Scientists use machine learning to develop molecular can opener

In an era of medicine that strives for more targeted drug therapies, more individualized therapies, and more effective treatments, doctors and scientists want to introduce molecules into biological systems that can perform specific actions. Masu.

Examples include gene therapy and drug delivery, which need to be effective and inexpensive for widespread use. To achieve this goal, three researchers used machine learning to design a method to remove molecules inside molecular cages. Their research is published in Physical Review Letters.

The study, led by Harvard University’s Ryan K. Krueger but with equal contributions from each co-author, uses differentiable molecular dynamics to design complex reactions and transform systems into specific lead to results.

As an example, they worked to design molecules capable of controlled disassembly of colloidal structures, specifically removing particles that are surrounded and bound by a complete shell or “cage” of colloidal particles. (A colloid is a mixture of substances in which nanoscale or microscale insoluble particles are dispersed throughout another substance. Examples include milk, smoke, and gelatin.)

Machine learning was used to optimize the design of the shell’s “opener” molecule, which they call a “spider” because of its shape. As they write, “degradation is central to the dynamic functions of living systems, including defect repair, self-replication, and catalysis.”

In particular, they designed a controlled decomposition of an icosahedral shell, a collection of 12 particles with 30 outer edges connecting the shell particles. This organization is very similar to the protein capsids that house viruses.

Shell particles are considered “speckled”. Interactions with other shell particles and particles within the cage have specific values ​​of parameters that determine the directionality and relative strength of the interactions. Patchiness, introduced into soft materials research two decades ago, has been aided by recent developments in patchy particle simulations within differentiable libraries, providing versatile tuning capabilities for designed interactions and Achieve a specific behavior.

Spots can even vary across the surface of the patchy particles. We have 12 individual shell particles here. The goal was to disassemble the shell, and there was an inherent tension between completing the disassembly while maintaining the integrity of the remaining substructure.

The researchers assumed a Morse potential as the potential energy of the interacting shell particles. It is often used as a model for interactions between two atoms in diatomic molecules and with caged molecules.

Morse potential is simple, with three free parameters that can (and should) be chosen to suit the desired situation. To remove the cage particle, one of the shell particles must be removed.

For their analysis, the researchers assumed that the object removing the shell particles was a rigid pyramid-shaped structure that fit on top of a cluster of 12 spheres. They called this object a “spider.” It consisted of a ring of pentagonal particles forming the base of a pyramid, and a single “head particle” at the top of the pyramidal mass.

In their simulation, an icosahedral shell is given and fixed, and the spider is free to land on and interact with any shell particle.

The patch parameters were adjusted so that the entire spider was neither attracted nor repelled by the cluster of shells, but the particles at the top of the pyramid were attracted to the patch on the shell particle by distance and a force that varied with distance. Ta. strength. You can also adjust the spider’s dimensions and the radius of its head and base particles.

Krueger and his collaborators used molecular dynamics. This is a standard technique that calculates the motion of each particle by the forces of interaction it experiences with other particles. They wanted to determine which specific parameters of the spider cause caged molecules to be pulled out of the shell.

Doing this in a brute-force manner on a computer would require an enormous amount of computing power and time, calculating all possible parameters for each particle until the desired result is obtained. So the group turned to machine learning to minimize a loss function that represents the tension between decomposition and the integrity of the remaining underlying structure.

Through this process, they were able to create a rigid spider that could complete the removal task. Next, we made the spider flexible and introduced a new free parameter that represents “configurable entropy.”

A similar optimization reduced the energy required to release the particles inside the cage. They show that spiders with asymmetrically flexible base legs require less energy to release caged particles compared to spiders with symmetric pentagonal bases as originally envisioned. I discovered that there are few.

They noted that their methodology is broadly applicable. “Because we are optimizing directly with respect to numerically integrated dynamics, our method is general enough to study a wide range of systems,” they write.

“First and foremost, it could enable the experimental realization of theoretical models that have been limited by the inability to fine-tune the interaction energies.”

© 2024 Science X Network