Generative Synthetic Data Models for Pre-Clinical Drug Discovery

dc.contributor.authorHU, BING
dc.date.accessioned2026-08-28T15:24:57Z
dc.date.issued2026-08-28
dc.date.submitted2026-08-24
dc.description.abstractBringing a new drug to market typically requires 2-3 billion and 10-15 years, with the majority of candidates failing due to poor pharmacokinetic properties or insufficient efficacy. Artificial intelligence has the potential to dramatically accelerate pre-clinical drug discovery, but a fundamental obstacle stands in the way: sparse overlap between datasets. Drug discovery datasets are collected independently, covering different molecules and their properties with little overlap, making it impossible to train the kind of large, unified foundation models that have revolutionized natural language processing and computer vision. This thesis presents a systematic research approach to overcome data overlap sparsity through generative synthetic data, developing four interconnected models that progressively advance toward a generative foundation model for pre-clinical drug discovery. Imagand is a novel SMILES-to-Pharmacokinetic (S2PK) diffusion model that simultaneously generates 12 pharmacokinetic (PK) target properties conditioned on learned molecular SMILES embeddings. The key technical contribution is the Discrete Local Gaussian Noise (DLGN) model, which creates a prior distribution closer to the true data distribution, improving synthetic data quality. Imagand achieves an average Hellinger Distance of 0.18 across 12 PK properties and demonstrates statistically significant machine learning efficiency improvements over real data for multiple PK datasets. DELBERT-2 is a pretrained fingerprint language model for DNA-Encoded Library (DEL) virtual screening, building on the prior work of DELBERT. By converting sparse molecular fingerprints into unified token sequences and pretraining a ModernBERT encoder via masked language modeling on 2.5M molecules from the AIRCHECK DEL corpus, DELBERT-2 learns transferable molecular representations that generalize to out-of-distribution chemical space. Evaluated across six protein targets and three complementary out-of-distribution protocols, DELBERT-2 achieves 13.28$\pm$3.72$\times$ increased enrichment at top-100, representing a 7.4\% improvement over no-pretraining baselines. xImagand-DKI is a novel multi-view SMILES/Protein-to-PK/DTI diffusion model that jointly generates pharmacokinetic and drug-target interaction (DTI) synthetic data, conditioned on molecular and protein representations enriched with domain knowledge. Domain knowledge integration via molecular fingerprints (FPFormer, precursor to DELBERT) and Gene Ontology embeddings (PO2Vec) consistently improves synthetic data quality, achieving an average Hellinger Distance of 0.11. xImagand-DKI is the first model to jointly generate synthetic data spanning both PK and DTI datasets, directly addressing data overlap sparsity across task boundaries. Syngand is an end-to-end diffusion model that jointly generates novel molecular structures and their pharmacokinetic properties in a single pipeline, extending the DiGress molecular graph diffusion model. A novel sampling methodology enables the generation of PK properties for any existing ligand. Syngand achieves statistically significant machine learning efficiency improvements over real data for 5 of 6 PK datasets. Together, these four components establish a progressive advance toward a foundation generative AI model for pre-clinical drug discovery, providing a practical toolkit for overcoming data overlap sparsity and enabling new research questions in polypharmacy, drug combination, and high-throughput screening.
dc.identifier.urihttps://hdl.handle.net/10012/24103
dc.language.isoen
dc.pendingfalse
dc.publisherUniversity of Waterlooen
dc.subjectDrug Discovery
dc.subjectArtificial Intelligence
dc.subjectSynthetic Data
dc.subjectGenerative Models
dc.subjectDeep Learning
dc.titleGenerative Synthetic Data Models for Pre-Clinical Drug Discovery
dc.typeDoctoral Thesis
uws-etd.degreeDoctor of Philosophy
uws-etd.degree.departmentDavid R. Cheriton School of Computer Science
uws-etd.degree.disciplineComputer Science
uws-etd.degree.grantorUniversity of Waterlooen
uws-etd.embargo.terms0
uws.contributor.advisorChen, Helen
uws.contributor.advisorLayton, Anita
uws.contributor.affiliation1Faculty of Mathematics
uws.peerReviewStatusUnrevieweden
uws.published.cityWaterlooen
uws.published.countryCanadaen
uws.published.provinceOntarioen
uws.scholarLevelGraduateen
uws.typeOfResourceTexten

Files

Original bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
Hu_Bing.pdf
Size:
6.02 MB
Format:
Adobe Portable Document Format

License bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
license.txt
Size:
6.4 KB
Format:
Item-specific license agreed upon to submission
Description: