{"id":{"repo_id":"cambridge","oai_identifier":"oai:www.repository.cam.ac.uk:1810/386730"},"canonical_url":"https://search.dev.ndltd.org/etd/cambridge/oai:www.repository.cam.ac.uk:1810/386730","repository":{"repo_id":"cambridge","name":"Cambridge University","base_url":"https://api.repository.cam.ac.uk/server/oai/request"},"display":{"title":"Improving ligand discovery using deep learning on three-dimensional structural data","abstract":"Deep learning methods offer the opportunity to improve precision and speed in various stages of drug discovery, such as molecular property prediction or goal-directed molecular generation. Incorporating three-dimensional (3D) structural data allows to steer prediction and generation towards 3D-dependant properties that are fundamental in hit identification and lead optimization. In this thesis, I first introduce existing deep learning applications in drug discovery, emphasizing on structure-based 3D generative models that produce molecules based on binding pocket context for high-affinity molecule inception. In a first research chapter, I describe the application of atomistic neural network to rank conformations of molecules tested in ligand-based or structure-based virtual screening, showing an early enrichment of bioactivelike conformations, allowing to accelerate virtual screening by testing less conformations. Secondly, I describe the benchmark I elaborated to assess the performances of structure-based 3D generative models, highlighting that the diverse set of tested models generate molecules with a poor structural quality undetected by scoring functions estimating binding affinity. Thirdly, I detail my reinforcement learning model for fragment-based generation within a binding pocket, showing better conformation quality for similar estimated binding affinity compared to previous models, while ensuring good synthesizability to allow further in vitro activity assessment. This research shows the importance of setting up appropriate benchmark for fair model performance evaluations, while proposing effective and innovative solutions to overcome known limitations of earlier models relying on deep learning of known structural data.","abstract_html":"Deep learning methods offer the opportunity to improve precision and speed in various stages of drug discovery, such as molecular property prediction or goal-directed molecular generation. Incorporating three-dimensional (3D) structural data allows to steer prediction and generation towards 3D-dependant properties that are fundamental in hit identification and lead optimization. In this thesis, I first introduce existing deep learning applications in drug discovery, emphasizing on structure-based 3D generative models that produce molecules based on binding pocket context for high-affinity molecule inception. In a first research chapter, I describe the application of atomistic neural network to rank conformations of molecules tested in ligand-based or structure-based virtual screening, showing an early enrichment of bioactivelike conformations, allowing to accelerate virtual screening by testing less conformations. Secondly, I describe the benchmark I elaborated to assess the performances of structure-based 3D generative models, highlighting that the diverse set of tested models generate molecules with a poor structural quality undetected by scoring functions estimating binding affinity. Thirdly, I detail my reinforcement learning model for fragment-based generation within a binding pocket, showing better conformation quality for similar estimated binding affinity compared to previous models, while ensuring good synthesizability to allow further in vitro activity assessment. This research shows the importance of setting up appropriate benchmark for fair model performance evaluations, while proposing effective and innovative solutions to overcome known limitations of earlier models relying on deep learning of known structural data.","abstract_has_math":false,"creators":["Baillif, Benoît"],"institution":"University of Cambridge","degree_name":"Doctor of Philosophy (PhD)","degree_level":"Doctoral","degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":["Bender, Andreas"],"committee_chairs":[],"committee_members":[],"year":2024,"date_issued":"2024-11-30","date_published":"2024-11-30","updated_at":"2026-07-22T22:24:25Z","subjects":["deep learning","three-dimensional data","generative model","bioactive","drug design"],"languages":[],"rights":[],"rights_urls":["https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/cc3a7a44-5f8d-41da-ae97-ee4fbfbfebc3/download","https://creativecommons.org/licenses/by/4.0/"],"identifier_entries":[{"key":"dc:creator.authoridentifier","label":"Author Identifier","values":["0000000211398152"],"render_values":[{"text":"0000-0002-1139-8152","href":"https://orcid.org/0000-0002-1139-8152","code":true}]}]},"links":{"outbound_url":"https://doi.org/10.17863/CAM.119823","outbound_label":"DOI","outbound_source":"dc:identifier.doi"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Bender, Andreas"]},{"key":"dc:contributor.sponsor","label":"Sponsor","values":["Cambridge Crystallographic Data Centre"]},{"key":"dc:creator","label":"Author","values":["Baillif, Benoît"]},{"key":"dc:creator.authoridentifier","label":"Author Identifier","values":["0000000211398152"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.issued","label":"Date","values":["2024-11-30"]},{"key":"dc:publisher.institution","label":"Dc Publisher Institution","values":["University of Cambridge"]},{"key":"dc:relation.isreferencedby.uri","label":"Dc Relation Isreferencedby URI","values":["https://www.repository.cam.ac.uk/handle/1810/386730"]},{"key":"dc:type","label":"Dc Type","values":["Thesis"]},{"key":"dc:type.qualificationlevel","label":"Dc Type Qualificationlevel","values":["Doctoral"]},{"key":"dc:type.qualificationname","label":"Dc Type Qualificationname","values":["Doctor of Philosophy (PhD)"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["deep learning","three-dimensional data","generative model","bioactive","drug design"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:rights","label":"Dc Rights","values":["https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/cc3a7a44-5f8d-41da-ae97-ee4fbfbfebc3/download","https://creativecommons.org/licenses/by/4.0/"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.doi","label":"DOI","values":["https://doi.org/10.17863/CAM.119823"]},{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/dd125ecd-2068-4380-aada-78e07aec4e8a/download"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Deep learning methods offer the opportunity to improve precision and speed in various stages of drug discovery, such as molecular property prediction or goal-directed molecular generation. Incorporating three-dimensional (3D) structural data allows to steer prediction and generation towards 3D-dependant properties that are fundamental in hit identification and lead optimization. In this thesis, I first introduce existing deep learning applications in drug discovery, emphasizing on structure-based 3D generative models that produce molecules based on binding pocket context for high-affinity molecule inception. In a first research chapter, I describe the application of atomistic neural network to rank conformations of molecules tested in ligand-based or structure-based virtual screening, showing an early enrichment of bioactivelike conformations, allowing to accelerate virtual screening by testing less conformations. Secondly, I describe the benchmark I elaborated to assess the performances of structure-based 3D generative models, highlighting that the diverse set of tested models generate molecules with a poor structural quality undetected by scoring functions estimating binding affinity. Thirdly, I detail my reinforcement learning model for fragment-based generation within a binding pocket, showing better conformation quality for similar estimated binding affinity compared to previous models, while ensuring good synthesizability to allow further in vitro activity assessment. This research shows the importance of setting up appropriate benchmark for fair model performance evaluations, while proposing effective and innovative solutions to overcome known limitations of earlier models relying on deep learning of known structural data."]},{"key":"dc:format.checksum.md5","label":"Dc Format Checksum Md5","values":["1181bf7091590140f2cefa737b0434cb","87eda9de84448d1f82354d60eee3eb5f"]},{"key":"dc:title","label":"Title","values":["Improving ligand discovery using deep learning on three-dimensional structural data"]}]}],"canonical_facts":{"dc:contributor.advisor":["Bender, Andreas"],"dc:contributor.sponsor":["Cambridge Crystallographic Data Centre"],"dc:creator":["Baillif, Benoît"],"dc:creator.authoridentifier":["0000000211398152"],"dc:date.issued":["2024-11-30"],"dc:description.abstract":["Deep learning methods offer the opportunity to improve precision and speed in various stages of drug discovery, such as molecular property prediction or goal-directed molecular generation. Incorporating three-dimensional (3D) structural data allows to steer prediction and generation towards 3D-dependant properties that are fundamental in hit identification and lead optimization. In this thesis, I first introduce existing deep learning applications in drug discovery, emphasizing on structure-based 3D generative models that produce molecules based on binding pocket context for high-affinity molecule inception. In a first research chapter, I describe the application of atomistic neural network to rank conformations of molecules tested in ligand-based or structure-based virtual screening, showing an early enrichment of bioactivelike conformations, allowing to accelerate virtual screening by testing less conformations. Secondly, I describe the benchmark I elaborated to assess the performances of structure-based 3D generative models, highlighting that the diverse set of tested models generate molecules with a poor structural quality undetected by scoring functions estimating binding affinity. Thirdly, I detail my reinforcement learning model for fragment-based generation within a binding pocket, showing better conformation quality for similar estimated binding affinity compared to previous models, while ensuring good synthesizability to allow further in vitro activity assessment. This research shows the importance of setting up appropriate benchmark for fair model performance evaluations, while proposing effective and innovative solutions to overcome known limitations of earlier models relying on deep learning of known structural data."],"dc:format.checksum.md5":["1181bf7091590140f2cefa737b0434cb","87eda9de84448d1f82354d60eee3eb5f"],"dc:identifier.doi":["https://doi.org/10.17863/CAM.119823"],"dc:identifier.uri":["https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/dd125ecd-2068-4380-aada-78e07aec4e8a/download"],"dc:publisher.institution":["University of Cambridge"],"dc:relation.isreferencedby.uri":["https://www.repository.cam.ac.uk/handle/1810/386730"],"dc:rights":["https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/cc3a7a44-5f8d-41da-ae97-ee4fbfbfebc3/download","https://creativecommons.org/licenses/by/4.0/"],"dc:subject":["deep learning","three-dimensional data","generative model","bioactive","drug design"],"dc:title":["Improving ligand discovery using deep learning on three-dimensional structural data"],"dc:type":["Thesis"],"dc:type.qualificationlevel":["Doctoral"],"dc:type.qualificationname":["Doctor of Philosophy (PhD)"]},"updated_at":"2026-07-22T22:24:25Z"}