Design of an intelligent material
A human engineer strives to create a system to best perform a needed function. For example, an engineer might design a new shingle to minimize energy consumption in a building. The engineer must make decisions about the design in the face of uncertainty. Often the engineer uses a mathematical model to predict the performance of various design choices, knowing that these model predictions will not be perfectly accurate due to a lack of full understanding of the relevant physics. After selecting a preliminary design, the engineer may build a prototype part and measure the actual performance. If the performance does not match the prediction, the model may be refined using the new information gained via the measurements. The new model can then be used to update the design and build a new prototype. Such is the iterative nature of human-directed design.
Human-directed designs do not match the level of complexity present in biological systems. A modern microprocessor is an extremely sophisticated engineering triumph, yet it cannot rival the flexible and robust behavior of the human brain. One characteristic feature of human-directed design is the limited number of options that are considered simultaneously. In the face of imperfect models (as all models are), the engineer desires to intuitively reason out the individual effect of each design parameter. As a result, the most complicated human-designed systems, such as a computer or an airplane, are built up from smaller subsystems, which have already been characterized prior to inclusion in the larger system. This hierarchical design framework reduces the complexity of the design process, and therefore also limits the set of possible designs and the final system performance.
The aim of the proposed research program is to create materials that embody biological design principles, so that they can learn and adapt to optimize themselves. Biological systems are designed through the simultaneous processes of random mutation and selection. Unlike human-directed design, there is no explicit objective that is externally specified. Biological systems that are best able to survive and replicate persist over time, whereas those that are less fit do not. A community of organisms will, over time, become populated with the organisms that are the most efficient at reproducing themselves, in the face of uncertain external conditions and internal competition for limited resources. The community itself can even be viewed as the evolving species. Its members may eventually differentiate themselves to perform specialized functions, such that the community is more fit to reproduce, to such an extent that the individual members can no longer survive apart from the larger community system.
Due to the success of biology in creating complex systems with diverse functionality, biological design principles have been borrowed by human engineers. Two notable examples are genetic algorithms and directed evolution. Both methods are based on mutation and selection, in which the criterion for selection is the desired function of the system. Thus both cases are still human-directed design methods. Directed evolution is applied to an experiment, in which a pool of variants is created via random mutation. For example, a chemical agent can induce random mutations in geneomic DNA. Each new DNA sequence is then characterized, based on its fitness to perform the human-specified function, and a sequence with high fitness is more likely to be included in the next pool. The new pool is copied, mutated again, and re-screened for fitness. At each iteration, the pool is expected to contain, on average, more highly fit sequences.
The concept of a genetic algorithm is the same as directed evolution, but now the cycle of mutation and selection is applied to a mathematical model, not an experiment. Each sequence represents a candidate design, and a model is used to predict the fitness of each sequence. Random “mutations” are made to the designs in the pool, and better (higher fitness) designs are more likely to be carried into the next round. The final design corresponds to the sequence in the final pool having the highest fitness.
The ideas of random mutation and selection used in directed evolution and genetic algorithms can complement more traditional methods for human-directed design. For example, directed evolution can be applied without the need for a mathematical model of the sequence-fitness relationship. Directed evolution and genetic algorithms both enable the search of a wider set of design options than would be practical with deterministic optimization algorithms, which are likely to find a nearby “locally optimal” design (i.e. incremental change to a preliminary design). Deterministic algorithms that guarantee the true “global” optimum also exist, but the computation required is extremely high, limiting their applicability in large design problems.
The design framework we propose here goes beyond this cycle of mutation and selection. Rather, we propose a design framework to create systems that can optimize themselves. We will design into these systems the ability to respond and learn via the mechanisms of mutation and selection. The material could be initially ”trained,” but will continue to learn throughout its lifetime. Traditional human-directed designs yield rigid systems that can only perform the tasks for which they are designed or programmed, and only under conditions assumed during the design phase. Here we aim to create more complex systems that exhibit adaption and robustness, and we describe these systems as “intelligent.” As defined by notions of algorithmic complexity, the intelligence of a system (living or non-living) can be measured by the shortest length of computer code that can reproduce the input-output behavior of the system. Human beings are still arguably more intelligent than any computer that has been built and programmed, so it would be impractical to compute this metric for a person. However, the materials we design and build will be much less complex than a human being, so this metric should be practical here, computed from our system-level models and codes.
In particular, we will design intelligent materials composed of peptide building blocks. Since proteins are the building blocks of life, peptides, or “short proteins,” give us a vast array of possible function, but with a discrete number of design choices. There are twenty natural amino acids, and thus 20n number of different peptide sequences of length n. For example, if n = 7, there are over a billion different peptides for us to choose from. A general challenge in protein design is the lack of understanding and models relating sequence to function. This motivates our focus on the simpler case of peptides, and our close coupling of modeling and experiments throughout the project. We will focus our initial designs around assemblies inspired by the cross-β motif, which is a three-dimensional assembly composed of two-dimensional β-sheets, which are themselves composed of (linear, one-dimensional) peptides. Cross-β assemblies possess the ability to catalyze reactions and execute morphology changes in response to environment cues, which may hint at their role in Alzheimer's and other amyloid diseases. Their structural complexity will be used to build algorithmic complexity into our intelligent materials.
We will optimize our intelligent material such that it can respond and learn in a way that is beneficial to the goal specified by the human engineer. Two essential features of biological systems that must be included in the material design are reversibility and diversity. Without reversibility, a living system can only move in one direction, and thus cannot flexibly respond to a changing environment - once a move is made, it is impossible to move back. Diversity is equally critical in the design of a complex system. Without diversity, a system can reversibly respond to its environment, but responsiveness alone is not intelligence. Intelligence requires the ability to learn, building additional functionality over time and creating greater robustness to an uncertain environment.
We will design our intelligent materials using the framework of dynamic combinatorial libraries (DCL), which have been proposed recently under the umbrella of “systems chemistry.” A DCL is a collection of molecules that can associate in various combinations to create larger molecular assemblies. A DCL is characterized by reversible associations, which enable the overall system to equilibrate, creating a diverse set of species in the library. These libraries have been proposed as a tool for materials design, and also as a model system in which to study origins-of-life chemistry. To date, interesting behaviors have emerged from DCL experiments, but no model predictions, and thus no model-based library designs, have been reported.
In the proposed research program we will design the library of initial peptide sequences such that the material will embody selective pressures that align with the human-specified objective. In particular, the assemblies that are most fit should also possess the highest ability to aggregate limited resources and thus be most successful at growth and reproduction. To ground our work, we will initially focus on the interplay between temperature and light, for adaptation to the environmental conditions, as motivated scientific questions about the emergence of life on the prebiotic Earth. However, our design framework is general, and can enable the design of intelligent materials for applications as diverse as alternative energy and disease therapies.