{"status":"OK","data":{"id":565081,"identifier":"SWCLWE","persistentUrl":"https://doi.org/10.34894/SWCLWE","protocol":"doi","authority":"10.34894","separator":"/","publisher":"DataverseNL","publicationDate":"2025-11-28","storageIdentifier":"surf://10.34894/SWCLWE","effectiveDatasetFileCountLimit":10000,"datasetFileUploadsAvailable":9997,"datasetType":"dataset","locks":[],"latestVersion":{"id":31948,"datasetId":565081,"datasetPersistentId":"doi:10.34894/SWCLWE","datasetType":"dataset","storageIdentifier":"surf://10.34894/SWCLWE","versionNumber":1,"internalVersionNumber":24,"versionMinorNumber":0,"versionState":"RELEASED","latestVersionPublishingState":"RELEASED","deaccessionLink":"","productionDate":"2025-11-14","lastUpdateTime":"2025-11-28T09:16:57Z","releaseTime":"2025-11-28T09:16:57Z","createTime":"2025-11-20T09:28:21Z","publicationDate":"2025-11-28","citationDate":"2025-11-28","effectiveDatasetFileCountLimit":10000,"datasetFileUploadsAvailable":9997,"license":{"name":"CC-BY-4.0","uri":"http://creativecommons.org/licenses/by/4.0","iconUri":"https://i.creativecommons.org/l/by/4.0/88x31.png"},"fileAccessRequest":true,"metadataBlocks":{"citation":{"displayName":"Citation Metadata","name":"citation","fields":[{"typeName":"title","multiple":false,"typeClass":"primitive","value":"Uncovering Internal Prediction Mechanisms of Transformer-Based Chemical Foundation Models"},{"typeName":"author","multiple":true,"typeClass":"compound","value":[{"authorName":{"typeName":"authorName","multiple":false,"typeClass":"primitive","value":"Pollice, Robert"},"authorAffiliation":{"typeName":"authorAffiliation","multiple":false,"typeClass":"primitive","value":"https://ror.org/012p63287","expandedvalue":{"scheme":"http://www.grid.ac/ontology/","termName":"University of Groningen","@type":"https://schema.org/Organization"}},"authorIdentifierScheme":{"typeName":"authorIdentifierScheme","multiple":false,"typeClass":"controlledVocabulary","value":"ORCID"},"authorIdentifier":{"typeName":"authorIdentifier","multiple":false,"typeClass":"primitive","value":"0000-0001-8836-6266"}},{"authorName":{"typeName":"authorName","multiple":false,"typeClass":"primitive","value":"Cardenas-Cartagena, Juan"},"authorAffiliation":{"typeName":"authorAffiliation","multiple":false,"typeClass":"primitive","value":"https://ror.org/012p63287","expandedvalue":{"scheme":"http://www.grid.ac/ontology/","termName":"University of Groningen","@type":"https://schema.org/Organization"}},"authorIdentifierScheme":{"typeName":"authorIdentifierScheme","multiple":false,"typeClass":"controlledVocabulary","value":"ORCID"},"authorIdentifier":{"typeName":"authorIdentifier","multiple":false,"typeClass":"primitive","value":"0000-0001-9718-6929"}},{"authorName":{"typeName":"authorName","multiple":false,"typeClass":"primitive","value":"Müller, Alexander"},"authorAffiliation":{"typeName":"authorAffiliation","multiple":false,"typeClass":"primitive","value":"https://ror.org/012p63287","expandedvalue":{"scheme":"http://www.grid.ac/ontology/","termName":"University of Groningen","@type":"https://schema.org/Organization"}},"authorIdentifierScheme":{"typeName":"authorIdentifierScheme","multiple":false,"typeClass":"controlledVocabulary","value":"ORCID"},"authorIdentifier":{"typeName":"authorIdentifier","multiple":false,"typeClass":"primitive","value":"0009-0002-9592-5606"}}]},{"typeName":"datasetContact","multiple":true,"typeClass":"compound","value":[{"datasetContactName":{"typeName":"datasetContactName","multiple":false,"typeClass":"primitive","value":"Groningen Digital Competence Centre"},"datasetContactAffiliation":{"typeName":"datasetContactAffiliation","multiple":false,"typeClass":"primitive","value":"rug.nl"}}]},{"typeName":"dsDescription","multiple":true,"typeClass":"compound","value":[{"dsDescriptionValue":{"typeName":"dsDescriptionValue","multiple":false,"typeClass":"primitive","value":"Supervised deep learning has become a standard approach to\ndeliver competitive predictive tools that allow relating the structure\nof molecules and their physicochemical features to properties such as\nbinding to protein targets, performance as electronic materials, and\nreactivity. However, efforts to understand how these models learn and\nhow specific predictions can be explained are still limited in\nchemistry, which hinders trust in these predictions. To overcome both\nthe current widespread approach of considering supervised deep learning\nmodels as black boxes and the limitations of explainable deep learning\nbased on feature attribution, we introduce two general methods from the\nfield of mechanistic interpretability to molecular property prediction.\nSpecifically, we leverage ablation and adapt existing techniques into\nthe regression lens to inspect model predictions of the\nTransformer-based ChemBERTa foundation model. Our results for 3\nChemBERTa-based models finetuned on distinct datasets allow us to\npropose the internal mechanisms operable within each of the\ncorresponding model layers that lead to the final predictions. Our\nresults are a stepping stone towards more trustworthy deep learning\nmodels in the molecular domain."}}]},{"typeName":"subject","multiple":true,"typeClass":"controlledVocabulary","value":["Chemistry","Computer and Information Science"]},{"typeName":"keyword","multiple":true,"typeClass":"compound","value":[{"keywordValue":{"typeName":"keywordValue","multiple":false,"typeClass":"primitive","value":"Mechanistic Interpretability"}},{"keywordValue":{"typeName":"keywordValue","multiple":false,"typeClass":"primitive","value":"Chemical Foundation Models"}},{"keywordValue":{"typeName":"keywordValue","multiple":false,"typeClass":"primitive","value":"Molecular Property Prediction"}}]},{"typeName":"topicClassification","multiple":true,"typeClass":"compound","value":[{"topicClassValue":{"typeName":"topicClassValue","multiple":false,"typeClass":"primitive","value":"Chemistry"}},{"topicClassValue":{"typeName":"topicClassValue","multiple":false,"typeClass":"primitive","value":"Machine Learning"}}]},{"typeName":"publication","multiple":true,"typeClass":"compound","value":[{"publicationCitation":{"typeName":"publicationCitation","multiple":false,"typeClass":"primitive","value":"Müller A, Cardenas-Cartagena J, Pollice R. Uncovering Internal Prediction Mechanisms of Transformer-Based Chemical Foundation Models. ChemRxiv. 2025.\nThis content is a preprint. Citation will be updated when the peer-reviewed article is available."},"publicationURL":{"typeName":"publicationURL","multiple":false,"typeClass":"primitive","value":"https://doi.org/10.26434/chemrxiv-2025-d1j8k"}}]},{"typeName":"language","multiple":true,"typeClass":"controlledVocabulary","value":["English"]},{"typeName":"producer","multiple":true,"typeClass":"compound","value":[{"producerName":{"typeName":"producerName","multiple":false,"typeClass":"primitive","value":"Pollice, Robert"},"producerAffiliation":{"typeName":"producerAffiliation","multiple":false,"typeClass":"primitive","value":"University of Groningen"}}]},{"typeName":"productionDate","multiple":false,"typeClass":"primitive","value":"2025-11-14"},{"typeName":"productionPlace","multiple":true,"typeClass":"primitive","value":["Groningen, The Netherlands"]},{"typeName":"contributor","multiple":true,"typeClass":"compound","value":[{"contributorType":{"typeName":"contributorType","multiple":false,"typeClass":"controlledVocabulary","value":"Researcher"},"contributorName":{"typeName":"contributorName","multiple":false,"typeClass":"primitive","value":"Müller, Alexander"}},{"contributorType":{"typeName":"contributorType","multiple":false,"typeClass":"controlledVocabulary","value":"Supervisor"},"contributorName":{"typeName":"contributorName","multiple":false,"typeClass":"primitive","value":"Cardenas-Cartagena, Juan"}},{"contributorType":{"typeName":"contributorType","multiple":false,"typeClass":"controlledVocabulary","value":"Project Leader"},"contributorName":{"typeName":"contributorName","multiple":false,"typeClass":"primitive","value":"Pollice, Robert"}}]},{"typeName":"distributor","multiple":true,"typeClass":"compound","value":[{"distributorName":{"typeName":"distributorName","multiple":false,"typeClass":"primitive","value":"DataverseNL network"},"distributorAffiliation":{"typeName":"distributorAffiliation","multiple":false,"typeClass":"primitive","value":"DANS"},"distributorAbbreviation":{"typeName":"distributorAbbreviation","multiple":false,"typeClass":"primitive","value":"DVNL"},"distributorURL":{"typeName":"distributorURL","multiple":false,"typeClass":"primitive","value":"https://dataverse.nl/"}}]},{"typeName":"depositor","multiple":false,"typeClass":"primitive","value":"Silva Correia Alves, Ana"},{"typeName":"dateOfDeposit","multiple":false,"typeClass":"primitive","value":"2025-11-20"},{"typeName":"dateOfCollection","multiple":true,"typeClass":"compound","value":[{"dateOfCollectionStart":{"typeName":"dateOfCollectionStart","multiple":false,"typeClass":"primitive","value":"2025-02-03"},"dateOfCollectionEnd":{"typeName":"dateOfCollectionEnd","multiple":false,"typeClass":"primitive","value":"2025-11-10"}}]},{"typeName":"kindOfData","multiple":true,"typeClass":"primitive","value":["Python code"]}]},"dansDataVaultMetadata":{"displayName":"Data Vault Metadata","name":"dansDataVaultMetadata","fields":[{"typeName":"dansDataversePid","multiple":false,"typeClass":"primitive","value":"doi:10.34894/SWCLWE"},{"typeName":"dansDataversePidVersion","multiple":false,"typeClass":"primitive","value":"1.0"},{"typeName":"dansBagId","multiple":false,"typeClass":"primitive","value":"urn:uuid:db63b969-9247-448e-98eb-70f50fccefe5"},{"typeName":"dansNbn","multiple":false,"typeClass":"primitive","value":"urn:nbn:nl:ui:13-6fd02369-1c37-4706-9d8e-6b3eb1103a2d"}]}},"files":[{"description":"Main branch of the GitLab repository that was used to generate all the results in the associated publication.","label":"mole-main.zip","restricted":false,"directoryLabel":"mole","version":1,"datasetVersionId":31948,"dataFile":{"id":565515,"persistentId":"","filename":"mole-main.zip","contentType":"application/zip","friendlyType":"ZIP Archive","filesize":360688196,"description":"Main branch of the GitLab repository that was used to generate all the results in the associated publication.","storageIdentifier":"surf://store:19aa6c6e0e4-6a59b0a03c2c","rootDataFileId":-1,"checksum":{"type":"SHA-1","value":"f46203ddac8a7e3158baf4832f86381969e92a7b"},"tabularData":false,"creationDate":"2025-11-21","publicationDate":"2025-11-28","directoryLabel":"mole","lastUpdateTime":"2025-11-28T09:16:57Z","fileAccessRequest":true}},{"label":"README.md","restricted":false,"version":1,"datasetVersionId":31948,"dataFile":{"id":566759,"persistentId":"","filename":"README.md","contentType":"text/markdown","friendlyType":"Markdown Text","filesize":8707,"storageIdentifier":"surf://store:19ac1b7de07-e160c39802a6","rootDataFileId":-1,"checksum":{"type":"SHA-1","value":"0677ff46bd477a4c708b167d7350857732f6973a"},"tabularData":false,"creationDate":"2025-11-26","publicationDate":"2025-11-28","lastUpdateTime":"2025-11-28T09:16:57Z","fileAccessRequest":true}},{"label":"requirements.txt","restricted":false,"version":1,"datasetVersionId":31948,"dataFile":{"id":566760,"persistentId":"","filename":"requirements.txt","contentType":"text/plain","friendlyType":"Plain Text","filesize":473,"storageIdentifier":"surf://store:19ac1b7dee8-7626a22d4c8e","rootDataFileId":-1,"checksum":{"type":"SHA-1","value":"afab19f4b090bc8d35990ad6690ac0118dfead59"},"tabularData":false,"creationDate":"2025-11-26","publicationDate":"2025-11-28","lastUpdateTime":"2025-11-28T09:16:57Z","fileAccessRequest":true}}]}}}