Team: 27
School: Los Alamos High
Area of Science: Computational Biology
Interim:
Biological proteins are nature's molecular tools for doing everything. Science magazine named protein structure prediction the breakthrough of the year. The next level to this is protein design. Designing a protein could mean all sorts of extremely powerful biotech, like targeted drugs, treatments that bind COVID or other diseases, and more. There are unlimited possibilities in material science and chemsitry: we could control the very placement and arrangement of individual atoms!
A protein is a sequence of amino acids that folds to a specific structure. Structure prediction means predicting the 3d backbone conformation of a protein given a sequence of amino acids. Protein design means starting from the 3d shape you want for a specific function and predicting a sequence that would fold to that structure, so we could produce that protein with the intended function in a lab. However, in a typical protein with 200 or more amino acids, and 20 amino acid possibilities with perhaps 10 different rotations of each, the number of possible sequences is 200^(20*10), which is astronomically large. Protein design is a NP-complete problem, but we can't check this many possible sequences.
There are classical algorithms that can find decent solutions, but it may be impossible to find the optimal solution classically. Quantum computers, however, may be able to search the entire solution space. Dr. Mulligan et al. successfully designed proteins with a novel quantum computing algorithm. However, current quantum computing technology isn't perfectly adiabatic, leading to lots of noise in the outcomes, and has so few connections and qubits only small problems can be used with them.
Another technology has shown promise in solving these kind of combinatoric problems: neuromorphic computing. This hardware can be much more highly connected than a quantum annealer. Designed after the brain, neuromorphic computing involves a population of leaky integrate and fire neurons communicating asynchronously in massive parallel through transmission lines with voltage pulses. These systems use 1000s of times less energy than digital computers. In a project from a previous year, I showed the promise of this tech by simulating the brain's visual cortex neuromorphically to perform image stabilization and recognition.
The goal of this project is to investigate whether neuromorphic computing is a useful technology for protein design.
I've designed methods to process protein structure data to create Q-matrices, which define the design problem as Quadratic Unconstrained Binary Optimization (QUBO) in python. As the name suggests, a QUBO problem has no constraints, so I designed a method to incorporate constraints with regularizing penalty terms, also in python. To avoid confusion to the reader, quantum and neuromorphic computers are not programmed in python. They don't take any sort of programming language. Instead, the Q-matrix, a square tensor of values, defines a set of bits (neurons or qubits) and how they're connected, such that the hardware will settle into a solution with low energy. I've loaded these Q-matrices on the D-Wave quantum annealer and Intel Loihi neuromorphic computer , which produced proposed solutions to protein packing problems. I've compared the energy of the solutions from these machines to classical algorithms, to see which can produce lower energy (better, more likely to fold) sequences.
Currently, my results show classical algorithms outperforming quantum and neuromorphic hardware, finding solutions fitting the softly imposed constraint, while the other hardwares do not. However, Dr. Mulligan et al.'s work shows quantum algorithms can be made to produce constraint-fitting solutions. So I expect with more work and research, I may be able to design algorithms that accomplish this and produce valid sequences with the D-wave and Intel Loihi devices. The Loihi also has many more capabilities that can't be replicated on a quantum computer, such as asymetric bit connectivity and more advanced information encoding (spike signals, not just binary). With these benefits, I hope neuromorphic computing may outperform classical and quantum, and find solutions other methods couldn't. I hope to design actual novel protein sequences on these devices.
Team members: Robert Strauss
Mentors: Dr. Garret Kenyon (LANL biological computing expert), Kyle Henke (Graduate Student with neuromorphic/quantum knowledge)
Team Members:
Sponsoring Teacher: Nathaniel Morgan