Assessing the Accuracy of a Generative Model for Creating Novel Spider Silk Proteins

Assessing the Accuracy of a Generative Model for Creating Novel Spider Silk 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 Research Article Assessing the Accuracy of a Generative Model for Creating Novel Spider Silk Proteins Aiden z Ackerman This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7926039/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 Spider silk’s exceptional mechanical properties of extensibility, tensile strength, and Young’s Modulus originate from its unique amino acid sequence as well as its 3D-structure, consisting of a mix of amorphous, flexible regions, and tough, crystalline portions of beta sheets held together by hydrogen bonds. These outstanding physical qualities have made it an ideal material for use in fields such as clothing, body armor, and medicine. However, while there is much data on a variety of spider silk types (MaSp AgSp, Flag), including information on their amino acid sequences and mechanical properties, it has been difficult to quantitatively utilize amino acid sequence or protein structure to reliably predict attributes of spider silk or to replicate them as synthetic silks. Utilizing a prototypical generative model trained on a dataset of existing spider silk sequences, novel spider silks were generated that mimicked natural spider silk sequences in the MaSp1, AgSp1, and Flag classifications of silk. In order to assess the accuracy of the model, the sequences were compared, and their 3D-Structures were analyzed utilizing Root Mean Squared Deviation. The results of this study suggest that the model is inconsistent and produces sequences with a high variability of sequences and structures, meaning that significant improvements will have to be made. Biotechnology and Bioengineering Spider silk mechanical properties amino acid sequence protein structure generative model MaSp1 AgSp1 synthetic silk Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Summary The potential uses of spider silk in industries such as clothing, body armor, and medicine are due to its unique sequence and 3 dimensional structure, which contribute to its outstanding flexibility, toughness, and light weight. However, while there is much data on a variety of spider silk types, such as those which form different portions of the spiderweb, as well as the diversity between many species of spiders, it has been difficult to correlate the makeup of the protein with its 3D-structure to predict physical properties of spider silk or to produce them as synthetic silks. Utilizing a model that relies on artificial intelligence and which was fed data of previously existing spider silks, new amino acid sequences of spider silks were generated; in order to assess the accuracy of the model, the sequences were compared, and the change in their structures over time were analyzed. The results from this study suggest that the model is inconsistent and produces sequences with a high variability, meaning that significant improvements will have to be made. 1 Introduction There are many types of spider silk, ranging from drag-line, which provides the scaffolding of the web and works to absorb energy, to glob-like silks, which are deposited on the capture spiral of the spider web to limit the movement of prey [ 1 ]. Most variants possess several outstanding mechanical properties, including a strength to density ratio surpassing steel and on par with Kevlar [ 2 , 3 ], which is due in part to the numerous hydrogen bonds between the crystalline beta sheets that compose its structure [ 4 ]. These secondary structures partially account for spider silk’s excellent flexibility and toughness[ 3 , 5 ]. These structures further allow spider silks to be used for a variety of purposes, as different combinations grant each of the many types of silk a unique flexibility, shape, and toughness. Due to these unique attributes and its wide variation in makeup, spider silk has attracted considerable attention in industries ranging from clothing to armor to even medicine, where it is used to strengthen tendons or act as degradable micro-prostheses [ 6 , 7 , 8 ]. In part due to the development of significantly more sophisticated programs and simulations, many of which use artificial intelligence and deep learning networks to accurately predict and pinpoint results, there has been renewed interest in exploring the properties of spider silk, starting with uncovering the recurring amino acid sequences and secondary structures that give it its outstanding properties [ 9 , 10 ]. One of these softwares is ColabFold, which is trained with thousands of template protein structures and which can predict how a wide variety of polypeptides will fold by comparing them to an existing database [ 11 ]. There have also been developments in regards to molecular dynamics programs such as Visual Molecular Dynamics (VMD) and NanoScale Molecular Dynamics (NAMD), which simulate the protein’s environment in order to calculate its stability, structure, and reaction to the presence of other proteins [ 12 , 13 ] While there have been updates to databases of newly sequenced spider silk proteins in the past decade [ 14 ], this overflow of data currently only means that most of the gathered information is untapped and has not been explored in depth. In addition, even from current analysis of the categorized data, researchers still struggle to understand from what specific amino acid motifs and 3D structures spider silks inherit their properties. Furthermore, despite extensive research into spider silks and the development of methods to synthesize synthetic variants, there are limitations in the production methods, including ones pertaining to the production of hybrid materials based on silk and developing a reliable spinning process for the manufacturing of synthetic silks [ 15 , 16 ]. To address this problem and to establish a more concrete relationship between the sequence, structure, and mechanical properties of spider silks, a prototypical generative model has been developed at the Laboratory of Atomistic and Molecular Mechanics. This model is unique in that it suggests novel sequences of spider silk proteins based on trends and patterns between the sequence, structure, and mechanical properties it has found in a database of pre-existing instances [ 17 ]. If successful, this would ultimately facilitate the creation of silk protein sequences that we already know the properties of, without having to fold and simulate them in other programs in order to confirm such. A successful model could also help to clarify relationships between the aforementioned sequences, shapes, and traits of the spider silk, a problem currently at the forefront of this field. However, this generative model is also currently poorly understood, and it is unknown how accurate it is at generating silk sequences that are not only novel but also possess the properties the user desires. Therefore, this study aims to assess the efficacy by comparing natural spider silk proteins with spider silk proteins obtained from the generative model. We ultimately hope to establish how different the natural spider silk is from the generated spider silk in regards to both their sequences and 3D structures in regards to their stability. 2 Materials and Methods 2.1 Organizing and Generating Protein Samples Three samples of eight spider silk sequences with more than 100 and less than 300 amino acid residues were chosen at random from the Spider Silkome Database [ 14 ], with each sample containing only sequences from one of three variations of spider silk: Ampullate Silk 1 (MaSp1), Aggregate Glue Silk 1 (AgSp1), and Flagelliform Silk (Flag). For each silk sequence included in the three samples, a corresponding analog was generated, obtained by inserting the first 24 amino acid residues of the chosen silk into the generative model [ 18 ]. Generated analogs with lengths closest to those of their natural counterparts were chosen in order to avoid any potential data skewing while comparing sequences due to significant differences in length. 2.2 Comparison The first sample of natural and artificial silk sequences from MaSp1 were compared by feeding them to the SequenceMatcher.ratio() function from difflib Python library. The function returned a ”similarity score” obtained by adding up all the matching blocks of letters for both sequences and then using the formula: Score = 2.0M/T Where M was the number of matches between the sequences and T was how many total elements were in both sequences combined [ 19 ]. This similarity score was then stored as a decimal between zero and one, with zero indicating the two sequences were completely different and one indicating they were identical. The same process was then repeated with the other two samples, AgSp1 and Flag. This method of comparison was chosen not only because of ease of implementation, needing only one function, but also because of its accuracy when comparing sequences of different lengths; due to the variation of spider silks, many natural silk sequences as well as novel ones generated by the model were of differing lengths. 2.3 Folding and molecular dynamics From each of the MaSp1, AgSp1, and Flag silk samples, one natural spider silk protein and its artificial analog were input into ColabFold and folded into 3D structures to receive a total of six proteins. Using VMD, each folded silk protein was then solvated in a cube of water to mimic its natural environment inside the spider, with the padding of water around the protein in each of the x, y, and z directions being set to 5 Angstroms. Then, the solvated system was ionized with sodium and chloride to bring the net charge to 0 for ease of performing calculations. Finally, the ionized and solvated silk protein was equilibrated, using the NAMD2 software, at a temperature of 300 Kelvin for a 10 nanosecond period in the simulation; during this time the system was set to calculate the position of the atoms of the spider silk protein every 2 femtoseconds and store their coordinates. After equilibration was carried out, the model was allowed to run for a further in-simulation time of 25 nanoseconds, with the atomic positions being recorded and stored every 2 femtoseconds. Finally, the files containing the data were uploaded to VMD, and the Root Mean Standard Deviation (RMSD) – the average deviation of all the atoms from their initial positions at the beginning of the simulation – was calculated using the RMSD Trajectory Tool and graphed over the entire time period. The resulting graphs were then analyzed for the silk proteins’ relative stabilities. 3 Results 3.1 Sequence and Similarity Similarity percentages of the natural spider silk proteins in the three samples were significantly varied. The MaSp1 sample possessed the most variation out of any of the samples, with the largest range, as well as the most extreme lower and upper values. In comparison, the Flag sample possessed a somewhat smaller range and more moderate extrema, while the AgSp1 data set had the least variation. Additionally, the median of the AgSp1 data set was much lower than those of MaSp1 and Flag, with the lower 50% of the scores being between 0.12 and 0.23.The spread of data in the MaSp1 and Flag had the bottom 50% of their similarity scores between 0.12 and 0.46 and 0.12 and 0.38, respectively. However, for the top 50% of the data, the AgSp1 scores were between 0.23 and 0.85, a larger spread than 0.46 to 0.96 and 0.38 to 0.91 of the MaSp1 and Flag silk proteins. 3.2 Root Mean Squared Deviation (a) Natural MaSp1 spider silk protein from Cyclosa Monticola. (b) Generated MaSp1 spider silk protein from Cyclosa Monticola. (c) Natural AgSp1 spider silk protein from Cyrtarachne Akirai (d) Generated AgSp1 spider silk protein from Cyrtarachne Akirai (e) Natural Flag spider silk protein from Cyclose Bifida (f) Generated Flag spider silk protein from Cyclose Bifida Figure 4: Visualization in ColabFold of the secondary structures of the six MaSp1, AgSp1, and Flag spider silk proteins. Water box and ions not visualized. The RMSD graphs of the existing spider silk protein for MaSp1 and the model generated sequence generally followed similar paths up to frame 600, increasing slightly over that period. From frame 600 until approximately frame 1100, the difference in RMSD between the proteins increased, with the natural variant trending upwards and the generated one downwards. However, after frame 1400, the two RMSDs deviated from each other significantly, with the natural silk protein’s RMSD spiking until frame 1500, when it tapered off and reached equilibrium. The model-generated protein, on the other hand, retained a relatively stable RMSD starting from approximately frame 1000 until the end of the time frame. The AgSp1 natural and generated spider silks had similar RMSD values from frame 1 to 950, both increasing at a steady rate. However, at approximately 1000 frames, the RMSD (a) MaSp1 natural and generated silks of the spider species Cyclosa Monticola. (b) AgSp1 natural and generated silks of the spider species Cyrtarachne Akirai (c) Flag natural and generated silks of the spider species Cyclose Bifida Figure 5: The RMSD graphs for the natural and model artificial sequences the samples, corresponding to a total time frame of 35 nanoseconds or 1754 frames and subdivided into periods of 73 frames of the proteins deviated significantly, with the generated silk protein remaining relatively stable while climbing slightly. During the same period, the natural silk protein suddenly spiked before reaching equilibrium at a larger RMSD of 20 Angstroms, the highest value of any of the proteins in the RMSD graphs. The RMSD graph for the natural and artificial Flag spider silk proteins of the Cyclose Bifida species had significantly different values throughout the entire time frame of the simulation. Throughout the 1752 frames, the natural spider silk remained at an RMSD between 2 and 4 Angstroms with no major spikes at any point. On the other hand, the artificial silk protein RMSD spiked multiple times between 500 and 800 frames, stabilizing at 8 Angstroms but decreasing slightly starting at around 1500 frames. At no point in the simulation was there any overlap or crossing of RMSD values between the natural and artificial silk proteins, as opposed to the MaSp1 and AgSp1 RMSD graphs. 4 Discussion 4.1 Sequence Analysis The similarity scores of the three types of silk proteins/samples indicate that there is still much to be done in calibrating the generative model. In regards to the similarity scores, we decided on values between 0.3 and 0.6 as scores that best indicate that a novel new spider silk protein was generated. A score below a 0.3 meant that an entirely different protein than the one inputted was created by the model, and anything above 0.6 suggested that the artificial silk sequence was not as likely to be novel and could be too similar to the natural one [ 20 , 21 , 22 ]. With these thresholds, only four of the eight generated silk sequences satisfied the conditions of being not only novel but also potentially possessing similar properties to the existing/natural silk for MaSp1, and only three fit the criteria for both the AgSp1 and Flag samples. This level of inconsistency could point to the fact that there is a large degree of inaccuracy in the generative model, potentially caused by an inadequate data set or one with skewed data. However, it is more likely that the various types of spider silks across different species are exceptionally diverse. Therefore, it would require much more work to identify and correlate all of the relevant amino acid motifs or structures that make each of them unique and give the types of silk from each species their own distinctive mechanical properties. There will have to be success in the implementation of this new knowledge to the model before it will be able to produce a sequence that is both novel and still the type of spider silk that is desired, along with its corresponding mechanical properties. 4.2 RMSD Analysis The RMSD of the Cyclosa Monticola MaSp1 sequences indicates that over the period of the 35 nanoseconds the molecular dynamics simulation was run, the model-generated spider silk sequence was much more stable structurally than the natural pre-existing one. The lessened frequency of significant spikes and sudden changes in its RMSD overall as well as a generally lower RMSD score in comparison with its analog indicates that it was closer to its initial shape at the start of the simulation and was more stable. This suggests that the generated amino acid could potentially be able to fold into a simpler shape, thus explaining why it had a much more stable RMSD; whether this shape contributes positively or negatively to its mechanical properties will have to be tested in the future. Furthermore, this finding brings up the possibility of the generative model’s use for crafting more stable spider silk sequences, which could facilitate and quicken the process of studying spider silks in the future. However, it is equally likely that the software utilized in this experiment – either the ColabFold or Molecular Dynamics programs – could not accurately fold the natural version of the Cyclosa Monticola MaSp1 silk sequence, therefore creating an unstable structure that changed its structure sharply during the simulation. The RMSD graph of the AgSp1 silk protein from Cyrtarachne Akirai shows a much closer relationship between the RMSD of the natural and generated sequences, with significant overlap of their respective lines. However, the sudden leap of 5 Angstroms at 1000 frames indicates instability with the artificial sequence of the AgSp1 protein, the opposite situation compared to the the MaSp1 one. In addition, while the pre-existing Cyrtarachne Akirai silk’s RMSD did not spike as often, its steadily increasing RMSD during the latter half of the time frame suggests that it had not reached full equilibrium yet. Compared to the model generated protein, which leveled out after the spike at 1000 frames, the pre-existing silk steadily increased a total of 4 Angstroms after the 1000 frame mark, displaying its continued instability in relation to its artificial analog. This once again brings up the possibility of further study into the generative model, and that if the sequences it generates are consistently more stable, its uses in the further study of silk proteins. Out of all the types of spider silk proteins, the difference of the RMSD between the natural protein and generated analog of the Flag sample was the most significant. This could have been because the silk chosen for the RMSD calculations was extremely stable compared with other members of that same sample or that natural Flag proteins are generally more stable than the generated versions due to errors in the creation process. To support this, according to the RMSD graphs, the natural Flag silk protein peaked at just 4 Angstroms, while the silk proteins of MaSp1 and AgSp1 peaked at about 18 and 16 Angstroms respectively. Out of all the types of spider silk proteins, the difference of the RMSD between the natural protein and generated analog of the Flag sample was the most significant. This could have been because the silk chosen for the RMSD calculations was extremely stable compared with other members of that same sample or that natural Flag proteins are generally more stable than the generated versions due to errors in the creation process. To support this, according to the RMSD graphs, the natural Flag silk protein peaked at just 4 Angstroms, while the silk proteins of MaSp1 and AgSp1 peaked at about 18 and 16 Angstroms respectively. A possible explanation for why the RMSD of both of the Flag proteins were much lower in general than those in the MaSp1 and AgSp1 graphs could be due to the relative shortness of the chosen natural Flag silk sequence, which was at 115 amino acids. Compared to the 170 and 200 amino acid residues of the MaSp1 and AgSp1 silk sequences, respectively, it was significantly shorter. This considerably shortened length could have resulted in a less complex secondary and tertiary structure, and thus less deviation from the initial structure and more stability overall. Overall, the generative model seems to be the most accurate at mimicking the stability of the 3D-Structure of the artificial silk protein compared to the natural analog for AgSp1, showing the least difference between the two RMSD values. However, the generative model struggles the most with suggesting sequences that possess similar stabilities and deviations from the natural spider silk protein in the Flag group, and further research will have to be done with larger sample sizes to establish a more concrete explanation and to improve the efficacy of the model in generating new spider silk sequences. The spider silk proteins from the MaSp1 sample are still somewhat accurate in terms of stability, but there are points where there is considerable deviation between the RMSD values between the pre-existing and artificial silk proteins. 5 Conclusion In this study, the accuracy of a generative model, used to produce novel spider silk sequences, was determined by analyzing the sequences it generated as well as their stabilities and structures. We utilized existing technologies such as NAMD, but also novel ones including ColabFold, which was used to fold the sequences of spider silk, both natural and artificial, as well as VMD, which was used to visualize the structure and dynamics of the various spider silk proteins. In the study, the similarity of eight natural silk sequences and eight generated ones for each sample were calculated, and the RMSD between two of them compared for the MaSp1, Flag, and AgSp1 samples. Ultimately, it was assessed that there was a wide variation and spread in similarity between the natural and artificial silk sequences, with only half or less of all three samples satisfying the conditions of being both novel and being the same type of spider silk as its natural analog. The results from the RMSD were also analyzed, providing insight into the accuracy of the generative model in mimicking the stability of real spider silk proteins. It was established that for the MaSp1 and AgSp1 silk proteins, the artificial versions were generally more stable, with a smaller average RMSD value, while the proteins from Flag possessed significantly lower RMSD values, with the generated analog’s structure fluctuating more. It was ultimately determined that while the generative model produced a large spread of sequences, some of which were too similar or different from the natural silks they were based off of, the stability of the generated proteins was to be studied as well as expanded sample sizes of sequences for both sequence analysis and comparison of 3D structures. 5.1 Ongoing Work Due to the substantial differences and variations between not only the real and generated spider silk proteins, but also between the types of silk examined in the project, a larger sample size is currently being studied, and significantly more proteins are being processed through simulations to gather data on RMSD depending on silk type and to determine what role silk sequence length plays in the RMSD values and structural stability. In addition, the sample sizes will be revised to only include silk proteins from the same genus to minimize variation caused by genetic differences. Finally, due to temporal restrictions and the fact that there is currently no general purpose statistical test that can determine if two time series graphs are significantly different, further work requires the development of a model which can do so for more in-depth quantitative analysis. 5.2 Future Work For future studies relating to the prototypical generative model, the mechanical properties of the generated spider silks examined in this study should also be tested in the future through simulations, which could possibly bring much more concrete data that we can analyze; testing the properties of the generated spider silks could be much more substantial than simply comparing differences in amino acid sequence, as a protein’s sequence is still a somewhat faulty indicator of its properties and structure. Additionally, the generative model utilized in this study should also have to be trained upon a much larger variation of data before future tests of its accuracy. A larger set of data would likely improve its accuracy in detecting amino acid motifs and quirks of each of the different types of spider silks, even when researchers would not have recognized them. 5.3 Implications and Potential Applications If improved upon enough, generative models like the one used in this project could be, as mentioned before, useful in establishing more concrete relationships between larger sections of amino acids and motifs in the spider silk proteins and their physical properties instead of a couple repeating patterns that are loosely associated with those aspects. This would not only allow us to create and modify existing silk to amplify its outstanding properties, but also allow us to know, in general, more about how to create stronger hybrid materials and mimetics. Long-term, such models would be able to create polypeptide “blueprints” quickly that we could then produce for many fields, such as medicine, with prosthetics, such as strengthening/replacing tendons, or construction. Declarations 6 Acknowledgments I am grateful to the Executive Vice President of The Center for Excellence in Education (CEE) Maite Ballestero, and the Director of RSI, mentor, Dr. Mark Kantrowitz, for their continued support. I would like to thank my mentor at the Laboratory of Atomistic and Molecular Mechanics, Professor Markus Buehler, and lab members Amadeus Cavalcanti Salvador de Alcantara and Wei Lu. I would also like to thank Catherine Xue, PhD of bio-statistics at Harvard, who was my tutor and provided feedback on my project’s initial drafts. And thanks to the Regeneron Science Talent Search (STS) for the Scholar recognition and awards. Finally, I am very grateful for my generous sponsors and the Department of Defense, as well as CEE, RSI, MIT, and STS for providing me this valuable opportunity, and Awards! References Eisoldt L, Smith A, Scheibel T (2011) Decoding the secrets of spider silk. Materials To day , 14(3):pp. 80–86. https://doi.org/10.1016/S1369-7021(11)70057-8 . URL https://www.sciencedirect.com/science/article/pii/S1369702111700578 Blamaires SJ, Blackledge TA (2017) and T. I-Min. Physicochemical property variation in spider silk: Ecology, evolution, and synthetic production. Annual Review of Entomology , 62(1):pp. 443–460. 10.1146/annurev-ento-031616-035615 . PMID: 27959639 Morin A, Pahlevan M, Alam P (2017) Silk Biocomposites: Structure and Chemistry , Chap. 8, pp. 189–219. John Wiley Sons, Ltd. https://doi.org/10.1002/9781119441632.ch8 . URL https://onlinelibrary.wiley.com/doi/abs/10.1002/9781119441632.ch8 Keten S, Xu Z, Ihle B et al (2010) Nanoconfinement controls stiffness, strength and mechanical toughness of β-sheet crystals in silk. Nat News Gu Y, Yu L, Mou J et al (2020) Mechanical properties and application analysis of spider silk bionic material. e-Polymers , 20(1):pp. 443–457, doi: 10.1515/epoly-2020-0049 . URL https://doi.org/10.1515/epoly-2020-0049 Kluge JA, Rabotyagova O, Leisk GG et al (2008) Spider silks and their applications. Cell 26(5):244–251. https://doi.org/10.1016/j.tibtech.2008.02.006 Khan A, Jawaid M (2020) Novel spider silk fiber for high-performance textiles and eco-friendly armor applications. J Nat Fibers 19(5):1879–1884. https://doi.org/10.1080/15440478.2020.1788496 Milazzo M, Fitzpatrick V, Owens CE et al (2023) 3d printability of silk/hydroxyapatite composites for microprosthetic applications. ACS Biomaterials Sci Eng 9(3):1285–1295. 10.1021/acsbiomaterials.2c01357 Malay AD, Arakawa K, Numata K (2017) Analysis of repetitive amino acid motifs reveals the essential features of spider dragline silk proteins. PLoS ONE 12(8). 10.1371/journal.pone.0183397 Jenkins JE, Creager MS, Lewis RV et al (2010) Quantitative correlation between the protein primary sequences and secondary structures in spider dragline silks. ACS Publications 1(1):192–200. https://doi.org/10.1021/bm9010672 Mirdita M, Schutze K, Moriwaki Y et al (2022) Colabfold: making protein folding accessible to all. Nat Methods 19:679–682. https://doi.org/10.1038/s41592-022-01488-1 Phillips JC, Hardy DJ, Maia JD et al (2020) Scalable molecular dynamics on cpu and gpu architectures with namd. J Chem Phys 153(4). https://doi.org/10.1063/5.0014475 Humphrey W, Dalke A, Schulten K (1996) Vmd: visual molecular dynamics. J Mol Graph 14(1):33–38. https://doi.org/10.1016/0263-7855 Arakawa K, Kong N et al (2022) 1000 spider silkomes: Linking sequences to silk physical properties. ScienceAdvances 8(41). 10.1126/sciadv.abo6043 Blamaires SJ, Spicer PT (2020) and P. J. Flanagan. Spider silk biomimetics programs to inform the development of new wearable technologies. Frontiers, 7 An B, Jenkins JE, Sampath S et al (2012) Reproducing natural spider silks’ copolymer behavior in synthetic silk mimics. Biomacromolecules 13(12):3938–3948. 10.1021/bm301110s M. J. Buehler. Generative pretrained autoregressive transformer graph neural network applied to the analysis and discovery of novel proteins. arXiv, (2023) 2305.04934. M. J. Buehler. Generative model for conditioned silk sequences. Obtained from the Laboratory for Atomistic and Molecular Mechanics at MIT difflib — helpers for computing deltas URL https://docs.python.org/3/library/ difflib.html#module-difflib Ni B, Kaplan DL (2023) and M. J. Buehler. Generative design of de novo proteins based on secondary-structure constraints using an attention-based diffusion model. Chem , 9(7):pp. 1828–1849. https://doi.org/10.1016/j.chempr.2023.03.020 . URL https://www.sciencedirect.com/science/article/pii/S2451929423001390 Quan L, Wu T, Lyu Q (2023) Computational de novo protein design: From secondary to primary, then toward tertiary structures. Chem , 9(7):pp. 1625–1627. https://doi.org/10.1016/j.chempr.2023.06.010 . URL https://www.sciencedirect.com/science/article/pii/S2451929423003157 Pearson WR (2013) An introduction to sequence similarity (homology) searching. Curr Protocols 42(7). https://doi.org/10.1002/0471250953.bi0301s42 Additional Declarations The authors declare no competing interests. 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-7926039","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":533561171,"identity":"b25da906-11b6-414a-bea5-4d7d767ab426","order_by":0,"name":"Aiden z Ackerman","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA60lEQVRIie3PsWoCMRjA8Q8+yPSdt54E2leIBCJi8FkCgVsrdO1QEOrWWxV8Cd8gR8COtx44FddughSE0lzrZgnn1iF/CCEhP5IApFL/MQTRTSL/WekwHEDRiwyfMUxlHwIXIlxfki9xe5g/zaRsmvo4N/ph4LDeU4QUnj3K1c4q1VrkK1NO1o7ZaYyAJ8WJOa1aBJ59+vDCbici7j2Nz/TltKw8nsl0JD9FiQi3YPbilADL+C8hFiWj8BeevVpZtFZNyZRi6JmcbCLkrvHbI51mo6qqD3syWgzeFu/tR+z71+Ftx1OpVCr1R9/D8EV0moaPkgAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0002-5176-4188","institution":"Stanford University","correspondingAuthor":true,"prefix":"","firstName":"Aiden","middleName":"z","lastName":"Ackerman","suffix":""}],"badges":[],"createdAt":"2025-10-22 18:17:20","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":true,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":true},"doi":"10.21203/rs.3.rs-7926039/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7926039/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":94391918,"identity":"ff9469a4-7c10-4b2c-bfa2-c43a637b977e","added_by":"auto","created_at":"2025-10-27 13:53:12","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":274797,"visible":true,"origin":"","legend":"","description":"","filename":"MITRSISTSAwardsAidenAckermanStanfordUniversity.docx","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/bbed1bef5ed493e47d08421c.docx"},{"id":94392865,"identity":"4c617628-c3b4-414c-b34d-3237b1cfa1cd","added_by":"auto","created_at":"2025-10-27 13:53:40","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs7926039.json","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/19edab5654fc5f81161b9047.json"},{"id":94394350,"identity":"b754cfc2-73d8-43bc-8db8-3e3646e63820","added_by":"auto","created_at":"2025-10-27 13:54:35","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":66242,"visible":true,"origin":"","legend":"","description":"","filename":"rs79260390enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/dfab8b41407426847a36d2d8.xml"},{"id":94392181,"identity":"068162aa-20fc-43f3-a7de-8141d2da7765","added_by":"auto","created_at":"2025-10-27 13:53:23","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":78423,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/ad59e416d56d0746c82a4fbd.png"},{"id":94392808,"identity":"821046c3-fc35-477c-b548-f0ed177a5fcc","added_by":"auto","created_at":"2025-10-27 13:53:38","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20422,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/d877ac4434318cf1e79a6b9f.png"},{"id":94391606,"identity":"8109c7b6-59f7-4a03-ab3c-d193e016dc28","added_by":"auto","created_at":"2025-10-27 13:52:58","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17826,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/37fab507068003f40fc5864f.png"},{"id":94393874,"identity":"67ad8dbd-3266-483a-8cf6-a4d4eac381bd","added_by":"auto","created_at":"2025-10-27 13:54:10","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":20820,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/0e706a60b2a5b294a28520bc.png"},{"id":94392208,"identity":"eb7f9b14-ee7f-469e-87d0-93d3cf3fb00e","added_by":"auto","created_at":"2025-10-27 13:53:23","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":431770,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/d98c47cbae9d9c1ecd70a3e2.jpeg"},{"id":94392744,"identity":"3ba0f97e-88f3-4de7-a7e9-80693394a10c","added_by":"auto","created_at":"2025-10-27 13:53:37","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":124348,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/6886b97e27763fc9c12bd9f7.jpeg"},{"id":94392913,"identity":"86d48e4e-891a-4f94-b4c3-e6546eb92bec","added_by":"auto","created_at":"2025-10-27 13:53:41","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6278,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/a7342a0a93819c820ad3dea3.png"},{"id":94394096,"identity":"4cd61a4d-f547-46c5-848a-2c64ff768eb0","added_by":"auto","created_at":"2025-10-27 13:54:17","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4982,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/7eef40eb67294a11b4b1a450.png"},{"id":94392630,"identity":"6bba2bdb-c733-486b-8111-4d49d0e31f48","added_by":"auto","created_at":"2025-10-27 13:53:34","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6436,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/fd1d7cae23a9dc5ab4971c71.png"},{"id":94394355,"identity":"25483e0b-7848-48ab-830b-2295dd86e221","added_by":"auto","created_at":"2025-10-27 13:54:35","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7384,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/c4e1bcfdf779b7ad2cc20bc4.png"},{"id":94489253,"identity":"d5302c69-b52f-408b-baf4-e96021381a73","added_by":"auto","created_at":"2025-10-27 17:04:00","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5459,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/14329d50a3f4a7b0c0fc9897.png"},{"id":94391539,"identity":"6b249027-b7a2-48d9-b5db-1ece2723d4df","added_by":"auto","created_at":"2025-10-27 13:52:56","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5634,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/cc4192387e85c501fca71b6c.png"},{"id":94393554,"identity":"3fa3e603-9212-4690-806b-3be21f98d6aa","added_by":"auto","created_at":"2025-10-27 13:54:02","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16965,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/4264b06af456d91b7e174ec0.png"},{"id":94393687,"identity":"1361d6f0-079e-42ec-94ca-44a6bed99f6c","added_by":"auto","created_at":"2025-10-27 13:54:06","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5353,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/5072714e23c3ccf685019abe.png"},{"id":94392212,"identity":"2fdba715-a1d6-468c-9858-e3b0771ac42e","added_by":"auto","created_at":"2025-10-27 13:53:23","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4865,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/44b27504994596780ddc4536.png"},{"id":94393067,"identity":"f1706c69-2052-451d-9708-d2f4f989cdc8","added_by":"auto","created_at":"2025-10-27 13:53:48","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5228,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/1ec92ef7f5463040a7f58bfb.png"},{"id":94393177,"identity":"01c4268c-bcf7-43b5-8024-1a2c0af6e923","added_by":"auto","created_at":"2025-10-27 13:53:53","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":115505,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/ea27fc26f8ff094e75925d0f.png"},{"id":94393186,"identity":"36cc0809-6af4-4f8b-9705-5e19741d1053","added_by":"auto","created_at":"2025-10-27 13:53:53","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":21666,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/1681f68e841522ee804691a7.png"},{"id":94391612,"identity":"84ae7111-81ba-46fe-97cf-5cfd6eb0c98e","added_by":"auto","created_at":"2025-10-27 13:52:59","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2590,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/9272557632d7e0fa5e4ca156.png"},{"id":94393548,"identity":"66ecb970-c5ff-4fd4-8d59-f7820336eb31","added_by":"auto","created_at":"2025-10-27 13:54:02","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2237,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/5f672b8a15a707a0d98a53b0.png"},{"id":94391895,"identity":"c7309cf6-d740-4ea5-b0fe-f4c1fd5193aa","added_by":"auto","created_at":"2025-10-27 13:53:10","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2675,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/b62ebba301c0b013cb615487.png"},{"id":94393175,"identity":"0f0691e9-e44f-4797-92e0-eb6beb57ee87","added_by":"auto","created_at":"2025-10-27 13:53:52","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2196,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/4fff084ae770de4014623d5e.png"},{"id":94393763,"identity":"e4bb15cd-8d14-4dd6-87e0-f8c68eb04a75","added_by":"auto","created_at":"2025-10-27 13:54:08","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2368,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/79b77c10777ee6b54505df49.png"},{"id":94392915,"identity":"cb221791-31a6-4749-83ae-8b47cc39b07d","added_by":"auto","created_at":"2025-10-27 13:53:42","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2470,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/3c8c424f4829038a55f80626.png"},{"id":94394256,"identity":"8ca00ec8-e7b2-49b0-b612-1045b0d42843","added_by":"auto","created_at":"2025-10-27 13:54:27","extension":"xml","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":64533,"visible":true,"origin":"","legend":"","description":"","filename":"rs79260390structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/720832e5acb9bfd0ec039e3b.xml"},{"id":94393001,"identity":"0da6ab93-bd7f-40dc-9f3a-d2b7c64f0f1f","added_by":"auto","created_at":"2025-10-27 13:53:45","extension":"html","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":74127,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/374f77cf8ed0c22085b93e6b.html"},{"id":94394102,"identity":"b407ec86-cfe1-49a5-8e6e-4f393b340b95","added_by":"auto","created_at":"2025-10-27 13:54:18","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":28357,"visible":true,"origin":"","legend":"\u003cp\u003eTypes of silk produced by spiders, which includes Aggregate Glue, Major Ampullate, and Flagelliform silks, among others. (Eishold, Smith, and Scheibel, 2011)\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/72e3acfbe1238b5b656c9657.jpg"},{"id":94393678,"identity":"2570309b-4348-41e9-b5cf-31d7d6b691f5","added_by":"auto","created_at":"2025-10-27 13:54:06","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":36884,"visible":true,"origin":"","legend":"\u003cp\u003eMethod using Visual Molecular Dynamics software, with placeholder ubiquitin protein used as example. Also shown are the solvation and ionization tools and proper parameters used in this project and the water box in the background.\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/544dbcab239fe554ef2df3be.jpg"},{"id":94393543,"identity":"416d3954-ddf8-4d89-9917-cbd162520aa2","added_by":"auto","created_at":"2025-10-27 13:54:01","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":15790,"visible":true,"origin":"","legend":"\u003cp\u003eBox and whisker plot of each of the three samples and their similarity scores from 0 to 1, showcasing the median, quartiles, and extrema. The mean is shown as an X for each sample.\u003c/p\u003e","description":"","filename":"Picture3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/c3c42f9ceab2a34d2d89242b.jpg"},{"id":94393063,"identity":"c21a04a4-4231-4d4c-bc37-ab154d7932bd","added_by":"auto","created_at":"2025-10-27 13:53:48","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":92998,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization in ColabFold of the secondary structures of the six MaSp1, AgSp1, and Flag spider silk proteins. Water box and ions not visualized.\u003c/p\u003e","description":"","filename":"Picture4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/871a648a934dfc5a0181d61b.jpg"},{"id":94392598,"identity":"0bc563de-1971-4c5d-8275-d2523ae49387","added_by":"auto","created_at":"2025-10-27 13:53:34","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":218804,"visible":true,"origin":"","legend":"\u003cp\u003eThe RMSD graphs for the natural and model artificial sequences the samples, corresponding to a total time frame of 35 nanoseconds or 1754 frames and subdivided into periods of 73 frames\u003c/p\u003e","description":"","filename":"Picture5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/49311b232b949822aa354ca8.jpg"},{"id":94491206,"identity":"0fe6f051-31e4-47e0-8eb5-6b4d30e9500c","added_by":"auto","created_at":"2025-10-27 17:23:51","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":860226,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7926039/v1/38f61093-1702-490c-b540-307764aa4998.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eAssessing the Accuracy of a Generative Model for Creating Novel Spider Silk Proteins\u003c/p\u003e","fulltext":[{"header":"Summary ","content":"\u003cp\u003eThe potential uses of spider silk in industries such as clothing, body armor, and medicine are due to its unique sequence and 3 dimensional structure, which contribute to its outstanding flexibility, toughness, and light weight. However, while there is much data on a variety of spider silk types, such as those which form different portions of the spiderweb, as well as the diversity between many species of spiders, it has been difficult to correlate the makeup of the protein with its 3D-structure to predict physical properties of spider silk or to produce them as synthetic silks. Utilizing a model that relies on artificial intelligence and which was fed data of previously existing spider silks, new amino acid sequences of spider silks were generated; in order to assess the accuracy of the model, the sequences were compared, and the change in their structures over time were analyzed. The results from this study suggest that the model is inconsistent and produces sequences with a high variability, meaning that significant improvements will have to be made.\u003c/p\u003e"},{"header":"1 Introduction","content":"\u003cp\u003eThere are many types of spider silk, ranging from drag-line, which provides the scaffolding of the web and works to absorb energy, to glob-like silks, which are deposited on the capture spiral of the spider web to limit the movement of prey [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Most variants possess several outstanding mechanical properties, including a strength to density ratio surpassing steel and on par with Kevlar [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], which is due in part to the numerous hydrogen bonds between the crystalline beta sheets that compose its structure [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. These secondary structures partially account for spider silk\u0026rsquo;s excellent flexibility and toughness[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. These structures further allow spider silks to be used for a variety of purposes, as different combinations grant each of the many types of silk a unique flexibility, shape, and toughness.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eDue to these unique attributes and its wide variation in makeup, spider silk has attracted considerable attention in industries ranging from clothing to armor to even medicine, where it is used to strengthen tendons or act as degradable micro-prostheses [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. In part due to the development of significantly more sophisticated programs and simulations, many of which use artificial intelligence and deep learning networks to accurately predict and pinpoint results, there has been renewed interest in exploring the properties of spider silk, starting with uncovering the recurring amino acid sequences and secondary structures that give it its outstanding properties [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. One of these softwares is ColabFold, which is trained with thousands of template protein structures and which can predict how a wide variety of polypeptides will fold by comparing them to an existing database [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. There have also been developments in regards to molecular dynamics programs such as Visual Molecular Dynamics (VMD) and NanoScale Molecular Dynamics (NAMD), which simulate the protein\u0026rsquo;s environment in order to calculate its stability, structure, and reaction to the presence of other proteins [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/p\u003e\u003cp\u003eWhile there have been updates to databases of newly sequenced spider silk proteins in the past decade [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], this overflow of data currently only means that most of the gathered information is untapped and has not been explored in depth. In addition, even from current analysis of the categorized data, researchers still struggle to understand from what specific amino acid motifs and 3D structures spider silks inherit their properties. Furthermore, despite extensive research into spider silks and the development of methods to synthesize synthetic variants, there are limitations in the production methods, including ones pertaining to the production of hybrid materials based on silk and developing a reliable spinning process for the manufacturing of synthetic silks [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTo address this problem and to establish a more concrete relationship between the sequence, structure, and mechanical properties of spider silks, a prototypical generative model has been developed at the Laboratory of Atomistic and Molecular Mechanics. This model is unique in that it suggests novel sequences of spider silk proteins based on trends and patterns between the sequence, structure, and mechanical properties it has found in a database of pre-existing instances [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. If successful, this would ultimately facilitate the creation of silk protein sequences that we already know the properties of, without having to fold and simulate them in other programs in order to confirm such. A successful model could also help to clarify relationships between the aforementioned sequences, shapes, and traits of the spider silk, a problem currently at the forefront of this field.\u003c/p\u003e\u003cp\u003eHowever, this generative model is also currently poorly understood, and it is unknown how accurate it is at generating silk sequences that are not only novel but also possess the properties the user desires. Therefore, this study aims to assess the efficacy by comparing natural spider silk proteins with spider silk proteins obtained from the generative model. We ultimately hope to establish how different the natural spider silk is from the generated spider silk in regards to both their sequences and 3D structures in regards to their stability.\u003c/p\u003e"},{"header":"2 Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Organizing and Generating Protein Samples\u003c/h2\u003e\u003cp\u003eThree samples of eight spider silk sequences with more than 100 and less than 300 amino acid residues were chosen at random from the Spider Silkome Database [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], with each sample containing only sequences from one of three variations of spider silk: Ampullate Silk 1 (MaSp1), Aggregate Glue Silk 1 (AgSp1), and Flagelliform Silk (Flag). For each silk sequence included in the three samples, a corresponding analog was generated, obtained by inserting the first 24 amino acid residues of the chosen silk into the generative model [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Generated analogs with lengths closest to those of their natural counterparts were chosen in order to avoid any potential data skewing while comparing sequences due to significant differences in length.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Comparison\u003c/h2\u003e\u003cp\u003eThe first sample of natural and artificial silk sequences from MaSp1 were compared by feeding them to the SequenceMatcher.ratio() function from difflib Python library. The function returned a \u0026rdquo;similarity score\u0026rdquo; obtained by adding up all the matching blocks of letters for both sequences and then using the formula:\u003c/p\u003e\u003cp\u003e\u003cem\u003eScore\u0026thinsp;=\u0026thinsp;2.0M/T\u003c/em\u003e\u003c/p\u003e\u003cp\u003eWhere \u003cem\u003eM\u003c/em\u003e was the number of matches between the sequences and \u003cem\u003eT\u003c/em\u003e was how many total elements were in both sequences combined [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. This similarity score was then stored as a decimal between zero and one, with zero indicating the two sequences were completely different and one indicating they were identical. The same process was then repeated with the other two samples, AgSp1 and Flag. This method of comparison was chosen not only because of ease of implementation, needing only one function, but also because of its accuracy when comparing sequences of different lengths; due to the variation of spider silks, many natural silk sequences as well as novel ones generated by the model were of differing lengths.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Folding and molecular dynamics\u003c/h2\u003e\u003cp\u003eFrom each of the MaSp1, AgSp1, and Flag silk samples, one natural spider silk protein and its artificial analog were input into ColabFold and folded into 3D structures to receive a total of six proteins. Using VMD, each folded silk protein was then solvated in a cube of water to mimic its natural environment inside the spider, with the padding of water around the protein in each of the x, y, and z directions being set to 5 Angstroms. Then, the solvated system was ionized with sodium and chloride to bring the net charge to 0 for ease of performing calculations.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFinally, the ionized and solvated silk protein was equilibrated, using the NAMD2 software, at a temperature of 300 Kelvin for a 10 nanosecond period in the simulation; during this time the system was set to calculate the position of the atoms of the spider silk protein every 2 femtoseconds and store their coordinates. After equilibration was carried out, the model was allowed to run for a further in-simulation time of 25 nanoseconds, with the atomic positions being recorded and stored every 2 femtoseconds. Finally, the files containing the data were uploaded to VMD, and the Root Mean Standard Deviation (RMSD) \u0026ndash; the average deviation of all the atoms from their initial positions at the beginning of the simulation \u0026ndash; was calculated using the RMSD Trajectory Tool and graphed over the entire time period. The resulting graphs were then analyzed for the silk proteins\u0026rsquo; relative stabilities.\u003c/p\u003e\u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec7\"\u003e\n \u003ch2\u003e3.1 Sequence and Similarity\u003c/h2\u003e\n \u003cp\u003eSimilarity percentages of the natural spider silk proteins in the three samples were significantly varied. The MaSp1 sample possessed the most variation out of any of the samples, with the largest range, as well as the most extreme lower and upper values. In comparison, the Flag sample possessed a somewhat smaller range and more moderate extrema, while the AgSp1 data set had the least variation.\u003c/p\u003e\n \u003cp\u003eAdditionally, the median of the AgSp1 data set was much lower than those of MaSp1 and Flag, with the lower 50% of the scores being between 0.12 and 0.23.The spread of data in the MaSp1 and Flag had the bottom 50% of their similarity scores between 0.12 and 0.46 and 0.12 and 0.38, respectively. However, for the top 50% of the data, the AgSp1 scores were between 0.23 and 0.85, a larger spread than 0.46 to 0.96 and 0.38 to 0.91 of the MaSp1 and Flag silk proteins.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\"\u003e\n \u003ch2\u003e3.2 Root Mean Squared Deviation\u003c/h2\u003e\n \u003cp\u003e(a) Natural MaSp1 spider silk protein from \u003cem\u003eCyclosa Monticola.\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003e(b) Generated MaSp1 spider silk protein from \u003cem\u003eCyclosa Monticola.\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003e(c) Natural AgSp1 spider silk protein from \u003cem\u003eCyrtarachne Akirai\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003e(d) Generated AgSp1 spider silk protein from \u003cem\u003eCyrtarachne Akirai\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003e(e) Natural Flag spider silk protein from\u003c/p\u003e\n \u003cp\u003e\u003cem\u003eCyclose Bifida\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003e(f) Generated Flag spider silk protein from \u003cem\u003eCyclose Bifida\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003eFigure 4: Visualization in ColabFold of the secondary structures of the six MaSp1, AgSp1, and Flag spider silk proteins. Water box and ions not visualized.\u003c/p\u003e\n \u003cp\u003eThe RMSD graphs of the existing spider silk protein for MaSp1 and the model generated sequence generally followed similar paths up to frame 600, increasing slightly over that period. From frame 600 until approximately frame 1100, the difference in RMSD between the proteins increased, with the natural variant trending upwards and the generated one downwards. However, after frame 1400, the two RMSDs deviated from each other significantly, with the natural silk protein\u0026rsquo;s RMSD spiking until frame 1500, when it tapered off and reached equilibrium. The model-generated protein, on the other hand, retained a relatively stable RMSD starting from approximately frame 1000 until the end of the time frame.\u003c/p\u003e\n \u003cp\u003eThe AgSp1 natural and generated spider silks had similar RMSD values from frame 1 to 950, both increasing at a steady rate. However, at approximately 1000 frames, the RMSD\u003c/p\u003e\n \u003cp\u003e(a) MaSp1 natural and generated silks of the spider\u003c/p\u003e\n \u003cdiv\u003e\n \u003cp\u003especies \u003cem\u003eCyclosa Monticola.\u003c/em\u003e\u003c/p\u003e\n \u003c/div\u003e\n \u003cp\u003e(b) AgSp1 natural and generated silks of the spider\u003c/p\u003e\n \u003cp\u003especies \u003cem\u003eCyrtarachne Akirai\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003e(c) Flag natural and generated silks of the spider\u003c/p\u003e\n \u003cp\u003especies \u003cem\u003eCyclose Bifida\u003c/em\u003e\u003c/p\u003e\n \u003cp\u003eFigure 5: The RMSD graphs for the natural and model artificial sequences the samples, corresponding to a total time frame of 35 nanoseconds or 1754 frames and subdivided into periods of 73 frames\u003c/p\u003e\n \u003cp\u003eof the proteins deviated significantly, with the generated silk protein remaining relatively stable while climbing slightly. During the same period, the natural silk protein suddenly spiked before reaching equilibrium at a larger RMSD of 20 Angstroms, the highest value of any of the proteins in the RMSD graphs.\u003c/p\u003e\n \u003cp\u003eThe RMSD graph for the natural and artificial Flag spider silk proteins of the \u003cem\u003eCyclose Bifida\u003c/em\u003e species had significantly different values throughout the entire time frame of the simulation. Throughout the 1752 frames, the natural spider silk remained at an RMSD between 2 and 4 Angstroms with no major spikes at any point. On the other hand, the artificial silk protein RMSD spiked multiple times between 500 and 800 frames, stabilizing at 8 Angstroms but decreasing slightly starting at around 1500 frames. At no point in the simulation was there any overlap or crossing of RMSD values between the natural and artificial silk proteins, as opposed to the MaSp1 and AgSp1 RMSD graphs.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Sequence Analysis\u003c/h2\u003e\u003cp\u003eThe similarity scores of the three types of silk proteins/samples indicate that there is still much to be done in calibrating the generative model. In regards to the similarity scores, we decided on values between 0.3 and 0.6 as scores that best indicate that a novel new spider silk protein was generated. A score below a 0.3 meant that an entirely different protein than the one inputted was created by the model, and anything above 0.6 suggested that the artificial silk sequence was not as likely to be novel and could be too similar to the natural one [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. With these thresholds, only four of the eight generated silk sequences satisfied the conditions of being not only novel but also potentially possessing similar properties to the existing/natural silk for MaSp1, and only three fit the criteria for both the AgSp1 and Flag samples. This level of inconsistency could point to the fact that there is a large degree of inaccuracy in the generative model, potentially caused by an inadequate data set or one\u003c/p\u003e\u003cp\u003ewith skewed data. However, it is more likely that the various types of spider silks across different species are exceptionally diverse. Therefore, it would require much more work to identify and correlate all of the relevant amino acid motifs or structures that make each of them unique and give the types of silk from each species their own distinctive mechanical properties. There will have to be success in the implementation of this new knowledge to the model before it will be able to produce a sequence that is both novel and still the type of spider silk that is desired, along with its corresponding mechanical properties.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e4.2 RMSD Analysis\u003c/h2\u003e\u003cp\u003eThe RMSD of the \u003cem\u003eCyclosa Monticola\u003c/em\u003e MaSp1 sequences indicates that over the period of the 35 nanoseconds the molecular dynamics simulation was run, the model-generated spider silk sequence was much more stable structurally than the natural pre-existing one. The lessened frequency of significant spikes and sudden changes in its RMSD overall as well as a generally lower RMSD score in comparison with its analog indicates that it was closer to its initial shape at the start of the simulation and was more stable. This suggests that the generated amino acid could potentially be able to fold into a simpler shape, thus explaining why it had a much more stable RMSD; whether this shape contributes positively or negatively to its mechanical properties will have to be tested in the future. Furthermore, this finding brings up the possibility of the generative model\u0026rsquo;s use for crafting more stable spider silk sequences, which could facilitate and quicken the process of studying spider silks in the future. However, it is equally likely that the software utilized in this experiment \u0026ndash; either the ColabFold or Molecular Dynamics programs \u0026ndash; could not accurately fold the natural version of the \u003cem\u003eCyclosa Monticola\u003c/em\u003e MaSp1 silk sequence, therefore creating an unstable structure that changed its structure sharply during the simulation.\u003c/p\u003e\u003cp\u003eThe RMSD graph of the AgSp1 silk protein from \u003cem\u003eCyrtarachne Akirai\u003c/em\u003e shows a much closer relationship between the RMSD of the natural and generated sequences, with significant overlap of their respective lines. However, the sudden leap of 5 Angstroms at 1000 frames\u003c/p\u003e\u003cp\u003eindicates instability with the artificial sequence of the AgSp1 protein, the opposite situation compared to the the MaSp1 one. In addition, while the pre-existing \u003cem\u003eCyrtarachne Akirai\u003c/em\u003e silk\u0026rsquo;s RMSD did not spike as often, its steadily increasing RMSD during the latter half of the time frame suggests that it had not reached full equilibrium yet. Compared to the model generated protein, which leveled out after the spike at 1000 frames, the pre-existing silk steadily increased a total of 4 Angstroms after the 1000 frame mark, displaying its continued instability in relation to its artificial analog. This once again brings up the possibility of further study into the generative model, and that if the sequences it generates are consistently more stable, its uses in the further study of silk proteins.\u003c/p\u003e\u003cp\u003eOut of all the types of spider silk proteins, the difference of the RMSD between the natural protein and generated analog of the Flag sample was the most significant. This could have been because the silk chosen for the RMSD calculations was extremely stable compared with other members of that same sample or that natural Flag proteins are generally more stable than the generated versions due to errors in the creation process. To support this, according to the RMSD graphs, the natural Flag silk protein peaked at just 4 Angstroms, while the silk proteins of MaSp1 and AgSp1 peaked at about 18 and 16 Angstroms respectively.\u003c/p\u003e\u003cp\u003eOut of all the types of spider silk proteins, the difference of the RMSD between the natural protein and generated analog of the Flag sample was the most significant. This could have been because the silk chosen for the RMSD calculations was extremely stable compared with other members of that same sample or that natural Flag proteins are generally more stable than the generated versions due to errors in the creation process. To support this, according to the RMSD graphs, the natural Flag silk protein peaked at just 4 Angstroms, while the silk proteins of MaSp1 and AgSp1 peaked at about 18 and 16 Angstroms respectively.\u003c/p\u003e\u003cp\u003eA possible explanation for why the RMSD of both of the Flag proteins were much lower in general than those in the MaSp1 and AgSp1 graphs could be due to the relative shortness of the chosen natural Flag silk sequence, which was at 115 amino acids. Compared to the 170 and 200 amino acid residues of the MaSp1 and AgSp1 silk sequences, respectively, it was significantly shorter. This considerably shortened length could have resulted in a less complex secondary and tertiary structure, and thus less deviation from the initial structure and more stability overall.\u003c/p\u003e\u003cp\u003eOverall, the generative model seems to be the most accurate at mimicking the stability of the 3D-Structure of the artificial silk protein compared to the natural analog for AgSp1, showing the least difference between the two RMSD values. However, the generative model struggles the most with suggesting sequences that possess similar stabilities and deviations\u003c/p\u003e\u003cp\u003efrom the natural spider silk protein in the Flag group, and further research will have to be done with larger sample sizes to establish a more concrete explanation and to improve the efficacy of the model in generating new spider silk sequences. The spider silk proteins from the MaSp1 sample are still somewhat accurate in terms of stability, but there are points where there is considerable deviation between the RMSD values between the pre-existing and artificial silk proteins.\u003c/p\u003e\u003c/div\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eIn this study, the accuracy of a generative model, used to produce novel spider silk sequences, was determined by analyzing the sequences it generated as well as their stabilities and structures. We utilized existing technologies such as NAMD, but also novel ones including ColabFold, which was used to fold the sequences of spider silk, both natural and artificial, as well as VMD, which was used to visualize the structure and dynamics of the various spider silk proteins. In the study, the similarity of eight natural silk sequences and eight generated ones for each sample were calculated, and the RMSD between two of them compared for the MaSp1, Flag, and AgSp1 samples. Ultimately, it was assessed that there was a wide variation and spread in similarity between the natural and artificial silk sequences, with only half or less of all three samples satisfying the conditions of being both novel and being the same type of spider silk as its natural analog. The results from the RMSD were also analyzed, providing insight into the accuracy of the generative model in mimicking the stability of real spider silk proteins. It was established that for the MaSp1 and AgSp1 silk proteins, the artificial versions were generally more stable, with a smaller average RMSD value, while the proteins from Flag possessed significantly lower RMSD values, with the generated analog\u0026rsquo;s structure fluctuating more. It was ultimately determined that while the generative model produced a large spread of sequences, some of which were too similar or different from the natural silks they were based off of, the stability of the generated proteins was to be studied as well as expanded sample sizes of sequences for both sequence analysis and comparison of 3D structures.\u003c/p\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e5.1 Ongoing Work\u003c/h2\u003e\u003cp\u003eDue to the substantial differences and variations between not only the real and generated spider silk proteins, but also between the types of silk examined in the project, a larger sample size is currently being studied, and significantly more proteins are being processed through simulations to gather data on RMSD depending on silk type and to determine what role silk sequence length plays in the RMSD values and structural stability. In addition, the sample sizes will be revised to only include silk proteins from the same genus to minimize variation caused by genetic differences. Finally, due to temporal restrictions and the fact that there is currently no general purpose statistical test that can determine if two time series graphs are significantly different, further work requires the development of a model which can do so for more in-depth quantitative analysis.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e5.2 Future Work\u003c/h2\u003e\u003cp\u003eFor future studies relating to the prototypical generative model, the mechanical properties of the generated spider silks examined in this study should also be tested in the future through simulations, which could possibly bring much more concrete data that we can analyze; testing the properties of the generated spider silks could be much more substantial than simply comparing differences in amino acid sequence, as a protein\u0026rsquo;s sequence is still a somewhat faulty indicator of its properties and structure. Additionally, the generative model utilized in this study should also have to be trained upon a much larger variation of data before future tests of its accuracy. A larger set of data would likely improve its accuracy in detecting amino acid motifs and quirks of each of the different types of spider silks, even when researchers would not have recognized them.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e5.3 Implications and Potential Applications\u003c/h2\u003e\u003cp\u003eIf improved upon enough, generative models like the one used in this project could be, as mentioned before, useful in establishing more concrete relationships between larger sections of amino acids and motifs in the spider silk proteins and their physical properties instead of a couple repeating patterns that are loosely associated with those aspects. This would not only allow us to create and modify existing silk to amplify its outstanding properties, but also allow us to know, in general, more about how to create stronger hybrid materials and mimetics. Long-term, such models would be able to create polypeptide \u0026ldquo;blueprints\u0026rdquo; quickly that we could then produce for many fields, such as medicine, with prosthetics, such as strengthening/replacing tendons, or construction.\u003c/p\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003e6 Acknowledgments\u003c/h2\u003e\u003cp\u003eI am grateful to the Executive Vice President of The Center for Excellence in Education (CEE) Maite Ballestero, and the Director of RSI, mentor, Dr. Mark Kantrowitz, for their continued support. I would like to thank my mentor at the Laboratory of Atomistic and Molecular Mechanics, Professor Markus Buehler, and lab members Amadeus Cavalcanti Salvador de Alcantara and Wei Lu. I would also like to thank Catherine Xue, PhD of bio-statistics at Harvard, who was my tutor and provided feedback on my project\u0026rsquo;s initial drafts. And thanks to the Regeneron Science Talent Search (STS) for the Scholar recognition and awards. Finally, I am very grateful for my generous sponsors and the Department of Defense, as well as CEE, RSI, MIT, and STS for providing me this valuable opportunity, and Awards!\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eEisoldt L, Smith A, Scheibel T (2011) Decoding the secrets of spider silk. \u003cem\u003eMaterials To day\u003c/em\u003e, 14(3):pp. 80\u0026ndash;86. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/S1369-7021(11)70057-8\u003c/span\u003e\u003cspan address=\"10.1016/S1369-7021(11)70057-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. URL \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.sciencedirect.com/science/article/pii/S1369702111700578\u003c/span\u003e\u003cspan address=\"https://www.sciencedirect.com/science/article/pii/S1369702111700578\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBlamaires SJ, Blackledge TA (2017) and T. I-Min. Physicochemical property variation in spider silk: Ecology, evolution, and synthetic production. \u003cem\u003eAnnual Review of Entomology\u003c/em\u003e, 62(1):pp. 443\u0026ndash;460. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1146/annurev-ento-031616-035615\u003c/span\u003e\u003cspan address=\"10.1146/annurev-ento-031616-035615\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. PMID: 27959639\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMorin A, Pahlevan M, Alam P (2017) \u003cem\u003eSilk Biocomposites: Structure and Chemistry\u003c/em\u003e, Chap. 8, pp. 189\u0026ndash;219. John Wiley Sons, Ltd. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/9781119441632.ch8\u003c/span\u003e\u003cspan address=\"10.1002/9781119441632.ch8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. URL \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://onlinelibrary.wiley.com/doi/abs/10.1002/9781119441632.ch8\u003c/span\u003e\u003cspan address=\"https://onlinelibrary.wiley.com/doi/abs/10.1002/9781119441632.ch8\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKeten S, Xu Z, Ihle B et al (2010) Nanoconfinement controls stiffness, strength and mechanical toughness of β-sheet crystals in silk. Nat News\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGu Y, Yu L, Mou J et al (2020) Mechanical properties and application analysis of spider silk bionic material. \u003cem\u003ee-Polymers\u003c/em\u003e, 20(1):pp. 443\u0026ndash;457, doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1515/epoly-2020-0049\u003c/span\u003e\u003cspan address=\"10.1515/epoly-2020-0049\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. URL https://doi.org/10.1515/epoly-2020-0049\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKluge JA, Rabotyagova O, Leisk GG et al (2008) Spider silks and their applications. Cell 26(5):244\u0026ndash;251. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.tibtech.2008.02.006\u003c/span\u003e\u003cspan address=\"10.1016/j.tibtech.2008.02.006\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKhan A, Jawaid M (2020) Novel spider silk fiber for high-performance textiles and eco-friendly armor applications. J Nat Fibers 19(5):1879\u0026ndash;1884. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/15440478.2020.1788496\u003c/span\u003e\u003cspan address=\"10.1080/15440478.2020.1788496\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMilazzo M, Fitzpatrick V, Owens CE et al (2023) 3d printability of silk/hydroxyapatite composites for microprosthetic applications. ACS Biomaterials Sci Eng 9(3):1285\u0026ndash;1295. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1021/acsbiomaterials.2c01357\u003c/span\u003e\u003cspan address=\"10.1021/acsbiomaterials.2c01357\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMalay AD, Arakawa K, Numata K (2017) Analysis of repetitive amino acid motifs reveals the essential features of spider dragline silk proteins. PLoS ONE 12(8). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0183397\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0183397\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJenkins JE, Creager MS, Lewis RV et al (2010) Quantitative correlation between the protein primary sequences and secondary structures in spider dragline silks. ACS Publications 1(1):192\u0026ndash;200. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1021/bm9010672\u003c/span\u003e\u003cspan address=\"10.1021/bm9010672\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMirdita M, Schutze K, Moriwaki Y et al (2022) Colabfold: making protein folding accessible to all. Nat Methods 19:679\u0026ndash;682. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41592-022-01488-1\u003c/span\u003e\u003cspan address=\"10.1038/s41592-022-01488-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePhillips JC, Hardy DJ, Maia JD et al (2020) Scalable molecular dynamics on cpu and gpu architectures with namd. J Chem Phys 153(4). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1063/5.0014475\u003c/span\u003e\u003cspan address=\"10.1063/5.0014475\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHumphrey W, Dalke A, Schulten K (1996) Vmd: visual molecular dynamics. J Mol Graph 14(1):33\u0026ndash;38. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/0263-7855\u003c/span\u003e\u003cspan address=\"10.1016/0263-7855\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eArakawa K, Kong N et al (2022) 1000 spider silkomes: Linking sequences to silk physical properties. ScienceAdvances 8(41). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1126/sciadv.abo6043\u003c/span\u003e\u003cspan address=\"10.1126/sciadv.abo6043\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBlamaires SJ, Spicer PT (2020) and P. J. Flanagan. Spider silk biomimetics programs to inform the development of new wearable technologies. Frontiers, 7\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAn B, Jenkins JE, Sampath S et al (2012) Reproducing natural spider silks\u0026rsquo; copolymer behavior in synthetic silk mimics. Biomacromolecules 13(12):3938\u0026ndash;3948. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1021/bm301110s\u003c/span\u003e\u003cspan address=\"10.1021/bm301110s\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eM. J. Buehler. Generative pretrained autoregressive transformer graph neural network applied to the analysis and discovery of novel proteins. arXiv, (2023) 2305.04934.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eM. J. Buehler. Generative model for conditioned silk sequences. Obtained from the Laboratory for Atomistic and Molecular Mechanics at MIT\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003edifflib \u0026mdash; helpers for computing deltas URL \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://docs.python.org/3/library/ difflib.html#module-difflib\u003c/span\u003e\u003cspan address=\"https://docs.python.org/3/library/ difflib.html#module-difflib\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNi B, Kaplan DL (2023) and M. J. Buehler. Generative design of de novo proteins based on secondary-structure constraints using an attention-based diffusion model. \u003cem\u003eChem\u003c/em\u003e, 9(7):pp. 1828\u0026ndash;1849. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.chempr.2023.03.020\u003c/span\u003e\u003cspan address=\"10.1016/j.chempr.2023.03.020\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. URL \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.sciencedirect.com/science/article/pii/S2451929423001390\u003c/span\u003e\u003cspan address=\"https://www.sciencedirect.com/science/article/pii/S2451929423001390\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQuan L, Wu T, Lyu Q (2023) Computational de novo protein design: From secondary to primary, then toward tertiary structures. \u003cem\u003eChem\u003c/em\u003e, 9(7):pp. 1625\u0026ndash;1627. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.chempr.2023.06.010\u003c/span\u003e\u003cspan address=\"10.1016/j.chempr.2023.06.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. URL \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.sciencedirect.com/science/article/pii/S2451929423003157\u003c/span\u003e\u003cspan address=\"https://www.sciencedirect.com/science/article/pii/S2451929423003157\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePearson WR (2013) An introduction to sequence similarity (homology) searching. Curr Protocols 42(7). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/0471250953.bi0301s42\u003c/span\u003e\u003cspan address=\"10.1002/0471250953.bi0301s42\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[{"identity":"3cf789ee-f749-4287-9556-d3be35676b76","identifier":"10.13039/100006919","name":"Massachusetts Institute of Technology","awardNumber":"50000","order_by":0}],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Massachusetts Institute of Technology","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":"Spider silk, mechanical properties, amino acid sequence, protein structure, generative model, MaSp1, AgSp1, synthetic silk","lastPublishedDoi":"10.21203/rs.3.rs-7926039/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7926039/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSpider silk\u0026rsquo;s exceptional mechanical properties of extensibility, tensile strength, and Young\u0026rsquo;s Modulus originate from its unique amino acid sequence as well as its 3D-structure, consisting of a mix of amorphous, flexible regions, and tough, crystalline portions of beta sheets held together by hydrogen bonds. These outstanding physical qualities have made it an ideal material for use in fields such as clothing, body armor, and medicine. However, while there is much data on a variety of spider silk types (MaSp AgSp, Flag), including information on their amino acid sequences and mechanical properties, it has been difficult to quantitatively utilize amino acid sequence or protein structure to reliably predict attributes of spider silk or to replicate them as synthetic silks. Utilizing a prototypical generative model trained on a dataset of existing spider silk sequences, novel spider silks were generated that mimicked natural spider silk sequences in the MaSp1, AgSp1, and Flag classifications of silk. In order to assess the accuracy of the model, the sequences were compared, and their 3D-Structures were analyzed utilizing Root Mean Squared Deviation. The results of this study suggest that the model is inconsistent and produces sequences with a high variability of sequences and structures, meaning that significant improvements will have to be made.\u003c/p\u003e","manuscriptTitle":"Assessing the Accuracy of a Generative Model for Creating Novel Spider Silk Proteins","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-25 17:11:22","doi":"10.21203/rs.3.rs-7926039/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":"29a5076e-8428-4c53-956b-33c5b74b04fd","owner":[],"postedDate":"October 25th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":56729826,"name":"Biotechnology and Bioengineering"}],"tags":[],"updatedAt":"2025-10-25T17:11:22+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-25 17:11:22","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7926039","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7926039","identity":"rs-7926039","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}