{"paper_id":"10cdb9c0-b660-45f8-8b2c-38ed1ef8f750","body_text":"PREPRINT\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025\nDOI: https://doi.org/10.3897/arphapreprints.e176591\nAutomated extraction of fungal trophic\nmodes from literature using BioBERT: an\nopen pilot workflow\n Beatrice Bock\n\nReviewed v 1\nAutomated extraction of fungal trophic modes\nfrom literature using BioBERT: an open pilot\nworkflow\nBeatrice Margareta Bock \n‡ Department of Biological Sciences, Northern Arizona University, Flagstaff, United States of America\n§ Center for Adaptable Western Landscapes, Northern Arizona University, Flagstaff, United States of America\nCorresponding author: Beatrice Margareta Bock (beabockm@gmail.com)\nAcademic editor: Editorial Secretary\nAbstract\nFungi exhibit diverse trophic strategies, ranging from obligate symbiosis to saprotrophy,\nwith  some  taxa  capable  of occupying  multiple  ecological  roles. Manually  identifying\ntrophic versatility from literature is time-consuming and difficult to scale. Here, we present\na  pilot  workflow  that  automates  the  classification  of  fungal  trophic  modes  using\ntransformer-based language models. A curated dataset of 56 fungal ecology abstracts\nwas manually labelled as dual (occupying multiple trophic modes) or solo (restricted to\none mode) and used to fine-tune four models: BioBERT, BERT-base-cased, BERT-base-\nuncased and BiodivBERT. Stratified 5-fold cross-validation revealed that BioBERT and\nBERT-base-cased  performed  equally  well  (~  89%  accuracy, balanced  precision  and\nrecall), highlighting the importance of case sensitivity in taxonomic text. BiodivBERT and\nuncased BERT models underperformed, indicating that domain adaptation alone is not\nsufficient.  This  pilot  study  emphasises  reproducibility,  transparency  and  open  data\nintegration,  offering  a  generalisable  proof-of-concept  for  linking  literature-derived\necological  information  to  existing  fungal  trait  databases such  as  FUNGuild  and\nFungalTraits. All code and data are openly available to support reuse and scaling to\nlarger datasets.\nKeywords\nfungal  ecology, trophic  modes, natural  language  processing, machine  learning, trait\ndatabases, BioBERT, saprotrophy-symbiosis continuum\n‡, § \n© Bock B. This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY 4.0),\nwhich permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are\ncredited.\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\nIntroduction\nFungi play essential roles in ecosystems as decomposers, pathogens and symbionts,\nand many taxa exhibit flexibility across these trophic modes ( Berbee et al. 2017 ). Some\nspecies  occupy  multiple  ecological  strategies  depending  on  host  identity  or\nenvironmental context, while others remain specialised to a single mode ( Martin and Tan\n2025).  Understanding  this  trophic  versatility  is  critical  for  biodiversity  assessments,\necosystem modelling and agricultural applications where a fungus’ ecological role affects\nplant performance.\nTrait databases, such as FUNGuild ( Nguyen et al. 2016 ) and FungalTraits ( Põlme et al.\n2021),  have  advanced  fungal  functional  annotation  at  scale,  but  manual  literature\nextraction  remains  a  bottleneck.  Human-curated  classifications  are  time-intensive,\nsubjective  and  limited  by  the  accessibility  of trait-relevant language  in  publications.\nMoreover, trait databases are valuable but can be limited in applicability, as annotations\nare often performed at the genus or family level despite often substantial interspecific\nvariability ( Violle et al. 2015 ).\nNatural  language  processing  (NLP)  offers  a  scalable  way  to  extract  trait-relevant\ninformation directly from text. Transformer-based models, such as BioBERT, pretrained\non large biomedical corpora, excel at contextual understanding and have achieved state-\nof-the-art results in various text classification tasks ( Lee et al. 2019 ). However, their use in\nfungal ecology and trait data integration remains largely unexplored.\nThis pilot study tests the feasibility of fine-tuning BioBERT to classify fungal trophic modes\nfrom abstracts. The workflow is designed for transparency and future scaling, providing a\nreproducible  pipeline  that can  complement or benchmark existing  trait databases. By\nlinking automated text classification with open trait resources, this work demonstrates a\npath towards more consistent, interoperable fungal functional data.\nRelated Work\nRecent advances in domain-specific NLP illustrate the potential for scaling and refining\nworkflows  like  this  one. BiodivBERT ( Abdelmageed  et al. 2023 )  represents  the  first\npretrained  language  model  tailored  specifically  for  biodiversity  research,  achieving\nsignificant gains in named entity recognition and relation extraction. Its development from\nlife-sciences corpora  highlights how  domain  adaptation  can  substantially improve  the\nprecision and recall of ecological information retrieval. Similarly, Cornelius et al. (2025)\ndemonstrated a machine-learning framework for extracting arthropod organismal traits,\ntranslating  unstructured  text  into  a  machine-actionable  database  (ArTraDB).  Their\napproach  shows  how  targeted  trait extraction  can  efficiently  transform  literature  into\nstructured ecological data.\n2\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\nParallel work in plant functional ecology supports the broader utility of transformer-based\nextraction for trait data. Domazetoski et al. (2025)  developed a natural language pipeline\nthat automatically identifies both categorical and numerical plant traits from unstructured\ndescriptions with  high  precision  and  recall. This scalability across morphological, life\nhistory and  functional  trait types underscores the  potential  to  adapt such  methods to\nfungal traits, where similar data gaps persist.\nBeyond  ecological  trait extraction, innovations  in  model  architecture  and  pretraining\nfurther extend applicability. BioT5 ( Pei et al. 2023 ) introduced cross-modal pretraining to\nconnect  textual  and  molecular  data  using  chemically  informed  representations,  an\napproach that could eventually allow integration of genomic or metabolomic predictors\ninto ecological models. ModernBERT ( Warner et al. 2025 ) provides a computationally\nefficient encoder capable of handling long-context inputs, which is particularly relevant\nfor large-scale biodiversity corpora. Together, these efforts suggest practical pathways to\nscale fungal trait text mining beyond small datasets, while maintaining interpretability and\nreproducibility. Foundational work by Gu et al. (2021)  also reinforces the importance of\ndomain-specific pretraining, demonstrating  that models trained  entirely within  a  target\ndomain outperform general-domain models adapted later.\nMaterials and Methods\nDataset Curation\nFig. 1 shows  the  overall  pipeline  from  data  collection  to  evaluation. Fungal  research\nabstracts were retrieved from the Web of Science Core Collection on 11 September 2025,\nusing  two  Boolean  search  queries  designed  to  capture  fungi  with  distinct  lifestyle\nclassifications.  For  solo  (single  trophic  mode)  examples,  we  searched:  '(\"obligate\nmycorrhizal\"  OR  \"strictly  endophytic\"  OR  \"exclusive  saprotroph\")  AND  fungus'  (119\nresults). For dual (multiple trophic mode) examples, we searched: '(\"dual lifestyle\" OR\n\"facultative  lifestyle\" OR  \"dual  trophic mode\" OR  \"lifestyle  switching\" OR  \"endophyte-\nsaprotroph\" OR \"plant-associated saprotroph\") AND fungi' (70 results).\nFrom these 189 candidate articles, abstracts were manually reviewed by a single labeller\n(BMB) and 56 were selected, based on: (1) unambiguous description of trophic mode in\nthe abstract text; (2) English language and (3) no duplicates between searches. Abstracts\nwere labelled as:\n     Dual: taxa reported to occupy more than one trophic mode (e.g. facultative pathogens\nthat also decompose organic matter);   \n     Solo:  taxa  restricted  to  a  single  trophic  mode  (e.g.  obligate  symbionts  or  strict\nsaprotrophs).\nAmbiguous abstracts without explicit trophic mode statements were excluded. The final\ndataset contained 56 abstracts (28 dual, 28 solo). The dataset was balanced between\n3\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\ndual and solo classes, with abstract lengths ranging from 150–500 words (mean 360). A\nsupplementary file  listing  the  permissible  trophic mode  labels and  their definitions is\navailable in the repository (datasets/trophic_mode_labels.md).\nPreprocessing and Model Training\nAbstracts  were  cleaned  and  tokenised  using  model-specific  tokenisers  (maximum\nsequence length: 512 tokens). Token length analysis revealed that three of 56 abstracts\n(5.4%) exceeded this limit and were truncated, though truncation occurred at the end of\nabstracts where contextual information for classification is typically less concentrated.\nWe compared four transformer-based language models to assess the impact of domain-\nspecific pretraining and case sensitivity: (1) BERT-base-uncased ('google-bert/bert-base-\nuncased'; Devlin et al. (2018) ), (2) BERT-base-cased ('google-bert/bert-base-cased'), (3)\nBioBERT v.1.1 ('monologg/biobert_v.1.1_pubmed', biomedical domain-adapted; Lee et\nal.  (2019) )  and  (4)  BiodivBERT  ('NoY o25/BiodivBERT',  biodiversity  domain-adapted;\nAbdelmageed  et  al.  (2023) ).  All  models  were  fine-tuned  for  binary  sequence\nclassification  using  the  Hugging  Face  Transformers  library  (v.4.40.0; Wolf (2020) ) in\nPython 3.9 (Python Software Foundation 2020 ).\nTo maximise statistical robustness with the small dataset, we employed stratified 5-fold\ncross-validation on all 56 abstracts rather than a single train-test split. This approach\nensures that each abstract is used once for validation while maintaining class balance\nacross  folds  (seed  =  42  for  reproducibility).  Models  were  trained  using  identical\nhyperparameters: learning rate = 5 × 10⁻⁵, batch size = 8, maximum epochs = 20 with\nearly stopping  (patience  = 3  epochs), dropout = 0.2, optimiser = AdamW with  class-\nweighted loss to account for any fold-level imbalances (PyTorch v.2.0.1; Paszke et al.\n(2019)). Training was executed on NAU's Monsoon HPC cluster using Tesla K80 GPUs\nwith CUDA 11.4.\nEvaluation\nPerformance was evaluated across all five folds for each model. For each fold, accuracy,\nprecision, recall and F1-score were computed using scikit-learn v.1.4.2 ( Pedregosa et al.\n2011) and results were aggregated as mean ± standard deviation. Precision, recall and\nF1-score are reported as macro averages (unweighted mean across both classes), which\ntreats solo and dual classes equally and is appropriate for balanced binary classification.\nComparative  visualisations include: (1) model  performance  bar charts with  error bars\nrepresenting cross-fold variation (Fig. 2); (2) aggregated confusion matrices for BioBERT\nshowing cumulative predictions across all folds (Fig. 3) and (3) training time comparison\nacross  models  (Fig.  4).  Error  analysis  identified  abstracts  misclassified  by  multiple\nmodels, highlighting particularly ambiguous cases (results/error_analysis.csv).\n4\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\nCode and Reproducibility\nAll code and data are openly available via: Zenodo and GitHub (Bock 2026).  \nThe repositories include scripts for dataset curation, model fine-tuning and evaluation,\nalong with example outputs and documentation to facilitate reuse or adaptation for other\ntrait-related classification tasks.\nResults\nModel Performance\nWe  compared  four  transformer-based  language  models  using  stratified  5-fold  cross-\nvalidation  on  all  56  abstracts:  (1)  BERT-base-uncased;  (2)  BERT-base-cased;  (3)\nBioBERT v.1.1 (biomedical domain-adapted) and (4) BiodivBERT (biodiversity domain-\nadapted). All models were trained with identical hyperparameters (20 epochs maximum\nwith early stopping, learning rate 5e-5, batch size 8, dropout 0.2).\nBioBERT achieved  the  highest overall  performance  (F1  = 0.892  ±  0.120, Accuracy =\n0.894  ±  0.116),  marginally  outperforming  BERT-base-cased  (F1  =  0.892  ±  0.100,\nAccuracy = 0.892 ± 0.100). Case sensitivity proved critical: cased models substantially\noutperformed  their  uncased  counterparts  (BERT-base-uncased: F1  =  0.700  ±  0.241,\nAccuracy  =  0.749  ±  0.177),  likely  because  taxonomic  nomenclature  capitalisation\nprovides important classification signals. Surprisingly, BiodivBERT underperformed (F1 =\n0.747 ± 0.198, Accuracy = 0.771 ± 0.166) despite its biodiversity-specific pre-training,\nsuggesting that domain alignment alone does not guarantee superior performance on\nspecialised classification tasks.\nClassification metrics for BioBERT are summarised in Table 1. Results are reported as\nmean ± standard deviation across five folds. Comparative model performance (Fig. 2)\ndemonstrates  that  cased  models  substantially  outperform  uncased  variants,  while\nBioBERT  and  BERT-cased  achieve  statistically  equivalent  results.  The  aggregated\nconfusion  matrices  (Fig.  3)  shows  balanced  performance  across  both  classes  for\nBioBERT and BERT-cased. Training efficiency varied substantially (Fig. 4), with BioBERT\nand BERT-cased completing training in ~ 10-11 minutes while BiodivBERT and BERT-\nuncased required ~ 35 minutes, likely due to differences in tokenisation efficiency and\nconvergence patterns.\nDiscussion\nThis pilot study demonstrates that transformer-based NLP models can extract ecological\ninformation  embedded  in  scientific  text.  With  a  small  but  carefully  curated  dataset,\nBioBERT and BERT-base-cased achieved ~ 89% accuracy in classifying fungal trophic\n5\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\nmodes, indicating  that pretrained  language  models, whether  biomedical  or  general-\npurpose, can generalise to ecological contexts with minimal fine-tuning.\nThe  comparative  model  analysis  revealed  three  important insights. First, BioBERT's\nmarginal advantage over BERT-base-cased suggests that biomedical domain adaptation\nprovides limited benefit for this specific task, possibly because fungal ecology vocabulary\ndiffers substantially from clinical biomedical text. Second, case sensitivity proved critical:\nmodels trained on cased text outperformed uncased variants by ~ 15 percentage points,\nlikely because  taxonomic nomenclature  capitalisation  (e.g. *Fusarium*  vs. *fusarium*)\ncarries important classification signals. Third, BiodivBERT's underperformance, despite\nbiodiversity-specific  pre-training,  indicates  that  domain  alignment  alone  does  not\nguarantee  superior  performance;  the  pre-training  corpus  must  closely  match  the\ndownstream task domain.\nThe workflow presented here provides a proof-of-concept for trait-orientated text mining\nin  fungi. Discrepancies  between  automated  and  curated  classifications  can  highlight\nambiguous  or  conflicting  entries,  which  is  particularly  important  given  the  known\nlimitations in trait data provenance ( Violle et al. 2015 ). When scaled, this approach may\nsupport more dynamic updating and cross-validation of trait resources such as FUNGuild\nand FungalTraits.\nError  analysis  across  all  models  (figures/error_analysis.csv)  reveals  that  39  total\nmisclassifications occurred across 280 predictions (4 models × 5 folds × 56/5 samples\nper fold). Seven abstracts were misclassified by at least two models, suggesting inherent\nambiguity in how trophic modes are described in these cases. For example, abstracts\ndescribing endophytes with both plant-beneficial and saprotrophic capabilities proved\nchallenging for all models, likely because the text emphasises ecological context over\nexplicit  trophic  mode  labels.  These  consistently  problematic  cases  highlight  where\nhuman curation remains essential and where future annotation guidelines could improve\nclarity.  Improving  model  learning,  such  as  by  increasing  the  dataset  size,  refining\nannotation or using more advanced models, may also help resolve these ambiguous\ncases in future work.\nLimitations\nThis pilot study has several deliberate constraints. The small sample size (56 abstracts)\nlimits  statistical  power  and  generalisability,  though  stratified  cross-validation  helps\nmitigate overfitting concerns. The binary simplification (solo vs. dual) reduces the rich\nspectrum of fungal  trophic strategies to  a  coarse  categorisation; real  ecological  roles\noften exist along gradients rather than discrete classes. Manual label subjectivity by a\nsingle labeller may introduce bias, though this was mitigated by conservative inclusion\ncriteria requiring explicit trophic mode statements.\nFuture Work\nSeveral directions could extend this pilot into a more comprehensive tool:\n6\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\n1. Expanding dataset scope: Increasing the corpus to hundreds or thousands of abstracts\nwould improve model robustness and allow detection of rarer trophic patterns. This could\ninclude full-text articles rather than abstracts alone and taxa beyond fungi to establish\ncross-kingdom applicability;\n2.  Multi-label  classification  for  ecological  gradients:  Rather  than  binary  solo/dual\nclassification,  future  models  could  predict  specific  trophic  modes  (e.g.  saprotroph,\nsymbiont,  pathogen,  endophyte)  as  non-exclusive  labels.  This  would  capture  the\ncontinuous nature of ecological roles and better reflect biological reality where taxa may\nsimultaneously occupy multiple niches;\n3. Integration with environmental metadata: Linking text-derived traits with geographic,\nclimatic or substrate data could enable context-aware predictions; for example, predicting\nhow a taxon's trophic mode might shift under different environmental conditions;\n4. Genomic and metabolomic predictors: Combining textual information with molecular\ndata  (e.g.  gene  content,  secondary  metabolite  profiles)  could  improve  prediction\naccuracy and provide mechanistic insights into trophic flexibility;\n5. Domain-specific pretraining: Training a BERT-style model from scratch on ecological\nand  mycological  corpora  (following  the  BiodivBERT  approach)  could  yield  better\nperformance than adapting biomedical models.\nConclusions\nThis pilot study demonstrates that transformer-based NLP can successfully extract fungal\ntrophic mode information from scientific abstracts. BioBERT and BERT-cased achieved ~\n89% accuracy in classifying abstracts as describing single or multiple trophic modes,\nvalidating  the  feasibility  of  automated  trait  extraction  for  fungal  ecology.  The  key\ncontributions of this work are:\n1. Proof-of-concept: Pretrained biomedical language models can generalise to ecological\nclassification tasks with minimal fine-tuning;\n2. Reproducible workflow: All code and data are openly available, enabling replication\nand extension by other researchers;\n3.  Trait  database  integration:  The  approach  complements  existing  resources  like\nFUNGuild  and  FungalTraits by providing  a  scalable  method  to  flag  taxa  with  lifestyle\nplasticity.  As  fungal  trait  databases  continue  to  grow  in  importance  for  biodiversity\nassessments and  ecosystem modelling, automated  text mining  offers a  path  towards\nmore efficient, consistent and comprehensive trait annotation. This workflow provides a\nfoundation for scaling to larger datasets and more nuanced ecological classifications. \n7\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\nAcknowledgements\nThank you to Dr. Nancy Johnson and Dr. Kitty Gehring for guidance and support on this\nand other projects. Special thanks to Anne-Marie Cooper, as well.\nConflicts of interest\nThe authors have declared that no competing interests exist.\nReferences\n• Abdelmageed N, Löffler F , König-Ries B (2023) BiodivBERT : a Pre-Trained Language\nModel for the Biodiversity Domain. In: Yamaguchi A, et al. (Ed.) SWAT4HCLS. 62-71 pp.\nURL: https://ceur-ws.org/Vol-3415/paper-7.pdf\n• Berbee M, James T , Strullu-Derrien C (2017) Early Diverging Fungi: Diversity and Impact\nat the Dawn of T errestrial Life. Annual Review of Microbiology 71 (1): 41‑60. https://\ndoi.org/10.1146/annurev-micro-030117-020324\n• Bock B (2026) beabock/biobert_dualsolo: Reproducible BioBERT & BERT model\ncomparison (4 models, 5-fold CV). Zenodo https://doi.org/10.5281/zenodo.17343492\n• Cornelius J, Detering H, Lithgow-Serrano O, Agosti D, Rinaldi F , Waterhouse R (2025)\nFrom literature to biodiversity data: mining arthropod organismal traits with machine\nlearning. Biodiversity Data Journal 13 https://doi.org/10.3897/bdj.13.e153070\n• Devlin J, Chang M, Lee K, T outanova K (2018) BERT : Pre-training of Deep Bidirectional\nTransformers for Language Understanding. CoRR URL: http://arxiv.org/abs/1810.04805\n• Domazetoski V , Kreft H, Bestova H, Wieder P , Koynov R, Zarei A, Weigelt P (2025)\nUsing large language models to extract plant functional traits from unstructured text.\nApplications in Plant Sciences 13 (3). https://doi.org/10.1002/aps3.70011\n• Gu Y , Tinn R, Cheng H, Lucas M, Usuyama N, Liu X, Naumann T , Gao J, Poon H (2021)\nDomain-Specific Language Model Pretraining for Biomedical Natural Language\nProcessing. ACM Transactions on Computing for Healthcare 3 (1): 1‑23. https://doi.org/\n10.1145/3458754\n• Lee J, Y oon W, Kim S, Kim D, Kim S, So CH, Kang J (2019) BioBERT : a pre-trained\nbiomedical language representation model for biomedical text mining. Bioinformatics 36\n(4): 1234‑1240. https://doi.org/10.1093/bioinformatics/btz682\n• Martin F , T an H (2025) Saprotrophy-to-symbiosis continuum in fungi. Current Biology 35\n(11). https://doi.org/10.1016/j.cub.2025.01.032\n• Nguyen N, Song Z, Bates S, Branco S, T edersoo L, Menke J, Schilling J, Kennedy P\n(2016) FUNGuild: An open annotation tool for parsing fungal community datasets by\necological guild. Fungal Ecology 20: 241‑248. https://doi.org/10.1016/j.funeco.2015.06.006\n• Paszke A, Gross S, Massa F , Lerer G, Bradbury J, Chanan G, Killeen T , Lin Z,\nGimelshein N, Chintala S, et al. (2019) PyT orch: An Imperative Style, High-Performance\nDeep Learning Library. Advances in Neural Information Processing Systems 32:\n8024‑8035. \n8\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\n• Pedregosa F , Varoquaux G, Gramfort A, Michel V , Thirion B, Grisel O, Blondel M,\nPrettenhofer P , Weiss R, Dubourg V , Vanderplas J, Passos A, Cournapeau D, Brucher M,\nPerrot M, Duchesnay E (2011) Scikit-learn: Machine Learning in {P}ython. Journal of\nMachine Learning Research 12: 2825‑2830. \n• Pei Q, Zhang W, Zhu J, Wu K, Gao K, Wu L, Xia Y , Yan R (2023) BioT5: Enriching\nCross-modal Integration in Biology with Chemical Knowledge and Natural Language\nAssociations. Proceedings of the 2023 Conference on Empirical Methods in Natural\nLanguage Processing https://doi.org/10.18653/v1/2023.emnlp-main.70\n• Põlme S, Abarenkov K, Henrik Nilsson R, Lindahl B, Clemmensen KE, Kauserud H,\nNguyen N, Kjøller R, Bates S, Baldrian P , Frøslev TG, Adojaan K, Vizzini A, Suija A,\nPfister D, Baral H, Järv H, Madrid H, Nordén J, Liu J, Pawlowska J, Põldmaa K, Pärtel K,\nRunnel K, Hansen K, Larsson K, Hyde KD, Sandoval-Denis M, Smith M, T oome-Heller M,\nWijayawardene N, Menolli N, Reynolds N, Drenkhan R, Maharachchikumbura SN,\nGibertoni T , Læssøe T , Davis W, T okarev Y , Corrales A, Soares AM, Agan A, Machado\nAR, Argüelles-Moyao A, Detheridge A, de Meiras-Ottoni A, Verbeken A, Dutta AK, Cui B,\nPradeep CK, Marín C, Stanton D, Gohar D, Wanasinghe D, Otsing E, Aslani F , Griffith G,\nLumbsch T , Grossart H, Masigol H, Timling I, Hiiesalu I, Oja J, Kupagme J, Geml J,\nAlvarez-Manjarrez J, Ilves K, Loit K, Adamson K, Nara K, Küngas K, Rojas-Jimenez K,\nBitenieks K, Irinyi L, Nagy L, Soonvald L, Zhou L, Wagner L, Aime MC, Öpik M, Mujica\nMI, Metsoja M, Ryberg M, Vasar M, Murata M, Nelsen M, Cleary M, Samarakoon M,\nDoilom M, Bahram M, Hagh-Doust N, Dulya O, Johnston P , Kohout P , Chen Q, Tian Q,\nNandi R, Amiri R, Perera RH, dos Santos Chikowski R, Mendes-Alvarenga R, Garibay-\nOrijel R, Gielen R, Phookamsak R, Jayawardena R, Rahimlou S, Karunarathna S,\nTibpromma S, Brown S, Sepp S, Mundra S, Luo Z, Bose T , Vahter T , Netherway T , Yang\nT , May T , Varga T , Li W, Coimbra VRM, de Oliveira VRT , de Lima VX, Mikryukov V , Lu Y ,\nMatsuda Y , Miyamoto Y , Kõljalg U, T edersoo L (2021) FungalTraits: a user-friendly traits\ndatabase of fungi and fungus-like stramenopiles. Fungal Diversity 105 (1): 1‑16. https://\ndoi.org/10.1007/s13225-020-00466-2\n• Python Software Foundation (2020) Python 3.9.0. URL: https://docs.python.org/3.9\n• Violle C, Borgy B, Choler P (2015) Trait databases: misuses and precautions. Journal of\nVegetation Science 26 (5): 826‑827. https://doi.org/10.1111/jvs.12325\n• Warner B, Chaffin A, Clavié B, Weller O, Hallström O, T aghadouini S, Gallagher A,\nBiswas R, Ladhak F , Aarsen T , Adams GT , Howard J, Poli I (2025) Smarter, Better,\nFaster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long\nContext Finetuning and Inference. Proceedings of the 63rd Annual Meeting of the\nAssociation for Computational Linguistics (Volume 1: Long Papers)2526‑2547. https://\ndoi.org/10.18653/v1/2025.acl-long.127\n• Wolf T , et al. (2020) Transformers: State-of-the-Art Natural Language Processing.\nProceedings of the 2020 Conference on Empirical Methods in Natural Language\nProcessing: System Demonstrations. Online. Association for Computational Linguistics\nURL: https://www.aclweb.org/anthology/2020.emnlp-demos.6\n9\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\nFigure 1.  \nWorkflow diagram of the classification pipeline. The diagram summarises the full end-to-end\nprocess: literature search (Web of Science queries), manual curation (56 labelled abstracts),\npreprocessing (text cleaning, tokenisation with truncation at 512 tokens and token-length QC),\nstratified 5-fold cross-validation, model fine-tuning across four models (BERT-base-uncased,\nBERT-base-cased,  BioBERT  v.1.1,  BiodivBERT)  with  standardised  hyperparameters,\nevaluation  (metrics,  confusion  matrices,  learning  curves)  and  outputs  (fine-tuned  models,\npredictions).\n \n10\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\nFigure 2.  \nComparative  model performance  across four  transformer-based  architectures.  Bar  charts\nshow mean classification metrics (accuracy, precision, recall, F1-score) ± standard deviation\nfrom  stratified  5-fold  cross-validation  (n  =  56  abstracts).  BioBERT  (biomedical  domain-\nadapted)  and  BERT-base-cased  achieved  statistically  equivalent  performance  (~  89%\naccuracy), substantially outperforming BERT-base-uncased (~ 75%) and BiodivBERT (~ 77%).\nCase sensitivity proved critical, with cased models outperforming uncased by ~ 15 percentage\npoints. Metrics are calculated as macro averages (unweighted mean across dual and solo\nclasses).\n \n11\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\nFigure 3.  \nAggregated confusion matrices for all four models (BioBERT , BERT-base-cased, BERT-base-\nuncased, BiodivBERT) across all five folds (total 56 predictions per model). Each matrix shows\ntrue labels (Solo = single trophic mode, Dual = multiple trophic modes) versus predicted labels,\nallowing direct comparison of error patterns and class balance for each model. BioBERT and\nBERT-base-cased  show  balanced  performance,  while  uncased  and  BiodivBERT  models\ndisplay more misclassifications. Colour intensity indicates prediction frequency; diagonal cells\nrepresent correct predictions.\n \n12\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\nFigure 4.  \nTraining time comparison across models. BioBERT and BERT-base-cased completed 5-fold\ncross-validation  training  in  ~  10-11  minutes,  while  BiodivBERT and  BERT-base-uncased\nrequired  ~  35  minutes.  Differences  likely  reflect  tokenisation  efficiency  and  convergence\npatterns rather than model size (all models have ~ 110M parameters). Faster convergence in\ncased models correlates with higher classification accuracy, suggesting that case-preserving\ntokenisation  provides  stronger  learning  signals  for  this  taxonomic  text  classification  task.\nTraining performed on NAU Monsoon HPC cluster (T esla K80 GPU, CUDA 11.4).\n \n13\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591\n\nMetric Value Note \nAccuracy 89.4% ± 11.6% Fraction of correctly predicted labels\nPrecision 89.9% ± 11.5% Positive predictive value (macro average)\nRecall 88.8% ± 12.4% True positive rate (macro average)\nF1-Score 89.2% ± 12.0% Harmonic mean of precision and recall (macro average)\nTable 1. \nBioBERT classification performance on fungal trophic modes (5-fold CV). Precision, recall and F1-\nscore  are  reported  as  macro  averages  (unweighted  mean  across  both  classes),  which  is\nappropriate for balanced binary classification and treats both classes equally regardless of support.\n14\nAuthor-formatted, not peer-reviewed document posted on 03/11/2025. DOI:  \nhttps://doi.org/10.3897/arphapreprints.e176591","source_license":"CC-BY-4.0","license_restricted":false}