{"id":{"repo_id":"cambridge","oai_identifier":"oai:www.repository.cam.ac.uk:1810/395521"},"canonical_url":"https://search.dev.ndltd.org/etd/cambridge/oai:www.repository.cam.ac.uk:1810/395521","repository":{"repo_id":"cambridge","name":"Cambridge University","base_url":"https://api.repository.cam.ac.uk/server/oai/request"},"display":{"title":"Biomolecular Simulations with Machine Learning Potentials","abstract":"Computational modelling has become a key component of the drug discovery toolbox. Typically, these tools are driven by empirical forcefields, which are computationally efficient but approximate representations of the true quantum mechanical potential energy surface. Over the last several years, machine learned forcefields have emerged as a powerful paradigm, replacing manual parametrisation with a purely data driven approach and learning to reproduce the outputs of more expensive electronic structure calculations at greatly reduced computational cost. This thesis introduces new developments that enable modelling (bio)molecular systems with machine learned forcefields. First, the MACE-OFF series of forcefields are presented, which we show to be capable of highly accurate predictions for condensed phase systems. We then introduce a theoretical framework for performing condensed phase alchemical free energy calculations with MACE-OFF type forcefields. Proof of concept calculations are presented, demonstrating exceptional accuracy on small molecule hydration free energy benchmarks with modified MACE potentials. In the subsequent chapter, we further refine this approach by optimising the model hyperparameters and expanding the training data. We introduce an improved model, dubbed λ-MACE and study the accuracy and transferability of the potentials on challenging hydration free energy benchmarks. We then demonstrate the performance of the model on logP calculations on drug-like molecules taken from medicinal chemistry assays. Finally, selected applications of the MACE-MP0 potential are presented. The performance of this foundation model, which was fitted on inorganic crystals, is tested on out-of- distribution biomolecular systems, demonstrating qualitatively accurate predictions. We further demonstrate that these predictions can be improved by finetuning the foundation model on the system of interest.","abstract_html":"Computational modelling has become a key component of the drug discovery toolbox. Typically, these tools are driven by empirical forcefields, which are computationally efficient but approximate representations of the true quantum mechanical potential energy surface. Over the last several years, machine learned forcefields have emerged as a powerful paradigm, replacing manual parametrisation with a purely data driven approach and learning to reproduce the outputs of more expensive electronic structure calculations at greatly reduced computational cost. This thesis introduces new developments that enable modelling (bio)molecular systems with machine learned forcefields. First, the MACE-OFF series of forcefields are presented, which we show to be capable of highly accurate predictions for condensed phase systems. We then introduce a theoretical framework for performing condensed phase alchemical free energy calculations with MACE-OFF type forcefields. Proof of concept calculations are presented, demonstrating exceptional accuracy on small molecule hydration free energy benchmarks with modified MACE potentials. In the subsequent chapter, we further refine this approach by optimising the model hyperparameters and expanding the training data. We introduce an improved model, dubbed λ-MACE and study the accuracy and transferability of the potentials on challenging hydration free energy benchmarks. We then demonstrate the performance of the model on logP calculations on drug-like molecules taken from medicinal chemistry assays. Finally, selected applications of the MACE-MP0 potential are presented. The performance of this foundation model, which was fitted on inorganic crystals, is tested on out-of- distribution biomolecular systems, demonstrating qualitatively accurate predictions. We further demonstrate that these predictions can be improved by finetuning the foundation model on the system of interest.","abstract_has_math":false,"creators":["Moore, Harry"],"institution":"University of Cambridge","degree_name":"Doctor of Philosophy (PhD)","degree_level":"Doctoral","degree_discipline":null,"degree_department":null,"school":null,"contributors":[],"advisors":["Csanyi, Gabor"],"committee_chairs":[],"committee_members":[],"year":2025,"date_issued":"2025-06-27","date_published":"2025-06-27","updated_at":"2026-07-22T22:24:17Z","subjects":["Machine Learning","Molecular Simulation","Molecular Dynamics","Geometric Deep Learning","Statistical Thermodynamics","Drug Discovery","Computational Chemistry"],"languages":["eng"],"rights":[],"rights_urls":["https://www.repository.cam.ac.uk/bitstreams/337a03ab-a645-4a1b-ac6f-43bbe3de7551/download","http://purl.org/NET/rdflicense/allrightsreserved"],"identifier_entries":[{"key":"dc:creator.authoridentifier","label":"Author Identifier","values":["0000000151163027"],"render_values":[{"text":"0000-0001-5116-3027","href":"https://orcid.org/0000-0001-5116-3027","code":true}]}]},"links":{"outbound_url":"https://doi.org/10.17863/CAM.124993","outbound_label":"DOI","outbound_source":"dc:identifier.doi"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor.advisor","label":"Advisor","values":["Csanyi, Gabor"]},{"key":"dc:contributor.sponsor","label":"Sponsor","values":["AstraZeneca Non-Clinical Studentship"]},{"key":"dc:creator","label":"Author","values":["Moore, Harry"]},{"key":"dc:creator.authoridentifier","label":"Author Identifier","values":["0000000151163027"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date.issued","label":"Date","values":["2025-06-27"]},{"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/1810/395521"]},{"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":["Machine Learning","Molecular Simulation","Molecular Dynamics","Geometric Deep Learning","Statistical Thermodynamics","Drug Discovery","Computational Chemistry"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["eng"]},{"key":"dc:rights","label":"Dc Rights","values":["https://www.repository.cam.ac.uk/bitstreams/337a03ab-a645-4a1b-ac6f-43bbe3de7551/download","http://purl.org/NET/rdflicense/allrightsreserved"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier.doi","label":"DOI","values":["https://doi.org/10.17863/CAM.124993"]},{"key":"dc:identifier.uri","label":"Identifier URI","values":["https://www.repository.cam.ac.uk/bitstreams/e9aca783-d8e5-4e51-93bd-4c0690a3413b/download"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description.abstract","label":"Abstract","values":["Computational modelling has become a key component of the drug discovery toolbox. Typically, these tools are driven by empirical forcefields, which are computationally efficient but approximate representations of the true quantum mechanical potential energy surface. Over the last several years, machine learned forcefields have emerged as a powerful paradigm, replacing manual parametrisation with a purely data driven approach and learning to reproduce the outputs of more expensive electronic structure calculations at greatly reduced computational cost. This thesis introduces new developments that enable modelling (bio)molecular systems with machine learned forcefields. First, the MACE-OFF series of forcefields are presented, which we show to be capable of highly accurate predictions for condensed phase systems. We then introduce a theoretical framework for performing condensed phase alchemical free energy calculations with MACE-OFF type forcefields. Proof of concept calculations are presented, demonstrating exceptional accuracy on small molecule hydration free energy benchmarks with modified MACE potentials. In the subsequent chapter, we further refine this approach by optimising the model hyperparameters and expanding the training data. We introduce an improved model, dubbed λ-MACE and study the accuracy and transferability of the potentials on challenging hydration free energy benchmarks. We then demonstrate the performance of the model on logP calculations on drug-like molecules taken from medicinal chemistry assays. Finally, selected applications of the MACE-MP0 potential are presented. The performance of this foundation model, which was fitted on inorganic crystals, is tested on out-of- distribution biomolecular systems, demonstrating qualitatively accurate predictions. We further demonstrate that these predictions can be improved by finetuning the foundation model on the system of interest."]},{"key":"dc:format.checksum.md5","label":"Dc Format Checksum Md5","values":["cae9e6f15efaa5daa8aa1860d097c6ca","87eda9de84448d1f82354d60eee3eb5f"]},{"key":"dc:title","label":"Title","values":["Biomolecular Simulations with Machine Learning Potentials"]}]}],"canonical_facts":{"dc:contributor.advisor":["Csanyi, Gabor"],"dc:contributor.sponsor":["AstraZeneca Non-Clinical Studentship"],"dc:creator":["Moore, Harry"],"dc:creator.authoridentifier":["0000000151163027"],"dc:date.issued":["2025-06-27"],"dc:description.abstract":["Computational modelling has become a key component of the drug discovery toolbox. Typically, these tools are driven by empirical forcefields, which are computationally efficient but approximate representations of the true quantum mechanical potential energy surface. Over the last several years, machine learned forcefields have emerged as a powerful paradigm, replacing manual parametrisation with a purely data driven approach and learning to reproduce the outputs of more expensive electronic structure calculations at greatly reduced computational cost. This thesis introduces new developments that enable modelling (bio)molecular systems with machine learned forcefields. First, the MACE-OFF series of forcefields are presented, which we show to be capable of highly accurate predictions for condensed phase systems. We then introduce a theoretical framework for performing condensed phase alchemical free energy calculations with MACE-OFF type forcefields. Proof of concept calculations are presented, demonstrating exceptional accuracy on small molecule hydration free energy benchmarks with modified MACE potentials. In the subsequent chapter, we further refine this approach by optimising the model hyperparameters and expanding the training data. We introduce an improved model, dubbed λ-MACE and study the accuracy and transferability of the potentials on challenging hydration free energy benchmarks. We then demonstrate the performance of the model on logP calculations on drug-like molecules taken from medicinal chemistry assays. Finally, selected applications of the MACE-MP0 potential are presented. The performance of this foundation model, which was fitted on inorganic crystals, is tested on out-of- distribution biomolecular systems, demonstrating qualitatively accurate predictions. We further demonstrate that these predictions can be improved by finetuning the foundation model on the system of interest."],"dc:format.checksum.md5":["cae9e6f15efaa5daa8aa1860d097c6ca","87eda9de84448d1f82354d60eee3eb5f"],"dc:identifier.doi":["https://doi.org/10.17863/CAM.124993"],"dc:identifier.uri":["https://www.repository.cam.ac.uk/bitstreams/e9aca783-d8e5-4e51-93bd-4c0690a3413b/download"],"dc:language":["eng"],"dc:publisher.institution":["University of Cambridge"],"dc:relation.isreferencedby.uri":["https://www.repository.cam.ac.uk/handle/1810/395521"],"dc:rights":["https://www.repository.cam.ac.uk/bitstreams/337a03ab-a645-4a1b-ac6f-43bbe3de7551/download","http://purl.org/NET/rdflicense/allrightsreserved"],"dc:subject":["Machine Learning","Molecular Simulation","Molecular Dynamics","Geometric Deep Learning","Statistical Thermodynamics","Drug Discovery","Computational Chemistry"],"dc:title":["Biomolecular Simulations with Machine Learning Potentials"],"dc:type":["Thesis"],"dc:type.qualificationlevel":["Doctoral"],"dc:type.qualificationname":["Doctor of Philosophy (PhD)"]},"updated_at":"2026-07-22T22:24:17Z"}