# SFU Researchers Present CGFlow for Drug Design > Source: > Published: 2026-08-23 14:00:45+00:00 # SFU Researchers Present CGFlow for Drug Design Simon Fraser University researchers presented CGFlow in July and August 2025, introducing an AI framework that designs drug molecules alongside feasible chemical synthesis routes. SFU coverage describes the method as combining stepwise molecular construction with 3D modeling to address a recurring limitation in computational drug design: candidates that appear promising but cannot be made in a laboratory. Simon Fraser University researchers have presented **CGFlow**, an AI framework intended to generate drug-like molecules together with practical routes for synthesizing them. The Vancouver Sun reported on August 23, 2026 that SFU researchers are advancing AI tools for discovering treatments for diseases including cancer, while SFU's 2025 coverage identifies CGFlow as the underlying research method. According to SFU, a central problem in AI-assisted molecular design is **synthesizability**. Models can propose structures predicted to bind to a disease target, but many candidates cannot be produced through realistic laboratory chemistry. SFU professor Martin Ester said in the university's August 2025 media release that drug development is commonly described as taking 10 years and $1 billion, and expressed hope that the method could shorten that process. ### Jointly modeling structure and synthesis SFU describes CGFlow as a dual-design method that models both a molecule's compositional structure and its continuous 3D state. The framework builds molecules incrementally rather than generating a full structure in one pass. Its compositional component uses Generative Flow Networks, or GFlowNets, to explore high-reward molecular structures and associated reaction steps; the state component models molecular geometry in three dimensions. Lead author Tony Shen described the drug-design task as identifying a disease-causing protein and designing molecules that bind to it, potentially deactivating harmful activity. The practical challenge is that binding-related predictions alone do not provide the chemical recipe needed to create a candidate. For ML practitioners, the work illustrates a broader design pattern in scientific generation: optimizing a target property and optimizing operational feasibility are distinct tasks. Generative systems that explicitly represent downstream constraints can reduce the gap between computationally attractive outputs and candidates suitable for experimental follow-up. ### Evidence still needed beyond generation The SFU materials describe CGFlow as research published at a leading conference, but they do not report laboratory synthesis results, preclinical efficacy results, or clinical testing for drugs generated by the system. Those stages remain necessary to establish whether generated candidates are safe, effective, and manufacturable at pharmaceutical scale. In comparable drug-discovery workflows, synthesis-aware generation can be valuable because chemists need a tractable route before investing in validation assays. Whether a framework improves real development timelines, however, depends on prospective experimental validation and performance across target classes, not solely on molecular-generation benchmarks. ## Key Points - 1CGFlow combines molecular generation with synthesis-route modeling, targeting candidates that are both structurally plausible and feasible for laboratory production. - 2The framework uses GFlowNets and 3D molecular state modeling to represent synthesis and geometry during generation. - 3SFU's materials report no clinical or preclinical validation, so prospective synthesis and assay results remain the key measure of practical value. ## Scoring Rationale CGFlow addresses a technically important gap between molecular generation and synthetic feasibility, which is relevant to computational chemistry and drug-discovery teams. The available sources describe a research framework rather than validated drug candidates or demonstrated clinical outcomes, limiting its immediate practitioner impact. ## Sources Primary source and supporting public references used for this report. Practice interview problems based on real data 1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with. [Try 250 free problems](/problems)