# Fileset

[1-s2.0-S2666389923002258-main.pdf](https://mdr.nims.go.jp/filesets/b25bf4da-3883-4314-8250-7bff5304d85a/download)

## Creator

Weilin Yuan, [Yusuke Hibi](https://orcid.org/0000-0003-4006-1070), [Ryo Tamura](https://orcid.org/0000-0002-0349-358X), Masato Sumita, [Yasuyuki Nakamura](https://orcid.org/0000-0003-0078-6413), [Masanobu Naito](https://orcid.org/0000-0001-7198-819X), [Koji Tsuda](https://orcid.org/0000-0002-4288-1606)

## Rights

[Creative Commons BY Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/)

## Other metadata

[Revealing factors influencing polymer degradation with rank-based machine learning](https://mdr.nims.go.jp/datasets/563a06f0-5aaf-4276-80ba-d0b23e7e3947)

## Fulltext

Revealing factors influencing polymer degradation with rank-based machine learningArticleRevealing factors influencing polymer degradationwith rank-based machine learningGraphical abstractHighlightsd Existing datasets of polymer degradability are small andoften incompatibled A rank-based method for integrating diverse polymerdegradability datasets is presentedd Key molecular factors that determine degradability areidentifiedYuan et al., 2023, Patterns 4, 100846December 8, 2023 ª 2023 The Authors.https://doi.org/10.1016/j.patter.2023.100846AuthorsWeilin Yuan, Yusuke Hibi, Ryo Tamura,Masato Sumita, Yasuyuki Nakamura,Masanobu Naito, Koji TsudaCorrespondencetamura.ryo@nims.go.jp (R.T.),naito.masanobu@nims.go.jp (M.N.),tsuda@k.u-tokyo.ac.jp (K.T.)In briefFactors influencing polymer degradabilityare identified via machine learning-basedintegrated analysis. This approach opensa way to large-scale analysis of polymerdegradability, which has been hamperedby the extremely diverse measurementconditions of existing datasets.llmailto:tamura.ryo@nims.go.�jpmailto:naito.masanobu@nims.go.�jpmailto:tsuda@k.u-tokyo.ac.�jphttps://doi.org/10.1016/j.patter.2023.100846OPEN ACCESSPlease cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846llArticleRevealing factors influencing polymerdegradation with rank-based machine learningWeilin Yuan,1 Yusuke Hibi,2 Ryo Tamura,1,3,4,* Masato Sumita,4 Yasuyuki Nakamura,2 Masanobu Naito,2,*and Koji Tsuda1,3,4,5,*1Graduate School of Frontier Sciences, The University of Tokyo, 5-1-5 Kashiwanoha, Kashiwa, Chiba 277-8561, Japan2Research Center for Macromolecules and Biomaterials, National Institute for Materials Science, 1-2-1 Sengen, Tsukuba, Ibaraki305-0047, Japan3Center for Basic Research on Materials, National Institute for Materials Science, 1-1 Namiki, Tsukuba, Ibaraki 305-0044, Japan4RIKEN Center for Advanced Intelligence Project, 1-4-1 Nihonbashi, Chuo-ku, Tokyo 103-0027, Japan5Lead contact*Correspondence: tamura.ryo@nims.go.jp (R.T.), naito.masanobu@nims.go.jp (M.N.), tsuda@k.u-tokyo.ac.jp (K.T.)https://doi.org/10.1016/j.patter.2023.100846THE BIGGER PICTURE In view of marine environmental protection, degradation control of polymers is anincreasingly important topic. Existing datasets about polymer degradability are small and diverse in mea-surement conditions, which makes direct comparison across datasets notoriously difficult. Improved ma-chine learning methods that better leverage limited and incompatible data, such as the RankSVM-basedmethod described in this work, may help chemists develop new and better polymers.Development/Pre-production:Data science output has beenrolled out/validated across multiple domains/problemsSUMMARYThe efficient treatment of polymer waste is amajor challenge for marine sustainability. It is useful to reveal thefactors that dominate the degradability of polymer materials for developing polymer materials in the future.The small number of available datasets on degradability and the diversity of their experimental means andconditions hinder large-scale analysis. In this study, we have developed a platform for evaluating the degrad-ability of polymers that is suitable for such data, using a rank-based machine learning technique based onRankSVM. We then made a ranking model to evaluate the degradability of polymers, integrating three data-sets on the degradability of polymers that are measured by different means and conditions. Analysis of thisranking model with a decision tree revealed factors that dominate the degradability of polymers.INTRODUCTIONIn view of marine environmental protection, degradation controlof polymers is an increasingly important topic.1,2 Plastic waste isa major problem because it takes time to degrade in the ocean.3Over a long time, plastic waste is gradually degraded to smallersizes by waves, wind, and some environmental factors thatcontribute to the molecular level.4 In particular, microplasticssmaller than 5 mm are known to cause a significant impact onthe ecosystem and are a major social problem.5 Therefore, thedevelopment of polymers with controllable degradability isimportant for protecting the marine environment.To develop polymers with controllable degradability, it isnecessary to consider the degradation at the chemical level.From the chemical viewpoint, three important degradationThis is an open access article undmechanisms are known.4 The first is photo-degradation by ultra-violet (UV) light. Impurities, heterostructures, or the polymer itselfare photodegraded, leading to a decrease in molecular weightand cracking.6 The second is hydrolysis in which polymers aredegraded by water. It is known that the higher temperaturecauses a shorter degradation time.6 The third is biodegradation.Bacteria attach to the surface of polymers and form biofilms,which accelerate degradation.7 In the ocean, a complex inter-play of these factors leads to the degradation of polymers.Therefore, it is difficult to conduct experiments in which thesefactors are uniquely controlled, and many of the data reportedon degradability were highly dependent on environments.In recent years, machine learning has been used to analyzepolymer data to extract factors that contribute to physical prop-erties.8 However, applying machine learning to analyze thePatterns 4, 100846, December 8, 2023 ª 2023 The Authors. 1er the CC BY license (http://creativecommons.org/licenses/by/4.0/).mailto:tamura.ryo@nims.go.jpmailto:naito.masanobu@nims.go.jpmailto:tsuda@k.u-tokyo.ac.jphttps://doi.org/10.1016/j.patter.2023.100846http://creativecommons.org/licenses/by/4.0/Table 1. Summary of literature dataset from Min et al., in which the ranking was created by weight loss per dayName Abbreviation SMILES Wt. loss (%/day)Poly(ethylene terephthalate)14 PET *CCOC(=O)c1ccc(cc1)C(=O)O* 0.0003Poly(lactic acid)15 PLA *OC(C)C(=O)* 0.00033Poly(L-lactic acid)16 PLLA *C([C@H](O*)C) = O 0.0014Polycaprolactone14 PCL *CCCCCC(=O)O* 0.0027Polystyrene17 PS *CC(*)c1ccccc1 0.0111Polyurethane18 PU *C(=O)NC1 = CC = C(CC2 = CC = C(NC(=O)OCCO*)C=C2)C=C10.0116Nylon619 Nylon6 *NCCCCCC(*) = O 0.0222Polycarbonate20,21 PC *Oc1ccc(C(C)(C)c2ccc(OC(*) = O)cc2)cc1 0.0238Poly(butylene succinate)22 PBS *OCCCCOC(=O)CCC(*) = O 0.0714Nylon6619 Nylon66 *NCCCCCCNC(=O)CCCCC(*) = O 0.0778Poly(ethylene sebacate)23 PESeb *OCCOC(=O)CCCCCCCCC(*) = O 0.0894Poly(butylene sebacate)23 PBSeb *OCCCCOC(=O)CCCCCCCCC(*) = O 0.114Poly(butylene azelate)23 PBAz *OCCCCOC(=O)CCCCCCCC(*) = O 0.2011Poly(3-hydroxybutyrate)24 P3HB CC(CC(*) = O)O* 0.2917Poly(ethylene azelate)23 PEAz *OCCOC(=O)CCCCCCCC(*) = O 0.3073Poly(butylene adipate)22 PBAdip *OCCCCOC(=O)CCCCC(*) = O 0.3929Nylon425 Nylon4 *NCCCC(*) = O 1.4286Poly(propylene succinate)23 PPS *OCCCOC(=O)CCCCCCC(=O)* 2.2405Poly(vinyl alcohol)26 PVA C(C(O)*)* 2.8333Poly(propylene azelate)23 PPAz *OCCCOC(=O)CCCCCCCC(=O)* 7.4969Poly(propylene sebacate)23 PPSeb *OCCCOC(=O)CCCCCCCCC(*) = O 7.5642Poly(propylene adipate)27 PPAd *OCCCOC(=O)CCCCC(=O)* 61.3197Poly(propylene glutarate)27 PPGI *OCCCOC(=O)CCCC(=O)* 100Poly(propylene pimalate)28 PPPIM *OCCCOC(=O)CCCCCC(*) = O 100llOPEN ACCESS ArticlePlease cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846degradability does not facilitate the analysis because the strongenvironmental dependence of the data inhibits the direct com-parison of them. To circumvent this obstacle, we used ranking-based machine learning9,10 to integrate datasets measured indifferent environments and have built a unified ranking of de-gradability through RankSVM,11,12 which can learn the pairwisepreference. After training RankSVM, degradability scores of allpolymers have been obtained. Based on this score, we haveanalyzed the important factors that dominate the degradabilityof polymers with tree-based analysis. Furthermore, using theconstructed RankSVM model, we have predicted the rankingof a polymer database (4,577 polymers) where the values ofthe degradability of polymers are not measured.RESULTS AND DISCUSSIONDegradation datasetsTo demonstrate the effectiveness of our method, we made a da-taset consisting of 15 polymers we measured (experimental da-tasets 1 and 2), and 24 polymers were obtained from the litera-ture,13 as the literature dataset, summarized in Table 1. Surely,the former and latter datasets are measured in different condi-tions. For preparing our dataset, seven polymer films were pro-duced by drying the homopolymer solutions on a glass slide asexperimental dataset 1. Each polymer film is placed in a centri-fuge tube and utterly immersed in artificial seawater. The sam-ples were placed on the roof of the National Institute forMaterials2 Patterns 4, 100846, December 8, 2023Science (Tsukuba, Japan) and applied to the exposure experi-ments. The results of degradation in experimental dataset 1are summarized in Table 2. Polymer degradability wasmeasuredvia the concentration of total organic carbon (TOC) in the artificialseawater after the exposure experiment. In detail, the degrad-ability dwas evaluated by Equation 6 from the weight of the poly-mers Wfilm, the surface area of the polymers Sfilm, the volume ofseawater Vwater, and the proportion of carbon elements Mc%.For the other dataset, eight commercial homopolymer film prod-ucts with uniform thickness were prepared as experimental da-taset 2, which are summarized in Table 3. For experimental data-sets 1 and 2, since the exposure experiments were conductedunder natural conditions, UV exposure and temperaturechanges are the main factors for the degradation.Unified ranking integrating three datasetsThe RankSVM model to predict degradability score was trainedby using the pairwise preferences prepared from each dataset.Note that preferences cannot be obtained between different da-tasets. In the RankSVM, there is a hyperparameter C. The valueof C is determined so that the prediction accuracy for prefer-ences, which is evaluated by cross-validation, is maximized.Here, 5-fold cross-validation is performed, and the accuracy de-pending onC is shown in Figure S1. When C = 0:007, the accu-racy reached a maximum value of 0.85, confirming the high pre-diction accuracy for the preferences. The RankSVM model wastrained using all preferences when C = 0:007, and the unifiedTable 2. Results of degradation in experimental dataset 1Name Abbreviation SMILES Wfilm [mg]TOC[mg/L] Vwater [mL] Mc [%] Sfilm [cm2] dPoly(isopropyl acrylate) PIPA *CC(*)C(=O)OC(C)C 34.1 0.5705 62 63.16 7.48 0.0002196Poly(isodecyl acrylate) PIDA *CC(*)C(=O)OCCCCCCCC(C)C16.1 0.5347 57 73.58 8.46 0.0003041Poly(benzyl acrylate) PBA *CC(*)C(=O)OCc1ccccc1 19.4 0.749 47 74.07 6.63 0.0003695Poly(2-methoxyethylacrylate)PMEA *CC(*)C(=O)OCCOC 12.6 0.5426 68 55.38 12.60 0.0004196Poly(2-butoxyethylacrylate)PBEA *CC(*)C(=O)OCCOCCCC 9.9 0.7745 50 62.79 14.00 0.0004450Poly(vinyl butyral) PVB *CC1CC(*)OC(CCC)O1 11 0.6276 50 67.61 7.20 0.0005861Poly(hexamethylenesebacate)PHMS *CCCCCCOC(=O)CCCCCCCCC(=O)O*8.8 0.4898 76 67.61 9.69 0.0006457llOPEN ACCESSArticlePlease cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846ranking was created by arranging the polymers in order of theirdegradability scores d, which are calculated by Equation 5.The unified rankingmade by RankSVM and individual rankingsof each dataset are shown in Figure 1. It is noteworthy that therankings are consistent, even though they are prepared viadifferent measurement conditions and experiments. Experi-mental dataset 2 contains many acrylate polymers, which are re-garded as undegradable polymers in the unified ranking. Thepolymers in experimental dataset 2 are scattered over the wholerange of the unified ranking. PS and PC, the common polymersof the literature dataset and experimental dataset 2, lie at thereasonable ranking in the unified ranking. In other words, PC ismore degradable than PS. This fact indicates that the unificationof different datasets has succeeded, and the reliability of unifiedranking is expected to be elevated by adding the datasets.Revealing factors influencing polymer degradationTo extract the factors that dominate the degradability of a poly-mer, we made a regression tree, as shown in Figure 2, with ninemolecular properties as input and the degradability score d asoutput. Here, we usedmolecular properties, including informationon chemical composition, which are familiar factors for syntheticchemists. According to this tree, the numbers of ester groups,alkyl carbons, hydroxyl groups, ether groups, and benzene ringsare selected as key descriptors. On the other hand, the numbersTable 3. Results of degradation in experimental dataset 2Name Abbreviation SMILES WPolystyrene PS *CC(*)c1ccccc1 1.Polypropylene PP *CC(*)C 0.Polyacetal POM *CO* 1.Poly(tetrafluoroethylene) PTFE *C(*)(F)F 0.Polycarbonate PC *Oc1ccc(C(C)(C)c2ccc(OC(*) = O)cc2)cc11.Poly(methylmethacrylate)PMMA *CC(*)(C)C(=O)OC 1.Poly(ethylenenaphthalate)PEN *CCOC(=O)c1ccc2cc(C(=O)O*)ccc2c11.Polychloroprene CR *CC/C=C(/*)Cl 6.of the amide group, carbonates, hetero atoms, and urethane arenot selected as key descriptors. In this tree, the number of leaf no-des is 10, and they are classified into three categories, ‘‘undegrad-able’’ (d% 0:303), ‘‘middle’’ (0:421%d%0:654), and ‘‘degrad-able’’ (dR 0:716), according to the average degradability scoresof polymers in each category, denoted by d.The prediction result of the decision tree mainly shows thefollowing tendency. (1) The polymers with more ester groupstend to be more degradable (node 1). (2) The polymers with ben-zene rings tend to have low degradability (nodes 3 and 7). (3) Thepolymers with more alkyl carbons tend to be more undegradable(node 6). Concerning (1), the polymers with ester groups areknown to be susceptible to alkaline environments of marine andmicrobial action.28 The decision tree successfully captured thesecharacteristics. Our analysis with the decision tree also repro-duces the characteristics. For (3), in general, more alkyl carbonsin monomer structures are known to promote crystallization,which tends to elevate the glass transition temperature (Tg).29Higher crystallinity and Tg would be favorable for reducing de-gradability.30 On the other hand, node 5 shows the opposite trendto node 6. The branching at node 5 can be considered as the pres-ence or absence of benzene rings, which happens to match thefactor for a number of alkyl carbons. Then, node5 extracts the fac-tor that the polymers without benzene rings are more degradable,which is consistent with the fact extracted as (2).film [mg]TOC[mg/L] Vwater [mL] Mc [%] Sfilm [cm2] d703 0.9217 32 0.05778 8.8 0.0065669463 0.8555 30 0.12047 6 0.02008071 0.3997 33 0.21463 10.08 0.021299634 0.2400 39 0.21141 7.6 0.02782804 0.7551 36 0.19936 7.02 0.0284128 0.5993 29 0.1631 3.8 0.04292565 0.6936 25 0.48537 2.31 0.2101878 0.5421 37 4.49634 9.66 0.4655Patterns 4, 100846, December 8, 2023 3Figure 1. Rankings of polymer degradabilityThe degradability scores in the unified ranking arederived from the RankSVM trained with the inte-gration of the literature dataset and two experi-mental datasets. In the literature dataset, 24 dataare contained from Min et al.13 For experimentaldatasets 1 and 2, seven polymer films were pro-duced, and eight commercial homopolymers wereprepared, respectively. The common polymers ofthe literature dataset and experimental dataset 2 arePS and PC.llOPEN ACCESS ArticlePlease cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846Contrary to factor (2), node 8 is a branch for classifying thepolymers with more benzene rings to degradable ones (likePET [benzene ring = 1] and PEN [benzene ring = 2]). PEN hastwo benzene rings, but these form a naphthalene ring. Thus,the degradation mechanism of PEN would be different fromthat by polymers with some independent benzene rings suchas PU and PC, and node 8 is not directly related to the mecha-nism of (2). Let us consider the difference between the degrada-tion mechanisms of PET and PEN. First, since PEN has higher Tgthan PET, PEN would have higher thermal stability.31 On theother hand, the higher Tg causes degradation in the PEN filmdur-ing high-temperature processing,32 and it is difficult to comparethem from their values of Tg. Second, PEN is susceptible tophoto-oxidative degradation under UV light conditions, whichcauses yellowing and gel formation on the surface. In addition,it is known that PET exhibits photo-stability higher than PEN.33Therefore, in terms of photo-degradation, PET would be rankedas more undegradable than PEN in the unified ranking. Note thatthese data come from different incompatible datasets, and thisnode is extracted by integrating some datasets.Interestingly, the decision tree shows that PVA categorized indegradable 3 is an exceptional degradable polymer despite theabsence of ester groups. We speculate that the impurity in PVAlargely affects the degradability of PVA. PVA is usually synthe-sized through hydrolysis of polyvinyl acetate resulting in impu-4 Patterns 4, 100846, December 8, 2023rities in PVA.34 Hence, the sample shouldinclude some impurities, and measuringthe accurate degradability of pure PVA isdifficult. The decision tree also classifiedPET in undegradable 3 as an undegradablepolymer in spite of its ester groups.We guess that the stabilization effect ofthe p-stacked structure can significantlyreduce the degradability even if the mono-mer structure contains two ester groups.On the other hand, it is also known thatthe degradability of PET increases withincreasing the fraction of the ester group.35Predictions of degradable orundegradable polymers in PoLyInfoThe degradability scores of polymersrecorded in PoLyInfo36 are predicted byusing our model. We obtained 17,771SMILES data from PoLyInfo, and 4,577polymers were extracted by consideringthe applicability domain. Here, the applicability domain is deter-mined by a k-nearest neighbors-based approach method.37 Thetop 10 undegradable and degradable polymers as prediction re-sults are shown in Tables 4 and 5. It can be found that the modeltends to predict that the acrylate polymers and polymers with anamide group are more undegradable. The polymers with anamide group show certain stability in a water environment andusually have high Tg due to the hydrogen bonds effect.38,39 Onthe other hand, the predicted degradability is widely distributed,even if they have the same functional group (see Figure S2), sinceit is not predicted only from certain functional groups. Therefore,the polymers shown in Tables 4 and 5 are ranked high not only byfunctional groups but also by other factors such as number ofatoms and types of neighbor atoms.The top 10 degradable polymers almost all contain a hydroxylgroup. On the other hand, the position of a hydroxyl group largelyinfluences the polymer degradability, and themechanism shouldbe different between the degradation of the side chain via hydro-lysis and that of the polymer backbone. The position of a hydroxylgroup largely influences the polymer degradability. However, it iscurrently difficult to distinguish between these degradationmechanisms because the training data do not contain sufficientdata. Since SMILES contains information on the polymerizationpoint, the Mol2vec descriptor can distinguish whether hydroxylgroups are on the side chains or on the polymer backbone.Degradable 1Degradable 3Undegradable 2Undegradable 36Alkyl carbons < 67Benzene rings < 18Benzene rings < 2 Sample size: 13  d = 0.8145Alkyl carbon < 6 3Benzene rings < 14Hydroxyl group < 12Ester groups < 11Ester groups < 2 TrueFalseDegradableMiddleUndegradableMiddle 2Degradable 2Middle 39Ether groups < 2 Middle 4Middle 1Undegradable 1Sample size: 1  d = 0.716Sample size: 1  d = 0.303Sample size: 1  d = 0.865Sample size: 6  d = 0.654Sample size: 2  d = 0.529Sample size: 3  d = 0.493Sample size: 7  d = 0.421Sample size: 2  d = 0.293Sample size: 1  d = 0.248Sample size: 37  d = 0.614Sample size: 22  d = 0.505Sample size: 10  d = 0.378Sample size: 12  d = 0.611Sample size: 15  d = 0.774Sample size: 9  d = 0.393Sample size: 11  d = 0.588Sample size: 2d = 0.51Sample size: 8  d = 0.623Rank label Undegradable1 Undegradable2 Undegradable3 Middle1 Middle2 Middle3 Middle4 Degradable1 Degradable2 Degradable30.248 0.293 0.303 0.421 0.493 0.529 0.654 0.716 0.814 0.865Polymer PBA PIDA, PBEA PETPLA, PLLA, PCL,P3HB, PIPAPMEA, PMMAPU, PS, PC PVB, POMnylon6, nylon66,nylon4, PP,PTFE, CRPENPBS, PESeb,PBSeb, PBAZ,PEAz, PBAdip,PPS, PPAz,PPSeb, PPAd,PPGI, PPPIM,PHMSPVAFigure 2. Regression tree trained using a degradability score derived from the RankSVM modelNine molecular properties of each polymer were used as descriptors, and five key properties were selected for regression tree construction. The polymers wereclassified into three categories according to average degradability score d (d% 0:303: ‘‘undegradable,’’ 0:421%d%0:654: ‘‘middle,’’ and dR 0:716:‘‘degradable’’).llOPEN ACCESSArticlePlease cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846Thus, we will be able to distinguish it by conducting experimentson the polymers shown in Table 5. In addition, in the training data,there are no polymerswithmultiple hydroxyl groups, and the pre-dicted degradability for such polymers would be unreliable. Toavoid various limitations for prediction, using these prediction re-sults, it is necessary to increase the amount of experimental data.ConclusionsIn conclusion, we have presented a platform that can evaluatepolymer degradability based on incompatible datasets throughrank-basedmachine learning. In this study, three types of degra-dation datasets have been prepared. See Tables 6 and 7 andFigure 3 for experimental details. For each dataset, we have eval-uated preferences, and the RankSVMmodel has been trained topredict preferences. Based on the degradability score obtainedby a trained RankSVM model, a unified ranking has been con-structed. To reveal effective factors for the polymer degradation,the decision tree has been trained. The characteristic of this de-cision tree is that the input is molecular properties, and theoutput is the degradability score, which is evaluated by inte-grating incompatible datasets. For input molecular properties,we have used chemical compositions such as the numbers ofPatterns 4, 100846, December 8, 2023 5Table 4. Prediction results of top 10 undegradable polymers in PoLyInfo, where the name of the polymer, degradability scored obtained by RankSVM, and chemical structure are shownName Degradability score Chemical structurePoly[1-(1-{[1-(ethoxycarbonyl)ethoxy]carbonyl}ethyl)pyrrole-2,5-diyl]�0.296575962Poly(N,N-diisopropylacrylamide) �0.210303998Poly(isopropyl crotonate) �0.108607135Poly{bis[(4-(methoxycarbonyl)phenoxy]phosphazene}�0.052889856Poly(diisopropyl fumarate) �0.024667207Poly[4-(isopropoxycarbonyl)styrene] �0.020469478Poly(4-isopropoxystyrene) �0.01198865(Continued on next page)llOPEN ACCESS Article6 Patterns 4, 100846, December 8, 2023Please cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846Table 4. ContinuedName Degradability score Chemical structurePoly(isopropyl acrylate) 0.00000003Poly{1-[(dioctylamino)carbonyl]ethylene} 0.033141513Poly{(butane-1,4-diamine)-alt-[4,4’-(propane-1,2-diyldioxy)dibenzoic acid]}0.036542042llOPEN ACCESSArticlePlease cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846ester groups, benzene rings, and alkyl carbons, which arefamiliar to synthetic chemists. From the decision tree, weconclude that the learning of rank-based models works wellbecause the well-known factors in polymer science are ex-tracted. On the other hand, there are no restrictions on the prop-erties as input for the decision tree, and morphology descriptorssuch as distance between rings and length of the main chain willbe target factors that are easy to understand. Although there isstill ambiguity and lack of knowledge on the factors of monomerstructure that influence the degradability of polymers, the predic-tion results are considerably consistent with existing knowledge.Because a higher accurate unified ranking helps to deepen theunderstanding of the polymer degradability, further experimentswill be applied to increase the data and improve the accuracy ofthe prediction model as future prospects.As discussed in Jablonka et al.,40 large-scale collection ofchemical data has the potential to change the way of materialdiscovery. The unified ranking we have shown in this study is stillsmall, but technically, it is possible to construct a ranking for mil-lions of polymers. In the future, our data integration approach willbe improved by a growing amount of collective data and willgrow into an important resource that chemists can easily referto and use for developing polymers.EXPERIMENTAL PROCEDURESResource availabilityLead contactFurther information and requests should be directed to and will be fulfilled bythe lead contact, Koji Tsuda (tsuda@k.u-tokyo.ac.jp).Materials availabilityThis study did not generate any new unique materials.Data and code availabilityAll original code has been deposited at Zenodo under the https://doi.org/10.5281/zenodo.8268022 and is publicly available as of the date of publication.All polymer degradability datasets used in this paper and the predicted de-gradability scores of PoLyInfo polymers are deposited at Zenodo under thehttps://doi.org/10.5281/zenodo.8268022 and are publicly available as of thedate of publication.Calculation details in RankSVMIn RankSVM, each dataset is converted to pairwise preferences by using mo-lecular fingerprints of polymer 1 (xð1Þi ) and polymer 2 (xð2Þi ) for comparison andsummarized into one training set,D =n�xð1Þi ; xð2Þi�; tioMi = 1; (Equation 1)ti =�1; if the degradability of polymer 1 is larger than polymer 2� 1;otherwise;where M is the number of total preferences. We employed 300-dimensionalMol2vec as the molecular fingerprint.41 RankSVM uses the following predic-tion function:y = wT�xð1Þi � xð2Þi�: (Equation 2)The parameterw is estimated by solving the following optimization problemsuch that the predicted results are maximally consistent with the labels ti :½w�; x�� = argminw;xkwk+CXMi = 1xi ; (Equation 3)tihwT�xð1Þi � xð2Þi�iR 1 � xi ði = 1;.;MÞ; (Equation 4)where xi is a slack parameter, and C is a trade-off parameter as hyperpara-meter. After training RankSVM, we can compute the degradability score das the following for any polymer:d = ðw�ÞTx (Equation 5)Through Equation 5, any polymer from the training data or from other data-bases such as PolyInfo36 can be included in the unified ranking if only the mo-lecular fingerprint can be generated.In our implementation, Python script degradability_ranking.py integrates thedegradability datasets and outputs the unified ranking shown in Figure 1. ThePatterns 4, 100846, December 8, 2023 7mailto:tsuda@k.u-tokyo.ac.jphttps://doi.org/10.5281/zenodo.8268022https://doi.org/10.5281/zenodo.8268022https://doi.org/10.5281/zenodo.8268022Table 5. Prediction results of top 10 degradable polymers in PoLyInfo, where the name of the polymer, degradability score d obtainedby RankSVM, and chemical structure are shownName Degradability score Chemical structurePoly{1-[3-(4-benzoyl-3-hydroxyphenoxy)-2-hydroxypropoxycarbonyl]-1-methylethylene}1.971466427Poly[3,3-bis(hydroxymethyl)oxetane] 1.822278633Poly{[4,5-bis(hydroxymethyl)cyclopentane-1,3-diyl]ethene-1,2-diyl}1.822115257Poly(glyceryl methacrylate) 1.817512141Poly[iminoglutarylimino(3,30-dihydroxybiphenyl-4,40-diyl)] 1.780371805Poly(1,4-dihydroxy-1-methylbutane-1,4-diyl) 1.637276638Poly(3-chloro-2-hydroxypropyl methacrylate) 1.611935588Poly{[2,2-dimethyl-4,5-bis(hydoxymethyl)-1,3-dioxolane]-alt-(diethyl carbonate)}1.603114876Poly{1-[2-(5-bromobenzofuran-2-yl)-2-oxoethoxycarbonyl]-1-methylethylene}1.559941678Poly{2-hydroxy-5-[1-(4-hydroxyphenyl)-1-methylethyl]-1,3-phenylene}1.548126408llOPEN ACCESS Article8 Patterns 4, 100846, December 8, 2023Please cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846Table 6. Production companies, product number, composition, and shape of raw materials used in experimental dataset 1Name Company Product number Composition ShapePIPA Scientific Polymer Products Inc.,New York, USA#475 poly(isopropyl acrylate): 25%–30%,toluene: 70–75%liquidPIDA Scientific Polymer Products Inc.,New York, USA#875 poly(isodecyl acrylate):25%–30%,toluene: 70%–75%liquidPBA Scientific Polymer Products Inc.,New York, USA#883 poly(benzyl acrylate): 30%–35%,toluene: 65–70%liquidPMEA Scientific Polymer Products Inc.,New York, USA#891 poly(2-methoxyethyl acrylate):20%–25%, toluene: 75%–80%liquidPBEA Scientific Polymer Products Inc.,New York, USA#896 poly(2-butoxyethyl acrylate):18%–22%, toluene: 78%–82%liquidPVB Scientific Polymer Products Inc.,New York, USA#043 poly(vinyl butyral): 96.0%, 1,1-diethoxybutane: %2.0%, water: %2.0%powderllOPEN ACCESSArticlePlease cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846input files for this script are literature.xlsx, exp1.xlsx, and exp2.xlsx, each ofwhich contains the SMILES strings of polymers and their degradability values.The rankings shown in Tables 4 and 5 were obtained using main.py.MaterialsThe rawmaterials for polymer film in experimental dataset 1 are summarized inTable 6. The commercially available polymer films in experimental dataset 2are summarized in Table 7. The polymer films were cut into small pieces andwashed with deionized water. The artificial seawater was prepared by dissolv-ing 18 g commercial powder of artificial seawater (Tomita Pharmaceutical Co.,Tokushima, Japan) into 500 g of deionized water.Film preparations in experimental dataset 1The procedure of film preparation for the polymer powders is as follows: PVBand PHMS are polymer powders. The polymer powder 100 mg was dissolvedin toluene 900 mg (Wako Pure Chemical Industries, Osaka, Japan). The result-ing 10 wt% solutions were cast on the glass slide, which was in advance tem-plated with Teflon tape to refine the film area (see Figure 3). After drying thecasted solution in the air for an hour, the Teflon tape was removed from theglass slide. The polymer film was further dried in the vacuum oven at 100�Cto completely remove the solvents. The PIPA, PIDA, PBA, PMEA, and PBEAare the samples as toluene solutions. For these samples, the same procedureswith polymer powders were applied for film preparation.Degradation experimentsIn an open-top glass container, the prepared films and commercially availablepolymer films were completely immersed in the artificial seawater. The top ofthe container was sealed with thin plastic wrap. The container was left on theroof of the National Institute for Materials Science (Tsukuba, Japan) for around20–30 days (experimental dataset 1: 7/30/2021–8/19/2021 and experimental da-taset 2: 11/5/2021–12/2/2021). After the degradation, the filmwas removed fromTable 7. Production companies and product number of polymersin experimental dataset 2Name Company Product numberPS Hikari Co., Ltd., Osaka, Japan PS2031-1PP Artec Co., Ltd., Osaka, Japan 20511POM ESCO Co., Ltd., Osaka, Japan EA441MD-0.3PTFE Chukoh Chemical Industries,Ltd., Tokyo, JapanASF-110 FRPC Hikari Co., Ltd., Osaka, Japan KPAC2005-1PMMA Asahi Kasei Co., Ltd., Tokyo, Japan K120913PEN Teijin DuPont Films JapanLimited., Co., Ltd., Tokyo, JapanQ51-A4CR P.D.R. Co., Ltd., Aichi, Japan crop barathe seawater. After vigorous mixing, 2 mL of the seawater was extracted anddiluted10 timeswithdeionizedwater. For this liquid sample, TOCwasmeasured.Total organic carbon measurementThe TOC measurements were conducted on TOC-L (Shimadzu). The calibra-tion was conducted using sodium carbonate and potassium hydrogen phtha-late aqueous solutions (100 C/L). The measurements were performed threetimes. For determining the absolute values of organic compounds dissolved/scattered in the seawater, we used the averaged TOC values to evaluate poly-mer degradability. The TOC was obtained by the difference between total car-bon value (TC) and inorganic carbon value (IC), that is, TOC = TC – IC.The weight Wfilm and surface area Sfilm of the polymers and the volume ofartificial seawater Vwater were measured before degradation experiments.The weight of carbon dissolved in water was obtained by multiplying Vwaterand TOC. The total weight of carbon elements in the film can be evaluatedby multiplying the proportion of carbon elements in molecular weight Mc%andWfilm. The density of polymers is assumed as 1 g/cm3. Considering the in-fluence of surface area Sfilm, the degradability d is defined as follows:d =TOC3VwaterWfilm 3Mc 3Sfilm: (Equation 6)The results of degradability for experimental dataset 1 and experimental da-taset 2 are shown in Tables 2 and 3, respectively.SUPPLEMENTAL INFORMATIONSupplemental information can be found online at https://doi.org/10.1016/j.patter.2023.100846.ACKNOWLEDGMENTSThis study was partially supported by a project subsidized by NEDO(P15009), AMED (JP20nk0101111), SIP (Technologies for Smart Bio-industryHomopolymer in organic solutionTemplateFilm❶ ❷ ❸Figure 3. Schematic pictures for the preparation process of polymerfilms in experimental dataset 1(1) Polymer solution is attached to the glass slide. (2) Drying is performed toobtain polymer films. (3) The template is removed.Patterns 4, 100846, December 8, 2023 9https://doi.org/10.1016/j.patter.2023.100846https://doi.org/10.1016/j.patter.2023.100846llOPEN ACCESS ArticlePlease cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846and Agriculture), JST ERATO (JPMJER1903), JST CREST (JPMJCR17J2,JPMJCR19J3, JPMJCR21O2), and JST SPRING (JPMJSP2108).AUTHOR CONTRIBUTIONSK.T., R.T., andM.N. conceived the idea and designed the research.W.Y., R.T.,M.S., and K.T. developed the method for machine learning analysis. W.Y.,Y.H., and Y.N. conducted the degradation experiments. K.T., R.T., and M.N.planned and supervised the project. All members contributed to the prepara-tion of the manuscript.DECLARATION OF INTERESTSThe authors declare no competing interests.INCLUSION AND DIVERSITYWe support inclusive, diverse, and equitable conduct of research.Received: April 12, 2023Revised: July 7, 2023Accepted: August 30, 2023Published: September 25, 2023REFERENCES1. Vikhareva, I.N., Buylova, E.A., Yarmuhametova, G.U., Aminova, G.K., andMazitova, A.K. (2021). An Overview of the Main Trends in the Creation ofBiodegradable Polymer Materials. J. Chem. 2021, 1–15. https://doi.org/10.1155/2021/5099705.2. Kuenneth, C., Lalonde, J., Marrone, B.L., Iverson, C.N., Ramprasad, R.,and Pilania, G. (2022). Bioplastic design using multitask deep neural net-works. Commun. Mater. 3. 96–10. https://doi.org/10.1038/s43246-022-00319-2.3. Chamas, A., Moon, H., Zheng, J., Qiu, Y., Tabassum, T., Jang, J.H., Abu-Omar, M., Scott, S.L., and Suh, S. (2020). Degradation Rates of Plastics inthe Environment. ACS Sustain. Chem. Eng. 8, 3494–3511. https://doi.org/10.1021/acssuschemeng.9b06635.4. Pandey, J.K., Raghunatha Reddy, K., Pratheep Kumar, A., and Singh, R.P.(2005). An overview on the degradability of polymer nanocomposites.Polym. Degrad. Stabil. 88, 234–250. https://doi.org/10.1016/j.polymde-gradstab.2004.09.013.5. Ivar do Sul, J.A., and Costa, M.F. (2014). The present and future of micro-plastic pollution in the marine environment. Environ. Pollut. 185, 352–364.https://doi.org/10.1016/j.envpol.2013.10.036.6. Lu, T., Solis-Ramos, E., Yi, Y., and Kumosa, M. (2018). UV degradationmodel for polymers and polymer matrix composites. Polym. Degrad.Stabil. 154, 203–210. https://doi.org/10.1016/j.polymdegradstab.2018.06.004.7. Ganesh Kumar, A., Anjana, K., Hinduja, M., Sujitha, K., and Dharani, G.(2020). Review on plastic wastes in marine environment –Biodegradation and biotechnological solutions. Mar. Pollut. Bull. 150,110733. https://doi.org/10.1016/j.marpolbul.2019.110733.8. Rasulev, B., and Casanola-Martin, G. (2020). QSAR/QSPR in Polymers:Recent Developments in Property Modeling. Int. J. Quant. Struct. Prop.Relatsh. 5, 80–88. https://doi.org/10.4018/IJQSPR.2020010105.9. Sun, X., Hou, Z., Sumita, M., Ishihara, S., Tamura, R., and Tsuda, K. (2020).Data integration for accelerated materials design via preference learning.New J. Phys. 22, 055001. https://doi.org/10.1088/1367-2630/ab82b9.10. Sun, X., Tamura, R., Sumita, M., Mori, K., Terayama, K., and Tsuda, K.(2022). Integrating Incompatible Assay Data Sets with Deep PreferenceLearning. ACS Med. Chem. Lett. 13, 70–75. https://doi.org/10.1021/acs-medchemlett.1c00439.11. Herbrich, R., Graepel, T., and Obermayer, K. (2000). Large margin rankboundaries for ordinal regression. In Advances in large margin classifiers,A.J. Smola, P. Bartlett, B. Schölkopf, and D. Schuurmans, eds. (The MITPress), pp. 115–132. https://doi.org/10.7551/mitpress/1113.003.0010.10 Patterns 4, 100846, December 8, 202312. Joachims, T. (2002). Optimizing search engines using clickthrough data. InProceedings of the eighth ACM SIGKDD international conference onKnowledge discovery and data mining KDD ’02 (Association forComputing Machinery), pp. 133–142. https://doi.org/10.1145/775047.775067.13. Min, K., Cuiffi, J.D., and Mathers, R.T. (2020). Ranking environmentaldegradation trends of plastic marine debris based on physical propertiesand molecular structure. Nat. Commun. 11, 727. https://doi.org/10.1038/s41467-020-14538-z.14. Bagheri, A.R., Laforsch, C., Greiner, A., and Agarwal, S. (2017). Fate of So-Called Biodegradable Polymers in Seawater and Freshwater. Glob. Chall.1, 1700048. https://doi.org/10.1002/gch2.201700048.15. Martin, R.T., Camargo, L.P., and Miller, S.A. (2014). Marine-degradablepolylactic acid. Green Chem. 16, 1768–1773. https://doi.org/10.1039/C3GC42604A.16. Tsuji, H., and Suzuyoshi, K. (2002). Environmental degradation of biode-gradable polyesters 1. Poly(e-caprolactone), poly[(R)-3-hydroxybutyrate],and poly(L-lactide) films in controlled static seawater. Polym. Degrad.Stabil. 75, 347–355. https://doi.org/10.1016/S0141-3910(01)00240-3.17. Syranidou, E., Karkanorachaki, K., Amorotti, F., Avgeropoulos, A.,Kolvenbach, B., Zhou, N.-Y., Fava, F., Corvini, P.F.-X., and Kalogerakis,N. (2019). Biodegradation of mixture of plastic films by tailored marineconsortia. J. Hazard Mater. 375, 33–42. https://doi.org/10.1016/j.jhaz-mat.2019.04.078.18. Muthukumar, T., Aravinthan, A., Lakshmi, K., Venkatesan, R.,Vedaprakash, L., and Doble, M. (2011). Fouling and stability of polymersand composites in marine environment. Int. Biodeterior. Biodegrad. 65,276–284. https://doi.org/10.1016/j.ibiod.2010.11.012.19. Sudhakar, M., Priyadarshini, C., Doble, M., Sriyutha Murthy, P., andVenkatesan, R. (2007). Marine bacteria mediated degradation of nylon66 and 6. Int. Biodeterior. Biodegrad. 60, 144–151. https://doi.org/10.1016/j.ibiod.2007.02.002.20. Artham, T., and Doble, M. (2009). Fouling and Degradation ofPolycarbonate in Seawater: Field and Lab Studies. J. Polym. Environ.17, 170–180. https://doi.org/10.1007/s10924-009-0135-x.21. Artham, T., and Doble, M. (2012). Bisphenol A andmetabolites released bybiodegradation of polycarbonate in seawater. Environ. Chem. Lett. 10,29–34. https://doi.org/10.1007/s10311-011-0324-4.22. Kasuya, K.i., Takagi, K.i., Ishiwatari, S.i., Yoshida, Y., and Doi, Y. (1998).Biodegradabilities of various aliphatic polyesters in natural waters.Polym. Degrad. Stabil. 59, 327–332. https://doi.org/10.1016/S0141-3910(97)00155-9.23. Papageorgiou, G.Z., Bikiaris, D.N., Achilias, D.S., Papastergiadis, E., andDocoslis, A. (2011). Crystallization and biodegradation of poly(butyleneazelate): Comparison with poly(ethylene azelate) and poly(propylene aze-late). Thermochim. Acta 515, 13–23. https://doi.org/10.1016/j.tca.2010.12.010.24. Sridewi, N., Bhubalan, K., and Sudesh, K. (2006). Degradation of commer-cially important polyhydroxyalkanoates in tropical mangrove ecosystem.Polym. Degrad. Stabil. 91, 2931–2940. https://doi.org/10.1016/j.polymde-gradstab.2006.08.027.25. Tachibana, K., Urano, Y., and Numata, K. (2013). Biodegradability of nylon4 film in a marine environment. Polym. Degrad. Stabil. 98, 1847–1851.https://doi.org/10.1016/j.polymdegradstab.2013.05.007.26. Vaclavkova, T., Ruzicka, J., Julinova, M., Vicha, R., and Koutny, M. (2007).Novel aspects of symbiotic (polyvinyl alcohol) biodegradation. Appl.Microbiol. Biotechnol. 76, 911–917. https://doi.org/10.1007/s00253-007-1062-1.27. Papageorgiou, G.Z., and Panayiotou, C. (2011). Crystallization andmeltingof biodegradable poly(propylene suberate). Thermochim. Acta 523,187–199. https://doi.org/10.1016/j.tca.2011.05.023.28. Bikiaris, D.N., Papageorgiou, G.Z., Giliopoulos, D.J., and Stergiou, C.A.(2008). Correlation between Chemical and Solid-State Structures andEnzymatic Hydrolysis in Novel Biodegradable Polyesters. The Case ofhttps://doi.org/10.1155/2021/5099705https://doi.org/10.1155/2021/5099705https://doi.org/10.1038/s43246-022-00319-2https://doi.org/10.1038/s43246-022-00319-2https://doi.org/10.1021/acssuschemeng.9b06635https://doi.org/10.1021/acssuschemeng.9b06635https://doi.org/10.1016/j.polymdegradstab.2004.09.013https://doi.org/10.1016/j.polymdegradstab.2004.09.013https://doi.org/10.1016/j.envpol.2013.10.036https://doi.org/10.1016/j.polymdegradstab.2018.06.004https://doi.org/10.1016/j.polymdegradstab.2018.06.004https://doi.org/10.1016/j.marpolbul.2019.110733https://doi.org/10.4018/IJQSPR.2020010105https://doi.org/10.1088/1367-2630/ab82b9https://doi.org/10.1021/acsmedchemlett.1c00439https://doi.org/10.1021/acsmedchemlett.1c00439https://doi.org/10.7551/mitpress/1113.003.0010https://doi.org/10.1145/775047.775067https://doi.org/10.1145/775047.775067https://doi.org/10.1038/s41467-020-14538-zhttps://doi.org/10.1038/s41467-020-14538-zhttps://doi.org/10.1002/gch2.201700048https://doi.org/10.1039/C3GC42604Ahttps://doi.org/10.1039/C3GC42604Ahttps://doi.org/10.1016/S0141-3910(01)00240-3https://doi.org/10.1016/j.jhazmat.2019.04.078https://doi.org/10.1016/j.jhazmat.2019.04.078https://doi.org/10.1016/j.ibiod.2010.11.012https://doi.org/10.1016/j.ibiod.2007.02.002https://doi.org/10.1016/j.ibiod.2007.02.002https://doi.org/10.1007/s10924-009-0135-xhttps://doi.org/10.1007/s10311-011-0324-4https://doi.org/10.1016/S0141-3910(97)00155-9https://doi.org/10.1016/S0141-3910(97)00155-9https://doi.org/10.1016/j.tca.2010.12.010https://doi.org/10.1016/j.tca.2010.12.010https://doi.org/10.1016/j.polymdegradstab.2006.08.027https://doi.org/10.1016/j.polymdegradstab.2006.08.027https://doi.org/10.1016/j.polymdegradstab.2013.05.007https://doi.org/10.1007/s00253-007-1062-1https://doi.org/10.1007/s00253-007-1062-1https://doi.org/10.1016/j.tca.2011.05.023llOPEN ACCESSArticlePlease cite this article in press as: Yuan et al., Revealing factors influencing polymer degradation with rank-based machine learning, Patterns (2023),https://doi.org/10.1016/j.patter.2023.100846Poly(propylene alkanedicarboxylate)s. Macromol. Biosci. 8, 728–740.https://doi.org/10.1002/mabi.200800035.29. Pan, C., Lu, J., Wu, B., Wu, L., and Li, B.-G. (2017). Effect ofMonomer Structure on Crystallization and Glass Transition of FlexibleCopolyesters. J. Polym. Environ. 25, 1051–1061. https://doi.org/10.1007/s10924-016-0881-5.30. Pantani, R., and Sorrentino, A. (2013). Influence of crystallinity on thebiodegradation rate of injection-moulded poly(lactic acid) samples incontrolled composting conditions. Polym. Degrad. Stabil. 98, 1089–1096. https://doi.org/10.1016/j.polymdegradstab.2013.01.005.31. Mackintosh, A.R., and Liggat, J.J. (2004). Dynamic mechanical analysis ofpoly(trimethylene terephthalate)—A comparison with poly(ethylene tere-phthalate) and poly(ethylene naphthalate). J. Appl. Polym. Sci. 92, 2791–2796. https://doi.org/10.1002/app.20290.32. Turnbull, L., Liggat, J.J., and MacDonald, W.A. (2013). Thermal degrada-tion chemistry of poly(ethylene naphthalate) – A study by thermal volatili-sation analysis. Polym. Degrad. Stabil. 98, 2244–2258. https://doi.org/10.1016/j.polymdegradstab.2013.08.018.33. Scheirs, J., and Gardette, J.-L. (1997). Photo-oxidation and photolysis ofpoly(ethylene naphthalate). Polym. Degrad. Stabil. 56, 339–350. https://doi.org/10.1016/S0141-3910(96)00199-1.34. Liu, B., Zhang, J., and Guo, H. (2022). Research Progress of PolyvinylAlcohol Water-Resistant Film Materials. Membranes 12, 347. https://doi.org/10.3390/membranes12030347.35. Bikiaris, D.N., and Karayannidis, G.P. (1999). Effect of carboxylic endgroups on thermooxidative stability of PET and PBT. Polym. Degrad.Stabil. 63, 213–218. https://doi.org/10.1016/S0141-3910(98)00094-9.36. Otsuka, S., Kuwajima, I., Hosoya, J., Xu, Y., and Yamazaki, M. (2011).PoLyInfo: Polymer Database for Polymeric Materials Design. In 2011International Conference on Emerging Intelligent Data and WebTechnologies, pp. 22–29. https://doi.org/10.1109/EIDWT.2011.13.37. Sahigara, F., Ballabio, D., Todeschini, R., and Consonni, V. (2013).Defining a novel k-nearest neighbours approach to assess the applicabilitydomain of a QSAR model for reliable predictions. J. Cheminf. 5, 27–29.https://doi.org/10.1186/1758-2946-5-27.38. Tokiwa, Y., Calabia, B.P., Ugwu, C.U., and Aiba, S. (2009).Biodegradability of Plastics. Int. J. Mol. Sci. 10, 3722–3742. https://doi.org/10.3390/ijms10093722.39. Seidi, F., Zhong, Y., Xiao, H., Jin, Y., and Crespy, D. (2022). Degradablepolyprodrugs: design and therapeutic efficiency. Chem. Soc. Rev. 51,6652–6703. https://doi.org/10.1039/D2CS00099G.40. Jablonka, K.M., Patiny, L., and Smit, B. (2022). Making the collectiveknowledge of chemistry open and machine actionable. Nat. Chem. 14,365–376. https://doi.org/10.1038/s41557-022-00910-7.41. Jaeger, S., Fulle, S., and Turk, S. (2018). Mol2vec: Unsupervised MachineLearning Approach with Chemical Intuition. J. Chem. Inf. Model. 58,27–35. https://doi.org/10.1021/acs.jcim.7b00616.Patterns 4, 100846, December 8, 2023 11https://doi.org/10.1002/mabi.200800035https://doi.org/10.1007/s10924-016-0881-5https://doi.org/10.1007/s10924-016-0881-5https://doi.org/10.1016/j.polymdegradstab.2013.01.005https://doi.org/10.1002/app.20290https://doi.org/10.1016/j.polymdegradstab.2013.08.018https://doi.org/10.1016/j.polymdegradstab.2013.08.018https://doi.org/10.1016/S0141-3910(96)00199-1https://doi.org/10.1016/S0141-3910(96)00199-1https://doi.org/10.3390/membranes12030347https://doi.org/10.3390/membranes12030347https://doi.org/10.1016/S0141-3910(98)00094-9https://doi.org/10.1109/EIDWT.2011.13https://doi.org/10.1186/1758-2946-5-27https://doi.org/10.3390/ijms10093722https://doi.org/10.3390/ijms10093722https://doi.org/10.1039/D2CS00099Ghttps://doi.org/10.1038/s41557-022-00910-7https://doi.org/10.1021/acs.jcim.7b00616 PATTER100846_proof.pdf Revealing factors influencing polymer degradation with rank-based machine learning Introduction Results and discussion Degradation datasets Unified ranking integrating three datasets Revealing factors influencing polymer degradation Predictions of degradable or undegradable polymers in PoLyInfo Conclusions Experimental procedures Resource availability Lead contact Materials availability Data and code availability Calculation details in RankSVM Materials Film preparations in experimental dataset 1 Degradation experiments Total organic carbon measurement Supplemental information Acknowledgments Author contributions Declaration of interests Inclusion and diversity References