descriptions that specified various functional characteristics, such as binding partners, catalytic activity, and subcellular localization. These descriptions were used to design novel sequences that were evaluated for stable structural folds and functional matches. A subset of these de novo designed proteins was then explored experimentally, with the majority stably expressing and exhibiting favorable thermodynamic properties. Thus, MP4 not only generates novel protein sequences, but also optimizes key functional and structural features, making it a powerful tool for protein design.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eOverview of the MP4 model\u003c/p\u003e\n\u003cp\u003eMP4 is a transformer-based text-to-protein AI model designed to translate natural language prompts into de novo protein sequences that align with specified functions and properties. Unlike traditional methods that often follow a conventional pipeline - first defining a backbone structure and then generating sequences to match, MP4 utilizes a text-to-protein approach. This allows it to generate proteins directly from functional text prompts, making it more flexible and capable of addressing complex design objectives simultaneously.\u003c/p\u003e\n\u003cp\u003eMP4 is designed to tackle some of the primary challenges in protein science, particularly the programmability of proteins - creating proteins that can perform specific functions. One of the key innovations in the MP4 model is the integration of conditional language models, such as the conditional transformer language, which allows the model to generate sequences based on specific annotated functions or properties\u003csup\u003e6\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eEach protein sequence generated by the MP4 model undergoes evaluation for amino acid com-position, structural confidence, and functional similarity to ensure that the proteins are not only theoretically feasible but also practically functional. This method enables a joint sequence-function distribution, making it easier to tailor proteins for desired characteristics.\u003c/p\u003e\n\u003cp\u003eMP4 generates protein sequences with high predicted foldability and functional activity\u003c/p\u003e\n\u003cp\u003eTo evaluate MP4\u0026rsquo;s ability to generate novel sequences from functional descriptions, we created over 1,000 prompts that specified diverse protein characteristics, including enzymatic activities, intracellular localizations, and binding partners. MP4 then generated diverse and unrelated se-quences based on these prompts (Fig 1), which were subsequently analyzed to assess their plausibility and realism. This evaluation focused on key metrics such as amino acid composition, predicted foldability, and alignment with known biochemical principles, providing insights into the model\u0026rsquo;s capacity to design biologically relevant proteins. The full repository can be explored at https://310.ai/mp/repo.\u003c/p\u003e\n\u003cp\u003eWe began by examining the amino acid composition of the de novo sequences generated by MP4, comparing their distributions to verified sequences from the non-redundant protein (NR) databases\u003csup\u003e7\u003c/sup\u003e. All amino acids were represented across the generated sequences, and their frequencies closely matched those observed in native UniProt sequences\u003csup\u003e8\u003c/sup\u003e (Fig 1A). MP4 ensures the natural-like distribution of amino acids in the generated sequences, with amino acid composition (AAcomp) scores ranging from 80 to 100 (Fig 1B). This metric identifies repetitive sequences, flagging potentially biologically implausible proteins.\u003c/p\u003e\n\u003cp\u003eA defining feature of the MP4 model is its ability to generate de novo protein sequences that significantly differ from natural sequences. Sequence novelty is assessed using the seqdif score, which quantifies how distinct a generated protein sequence is from known reference in the NR database. Seqdif scores range from 0 to 100, with higher values indicating greater novelty. According to the observed seqdif score distribution, the majority of the generated sequences cluster in the 50-60 score range, signifying sequences at least 50% different from any natural sequence (Fig 1C). A smaller subset of proteins exhibits seqdif scores approaching 70-80 score, representing sequences that are highly divergent from natural proteins (Fig 1C), highlighting MP4\u0026rsquo;s capacity to explore novel sequence space while maintaining a balance between sequence novelty and biological feasibility.\u003c/p\u003e\n\u003cp\u003eNext, we evaluated the structural stability of the generated sequences by predicting their folded\u003c/p\u003e\n\u003cp\u003estructures using ESMFold\u003csup\u003e9\u003c/sup\u003e. For each sequence, we calculated the average predicted local distance difference test (pLDDT) as a measure of structural confidence\u003csup\u003e10\u003c/sup\u003e. Similar to the amino acid distribution, most sequences were predicted to fold into stable protein structures (Fig 1D), with an average pLDDT of 82.6, indicating high local confidence in the predicted folds. Structural similarity was further evaluated using FoldSeek and the reported TM-score to compare the generated structures to those in the Protein Data Bank, reported as structdif\u003csup\u003e11\u003c/sup\u003e (Fig 1E). Despite their sequence novelty, most generated proteins adopted folds that are well-established in nature, consistent with the principle that structure is often tightly linked to function. These findings demonstrate that MP4 not only interprets intended functional descriptions but also designs novel sequences that adopt the necessary structural folds to perform those functions.\u003c/p\u003e\n\u003cp\u003eWe evaluated how well the generated sequences aligned with their input prompts using ProtNLM, a UniProt-supported method that predicts protein functions from amino acid sequences\u003csup\u003e12\u003c/sup\u003e. Nlmsim, a ChatGPT-based similarity score, compares the input prompt with ProtNLM\u0026rsquo;s output (OpenAI (2024)). Scores of 80\u0026ndash;100 indicate exact or subset matches, while 60\u0026ndash;80 suggests similar words, though synonyms or broader categories may score poorly. Many sequences showed keyword matches in ProtNLM outputs (Fig 1F), highlighting MP4\u0026rsquo;s ability to translate functional descriptions into protein designs.\u003c/p\u003e\n\u003cp\u003eProteins generated by MP4 have desirable experimental properties\u003c/p\u003e\n\u003cp\u003eTo validate the experimental properties of the sequences generated by MP4, we characterized a subset of these designs to assess whether they possessed favorable traits beyond computational predictions. Specifically, we cloned a representative subset of 94 sequences, emphasizing those with stable predicted structural and diverse functional properties. This selected subset maintained sequence diversity (Fig 2A), highlighting that MP4 is not converging onto a single solution, nor replicating natural proteins. Each protein was expressed in a prokaryotic cell-free system using a split-GFP tag, and relative protein levels were quantified through a split-GFP assay\u003csup\u003e13\u003c/sup\u003e. Notably, a significant proportion of the cloned sequences successfully translated into measurable protein yields, with 79 out of 94 sequences (84%) yielding detectable protein levels (Fig 2B). Full results can be explored at https://310.ai/mp/lab/1.\u003c/p\u003e\n\u003cp\u003eThermostability, a key property of rationally designed proteins, was also assessed. This characteristic is defined by a protein\u0026rsquo;s ability to maintain structural integrity under increasing temperatures\u003csup\u003e14\u003c/sup\u003e. Material from each expression construct was subjected to differential scanning fluorimetry (DSF) to determine the melting temperature (Tm), representing the temperature at which 50% of the protein remains folded\u003csup\u003e15\u003c/sup\u003e. However, due to small expression volumes and low tryptophan content for fluorescent detection\u003csup\u003e16\u003c/sup\u003e, reliable signals were obtained from only 17 protein samples (Table 1). Despite this limitation (which could be overcome by prioritizing buried tryptophans during design), the average thermostability measurement exceeds 62\u0026deg;C, with the most stable proteins approaching 90\u0026deg;C (Fig 2C). In addition to characterizing these 17 by DSF, we selected an additional 10 samples to quantify by dynamic light scattering (DLS)\u003csup\u003e17\u003c/sup\u003e. These samples, although providing no measurable signal by DSF, resulted in a uniform peak by DLS implying that stable protein characteristics were incorporated throughout the design panel.\u003c/p\u003e\n\u003cp\u003eThese findings, although limited to a representative subset of the MP4-designed proteins, suggest that the MP4 model not only generates sequences with intended functional properties but also accounts for additional attributes such as expression efficiency and thermostability. The results highlight the model\u0026rsquo;s capacity to design proteins with a high likelihood of successful experimental translation and robust structural properties.\u003c/p\u003e\n\u003cp\u003eProperty interrogation of MP4 designed proteins\u003c/p\u003e\n\u003cp\u003eWe next assessed how well commonly used computational metrics predict protein behavior and expression levels.\u003c/p\u003e\n\u003cp\u003eFirst, we examined the relationship between predicted secondary structure composition and expression levels. While no strong correlation was observed overall, MP4-designed proteins exhibited a broad range of alpha-helical content (20\u0026ndash;90%, Fig 3A). Notably, well-expressing designs were found across this spectrum, including some with minimal alpha-helical content (traditionally considered difficult for computational design) and others composed almost entirely of alpha helices. However, the two designs with the highest predicted alpha-helical content failed to express, likely due to prediction biases from ESMFold. These findings indicate that MP4 does not impose a strong preference for specific protein folds and is capable of generating diverse, viable scaffolds.\u003c/p\u003e\n\u003cp\u003eHydrophobicity is another commonly used metric for ranking and evaluating protein designs, as it is often linked to increased aggregation, which can negatively impact both expression levels and thermostability. A hydrophobicity prediction model indicated that most MP4-designed proteins exhibited minimal hydrophobic content\u003csup\u003e18\u003c/sup\u003e. However, only weak correlation was observed between predicted hydrophobicity and measured expression levels (Fig 3B), reinforcing the notion that while hydrophobicity plays a role in protein behavior, it is not the sole determinant of successful folding and expression.\u003c/p\u003e\n\u003cp\u003eGiven the multifaceted nature of protein developability, we next evaluated a composite \u0026lsquo;developability\u0026rsquo; predictor that integrates hydrophobicity, charge, and solubility into a weighted usability score, NetSolP\u003csup\u003e19\u003c/sup\u003e. Unlike hydrophobicity alone, MP4-designed proteins span a broad range of predicted usability, indicating that some sequences may lack optimal characteristics for experimental expression. Despite this variation, the usability score showed only a modest improvement over hydrophobicity in correlating with expression levels (Fig 3C), suggesting that even multi-parameter predictors struggle to fully capture the complexity of factors influencing experimental protein expression.\u003c/p\u003e\n\u003cp\u003eMP4 designs for function\u003c/p\u003e\n\u003cp\u003eAlthough the proteins described above were designed to maximize diversity across a wide range of protein families, we observed that a substantial subset (n = 32) were predicted to bind and catalyze adenosine triphosphate (ATP). ATP is a fundamental metabolic molecule, serving as a universal energy currency across all different cell types\u003csup\u003e20\u003c/sup\u003e. Due to its central role in cellular processes, ATP-binding proteins have been widely developed for diverse applications, including cancer diagnostics, biosensing, and targeted therapeutics\u003csup\u003e21,22\u003c/sup\u003e. Given the functional and translational relevance of this class, we selected this subset of MP4-designed proteins for deeper investigation and functional validation.\u003c/p\u003e\n\u003cp\u003eThe proteins in this ATP-associated subset were classified by ProtNLM as either ATP-binding cassette (ABC) transporters or adenylate kinases. Consistent with our broader findings, many of these sequences were successfully expressed in a cell free expression system, with 22 out of 32 (69%) yielding measurable expression levels. Furthermore, 11 of these proteins produced a detectable nanoDSF signal (with an average Tm of 64℃), which provided the basis for subsequent screening of ATP-binding activity. Full results can be explored at https://310.ai/mp/lab/2.\u003c/p\u003e\n\u003cp\u003eTo rapidly screen for ATP-binding activity, we repeated nanoDSF measurements in the presence of 2 mM ATP or AMP-PNP, a non-hydrolyzable analog of ATP. Binding of AMP-PNP is expected to stabilize the ATP-binding pocket, resulting in a measurable thermal shift. Indeed, we observed a 2\u0026ndash;5 \u0026deg;C increase in melting temperature in 8 proteins, with responsive candidates identified from both the adenylate kinase (Fig 4A-C) and ABC transporter (Fig 4D-E) groups.\u003c/p\u003e\n\u003cp\u003eCloser examination of the top-performing designs revealed notable divergence from native proteins. On average, MP4-generated sequences differed by 97 amino acids (55.3% identity) from their closest known structural analogs, indicating substantial novelty. Of particular interest were changes observed in the ATP-binding sites. While native adenylate kinases often conserve residues such as F35, present in 100% of aligned natural sequences and the majority of aligned natural structures (Fig. S9). MP4 introduced a rare variant seen in only a handful of kinases, F35L, into designs MS4BB and MBMLF (Fig. 4B,C). This substitution suggests that MP4 can infer viable structural solutions based on underlying biophysical principles rather than strict sequence homology. These findings demonstrate that MP4 is capable of generating functional proteins that meet core biochemical requirements from simple, natural-language prompts\u0026mdash;even when departing from canonical evolutionary patterns.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eOne of the key strengths of MP4 is its ability to generate protein sequences that can be translated into experimentally validated molecules. By optimizing multiple properties simultaneously, MP4 designs proteins that are structurally robust and stable under experimental conditions. This integrated approach highlights the potential of MP4 as a powerful tool to advance protein engineering and overcome practical challenges in the de novo protein design.\u003c/p\u003e \u003cp\u003eThis study demonstrates the capability of MP4 to generate protein sequences that exhibit desirable experimental properties, such as efficient expression and thermostability, while maintaining a high success rate in translation. The findings underscore the value of generalist protein design models, which consider a range of structural and functional properties simultaneously. By achieving measurable protein expression in 84% of the tested sequences and identifying several proteins with thermostability exceeding 65\u0026deg;C, MP4 highlights its potential as a versatile tool for rational protein design.\u003c/p\u003e \u003cp\u003eThe thermostability of the proteins designed by MP4 further underscores its utility for applications requiring robust protein performance under extreme conditions. Although only 17 proteins yielded reliable thermal melting curves due to a combination of low tryptophan content and technical constraints, the average melting temperature was 62\u0026deg;C, with the most stable protein nearing 90\u0026deg;C. These findings suggest that MP4 inherently considers stability as part of its design process. This is particularly significant for industrial and therapeutic applications, where proteins must remain functional under harsh environmental conditions.\u003c/p\u003e \u003cp\u003eUnlike tools based on structure (AlphaFold-, RFdiffussion-, ProteinMPNN-based models) or protein language models dependent on sequence (like ESM-, ProGen-based models), our generative approach enables the design of novel proteins from a simple, a programmable description. MP4 is deliberately designed to translate natural language into novel and functional protein sequences, allowing for more diverse and unanticipated solutions by avoiding the constraints of predefined structures or starting sequences. Thus, MP4 provides a unique and flexible framework for de novo protein design, capable of generating functional biomolecules from intent-driven instructions.\u003c/p\u003e \u003cp\u003eWhile the current vocabulary understood by MP4 is limited, future iterations will incorporate an expanded, precise, and technically sophisticated lexicon. This advancement will enable true molecular programming, where users can specify target protein properties\u0026mdash;function, stability, binding affinity, and more\u0026mdash;with fine-grained control. The model will then generate optimized protein sequences in a single inference step, transforming biological design into a deterministic, programmable process.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eExperimental lab work was done at Adaptyv Bio, the cloud lab for proteins, Tierra Biosciences, and Arctoris Ltd. AMD Instinct Team for GPUs. N.B. for amazing graphic design. \u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors are employees of 310 AI.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data generated or analyzed during this study are included in this published article, its supplementary information files, and in the following url links: https://310.ai/mp/lab/1, https://310.ai/mp/lab/2/\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe code and model used in this study are proprietary to 310 AI and are not publicly available due to commercial confidentiality.\u003cbr\u003e\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eLeaver-Fay, A. \u003cem\u003eet al.\u003c/em\u003e ROSETTA3: an object-oriented software suite for the simulation and design of macromolecules. \u003cem\u003eMethods Enzym. \u003c/em\u003e\u003cstrong\u003e487\u003c/strong\u003e, 545\u0026ndash;574 (2011).\u003c/li\u003e\n\u003cli\u003eAbramson, J. \u003cem\u003eet al.\u003c/em\u003e Accurate structure prediction of biomolecular interactions with AlphaFold 3. \u003cem\u003eNature \u003c/em\u003e\u003cstrong\u003e630\u003c/strong\u003e, 493\u0026ndash;500 (2024).\u003c/li\u003e\n\u003cli\u003eWohlwend, J. \u003cem\u003eet al.\u003c/em\u003e Boltz-1: Democratizing Biomolecular Interaction Modeling. (2024) doi:10.1101/2024.11.19.624167.\u003c/li\u003e\n\u003cli\u003eTunyasuvunakool, K. \u003cem\u003eet al.\u003c/em\u003e Highly accurate protein structure prediction for the human proteome. \u003cem\u003eNature \u003c/em\u003e\u003cstrong\u003e596\u003c/strong\u003e, 590\u0026ndash;596 (2021).\u003c/li\u003e\n\u003cli\u003eDauparas, J. \u003cem\u003eet al.\u003c/em\u003e Robust deep learning\u0026ndash;based protein sequence design using ProteinMPNN. \u003cem\u003eScience \u003c/em\u003e\u003cstrong\u003e378\u003c/strong\u003e, 49\u0026ndash;56 (2022).\u003c/li\u003e\n\u003cli\u003eKeskar, N. S., McCann, B., Varshney, L., Xiong, C. \u0026amp; Socher, R. CTRL - A Conditional Transformer Language Model for Controllable Generation. \u003cem\u003eArXiv Prepr. ArXiv190905858\u003c/em\u003e (2019).\u003c/li\u003e\n\u003cli\u003eBoratyn, G. M. \u003cem\u003eet al.\u003c/em\u003e BLAST: a more efficient report with usability improvements. \u003cem\u003eNucleic Acids Res. \u003c/em\u003e\u003cstrong\u003e41\u003c/strong\u003e, W29\u0026ndash;W33 (2013).\u003c/li\u003e\n\u003cli\u003eBateman, A. \u003cem\u003eet al.\u003c/em\u003e UniProt: the Universal Protein Knowledgebase in 2025. \u003cem\u003eNucleic Acids Res. \u003c/em\u003e\u003cstrong\u003e53\u003c/strong\u003e, D609\u0026ndash;D617 (2024).\u003c/li\u003e\n\u003cli\u003eLin, Z. \u003cem\u003eet al.\u003c/em\u003e Evolutionary-scale prediction of atomic-level protein structure with a language model. \u003cem\u003eScience \u003c/em\u003e\u003cstrong\u003e379\u003c/strong\u003e, 1123\u0026ndash;1130 (2023).\u003c/li\u003e\n\u003cli\u003eMariani, V., Biasini, M., Barbato, A. \u0026amp; Schwede, T. lDDT: a local superposition-free score for comparing protein structures and models using distance difference tests. \u003cem\u003eBioinformatics \u003c/em\u003e\u003cstrong\u003e29\u003c/strong\u003e, 2722\u0026ndash;2728 (2013).\u003c/li\u003e\n\u003cli\u003evan Kempen, M. \u003cem\u003eet al.\u003c/em\u003e Fast and accurate protein structure search with Foldseek. \u003cem\u003eNat. Biotechnol. \u003c/em\u003e\u003cstrong\u003e42\u003c/strong\u003e, 243\u0026ndash;246 (2023).\u003c/li\u003e\n\u003cli\u003eGane, A. \u003cem\u003eet al.\u003c/em\u003e ProtNLM: Model-based Natural Language Protein Annotation. \u003cem\u003ePreprint\u003c/em\u003e (2022).\u003c/li\u003e\n\u003cli\u003eBignon, C., Gruet, A. \u0026amp; Longhi, S. Split-GFP Reassembly Assay: Strengths and Caveats from a Multiparametric Analysis. \u003cem\u003eInt. J. Mol. Sci. \u003c/em\u003e\u003cstrong\u003e23\u003c/strong\u003e, 13167 (2022).\u003c/li\u003e\n\u003cli\u003eVihinen, M. Relationship of protein flexibility to thermostability. \u003cem\u003e\u0026ldquo;Protein Eng. Des. Sel. \u003c/em\u003e\u003cstrong\u003e1\u003c/strong\u003e, 477\u0026ndash;480 (1987).\u003c/li\u003e\n\u003cli\u003eHellman, L. M. \u003cem\u003eet al.\u003c/em\u003e Differential scanning fluorimetry based assessments of the thermal and kinetic stability of peptide\u0026ndash;MHC complexes. \u003cem\u003eJ. Immunol. Methods \u003c/em\u003e\u003cstrong\u003e432\u003c/strong\u003e, 95\u0026ndash;101 (2016).\u003c/li\u003e\n\u003cli\u003eWen, J., Lord, H., Knutson, N. \u0026amp; Wikstr\u0026ouml;m, M. Nano differential scanning fluorimetry for comparability studies of therapeutic proteins. \u003cem\u003eAnal. Biochem. \u003c/em\u003e\u003cstrong\u003e593\u003c/strong\u003e, 113581 (2020).\u003c/li\u003e\n\u003cli\u003eStetefeld, J., McKenna, S. A. \u0026amp; Patel, T. R. Dynamic light scattering: a practical guide and applications in biomedical sciences. \u003cem\u003eBiophys. Rev. \u003c/em\u003e\u003cstrong\u003e8\u003c/strong\u003e, 409\u0026ndash;427 (2016).\u003c/li\u003e\n\u003cli\u003eMalleshappa Gowder, S., Chatterjee, J., Chaudhuri, T. \u0026amp; Paul, K. Prediction and Analysis of Surface Hydrophobic Residues in Tertiary Structure of Proteins. \u003cem\u003eSci. World J. \u003c/em\u003e\u003cstrong\u003e2014\u003c/strong\u003e, 1\u0026ndash;7 (2014).\u003c/li\u003e\n\u003cli\u003eThumuluri, V. \u003cem\u003eet al.\u003c/em\u003e NetSolP: predicting protein solubility in Escherichia coli using language models. \u003cem\u003eBioinformatics \u003c/em\u003e\u003cstrong\u003e38\u003c/strong\u003e, 941\u0026ndash;946 (2021).\u003c/li\u003e\n\u003cli\u003eDunn, J. \u0026amp; Grider, M. H. Physiology, Adenosine Triphosphate. in \u003cem\u003eStatPearls\u003c/em\u003e (StatPearls Publishing, Treasure Island (FL), 2025).\u003c/li\u003e\n\u003cli\u003eWang, T., Ma, F. \u0026amp; Qian, H.-L. Defueling the cancer: ATP synthase as an emerging target in cancer therapy. \u003cem\u003eMol. Ther. Oncolytics \u003c/em\u003e\u003cstrong\u003e23\u003c/strong\u003e, 82\u0026ndash;95 (2021).\u003c/li\u003e\n\u003cli\u003eYegutkin, G. G. \u0026amp; Boison, D. ATP and Adenosine Metabolism in Cancer: Exploitation for Therapeutic Gain. \u003cem\u003ePharmacol. Rev. \u003c/em\u003e\u003cstrong\u003e74\u003c/strong\u003e, 797\u0026ndash;822 (2022).\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Table","content":"\u003cp\u003eTable 1 is not available with this version.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"de novo protein design, molecule programming, text-to-protein, protein AI model, machine learning","lastPublishedDoi":"10.21203/rs.3.rs-6683338/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6683338/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDespite significant advancements, the creation of functional proteins de novo remains a fundamental challenge. Although deep learning has revolutionized applications such as protein folding, a critical gap persists in integrating design objectives across structure and function. Here, we present MP4, a transformer-based AI model that generates novel sequences from functional text prompts, that enables the design of fully folded, functional proteins from minimal input specifications. Our approach demonstrates the ability to generate entirely novel proteins with high experimental success rates or effectively redesign existing proteins. This transformer-based model highlights the potential of generalist AI to address complex challenges in protein design, offering a versatile alternative to specialized approaches.\u003c/p\u003e","manuscriptTitle":"A generalized protein design ML model enables generation of functional de novo proteins","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-28 05:36:03","doi":"10.21203/rs.3.rs-6683338/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"2bde2911-cbbf-46e9-99b7-356eb2786446","owner":[],"postedDate":"May 28th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":48852131,"name":"Biological sciences/Computational biology and bioinformatics/Machine learning"},{"id":48852132,"name":"Biological sciences/Computational biology and bioinformatics/Protein design"}],"tags":[],"updatedAt":"2025-06-18T06:20:22+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-28 05:36:03","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6683338","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6683338","identity":"rs-6683338","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}