Machine Learning-Designed Molecule Removes Particles from Molecular Cage
Machine Learning-Designed Molecule Removes Particles from Molecular Cage
As medical care evolves to target more personalized and effective therapies, scientists aim to introduce specific molecules into biological systems to perform precise actions. Examples include gene therapy and drug delivery, both of which must be both cost-effective and efficient for broad usage. In pursuit of this goal, a team of researchers has utilized machine learning to develop a method for removing molecules from within a molecular cage. Their study is featured in Physical Review Letters.
Led by Ryan K. Krueger from Harvard University, the research employs differentiable molecular dynamics to design complex reactions that guide a system toward specific outcomes. For example, the researchers focused on the controlled disassembly of colloidal structures, particularly designing a molecule capable of removing a particle enclosed by a complete "cage" of colloidal particles. Colloids, such as milk or gelatin, consist of microscopic particles dispersed within another substance. Machine learning was employed to optimize the design of the "opener" molecule for the shell, which the team referred to as the "spider" due to its geometric shape. The researchers note that disassembly is key to biological processes like defect repair, self-replication, and catalysis.
The team specifically designed a molecule to disassemble icosahedral shells, which are composed of 12 particles with 30 outer edges connecting the shell components. This structure is similar to protein capsids that house viruses. The particles in these shells are "patchy," meaning their interactions with each other and with the caged particle have specific directional and strength-based parameters. First introduced in soft material research two decades ago, patchiness provides flexibility in tailoring interactions to achieve desired behaviors, aided by recent advancements in patchy particle simulations within a differentiable library. In this study, the patchiness was applied to the 12 shell particles. The goal was to disassemble the shell while maintaining the integrity of the remaining structure. The researchers used a Morse potential model to describe the energy of the interactions, which requires three free parameters that can be adjusted to fit the situation. Removing the caged particle necessitates removing one of the shell particles.
To test this, the team modeled the "spider" as a rigid pyramid structure with a base formed by a pentagon-shaped ring of particles and a single "head particle" at the top. This spider was designed to land on any shell particle and interact with it, with patch parameters tuned so that the spider itself would neither attract nor repel the shell cluster. The head particle, however, would interact with the shell particles based on the distance and strength of the force.
The researchers used molecular dynamics, a technique that calculates particle motion based on the forces acting between them, to determine which spider parameters would successfully extract the caged particle. Attempting this manually through brute force—calculating the interactions for all possible parameters—would have been computationally expensive. Instead, the team applied machine learning to minimize a loss function that balanced disassembly with the preservation of the remaining substructure.
The machine learning approach allowed the creation of a rigid spider capable of performing the removal task. The team then introduced a new parameter, "configurable entropy," which allowed the spider to flex and adjust its configuration.
