Optimization Approaches to Algorithmic Decision Making in Molecular Discovery
Name
fromer-jfromer-phd-cheme-2026-thesis.pdf
Size
18.61 MB
Format
Adobe PDF
Checksum (MD5)
41bf1ef4b34c681be55aca7e3dedb995
Author(s)
Fromer, Jenna C.
Advisor(s)
Coley, Connor W.
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
Drug discovery is an optimization problem that seeks to identify molecules that satisfy multiple design criteria. The discovery of small molecule drugs often involves computationally screening a molecular library to identify molecules predicted to bind to a target protein, followed by experimental validation and additional structural optimization of top-performing molecules. Each of these stages involves time-consuming physical and computational experiments. This dissertation involves the development and application of optimization tools to support hit identification, design-make-test cycles, and chemical space exploration.
This thesis begins with the development of Bayesian optimization methods to accelerate virtual screening. A novel acquisition function, multipoint Probability of Optimality, is introduced to improve sample efficiency in model-guided virtual screening. Further, while model-guided active learning has proven effective for optimizing individual predicted properties, single-objective optimization defers critical trade-offs to later stages of design. This thesis applies Pareto optimization to retrospective multi-objective virtual screening tasks.
Next, this thesis introduces an optimization framework, termed SPARROW, for simultaneous compound and synthetic route selection in design-make-test cycles. The selection of which compounds to synthesize and test from all possibilities is a complex decision-making process that must consider both experimental resource constraints and the perceived value of candidates. SPARROW formalizes this decision-making process by jointly optimizing compound utility and synthetic cost, accommodating flexible definitions of both objectives.
Compound downselection in design-make-test cycles is inherently limited to the chemical space known to be synthetically accessible; however, compounds with enhanced property profiles may exist in yet-unexplored chemical space. This thesis proposes a cheminformatics pipeline to aid in the systematic prioritization of unprecedented scaffolds for synthesis by assessing novelty, stability, and synthetic tractability. This approach aims to guide the exploration of new chemical space in a targeted and experimentally actionable manner.
Finally, this thesis concludes by describing two efforts to apply algorithmic design and downselection methods to prospective molecular discovery projects.
Altogether, this thesis demonstrates how optimization methods can support virtual screening, compound selection, and scaffold exploration in small molecule drug discovery.
MIT Department
Massachusetts Institute of Technology. Department of Chemical Engineering
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
