Machine-Learning-Assisted Inverse Design of Biopharmaceutical Formulations and Delivery Systems

Abstract

Pharmaceutical formulation and delivery-device development has historically proceeded by forward trial and error: a candidate material, formulation, or geometry is proposed, fabricated, and tested, and the cycle repeats until an acceptable product emerges. Machine learning enables an inverse alternative: given a target product profile — a desired release rate, gastric-residence time, transfection efficiency, or shape transformation — a model proposes the material composition or device geometry predicted to achieve it, collapsing what was previously dozens of design-build-test cycles into a handful. This review develops that inverse-design paradigm specifically for biopharmaceutical formulation and delivery, grounding each technical claim in verifiable, named research from the Massachusetts Institute of Technology and Harvard University. We examine four case studies: the directed message-passing graph neural network that Stokes, Yang, Collins, Barzilay, Jaakkola and colleagues at MIT used to discover the antibiotic halicin from a chemical library of over one hundred million candidates; the geometrically optimized, origami-inspired ingestible device that Traverso, Langer and colleagues at MIT and Brigham and Women’s Hospital/Harvard Medical School designed to self-deploy into a gastric-retentive torus; the algorithmically encoded, biomimetic 4D-printing platform that Lewis, Mahadevan and colleagues at the Harvard Wyss Institute developed to program anisotropic hydrogel swelling into predetermined shape changes; and the autonomous, machine-learning-guided closed-loop synthesis and formulation systems that Jensen and colleagues in the MIT Department of Chemical Engineering have advanced as “self-driving laboratories.” From these cases we distill a general method taxonomy — Bayesian optimization, generative models, graph neural networks, physics-informed machine learning coupled to finite-element simulation, and active-learning closed loops — and map each onto concrete biopharmaceutical inverse-design problems: ionizable-lipid and lipid-nanoparticle design for mRNA delivery, shape-memory device and 4D-printed geometry optimization, and controlled-release formulation screening. We close by examining the regulatory posture of the U.S. Food and Drug Administration toward artificial intelligence in drug development, including its January 2025 draft guidance and the January 2026 joint FDA-EMA guiding principles, and by outlining the data, validation, and interpretability barriers that stand between today’s proof-of-concept demonstrations and routine, regulator-accepted inverse design of biopharmaceutical products.