The hypernetworks are a theoretical model of learning and memory inspired by molecular computation (Zhang et al., DNA10-2004 & DNA12-2006). The entire network is represented as a test tube of DNA molecules, where each DNA molecule encodes a hyperedge of the hypernetwork. The thesis underlying the hypernetwork model is that the self-organizing process simulating the evolution of a high-density soup of massively interacting molecules is a powerful tool for making complex associative memory. This method does not require long DNA sequences, allows for tolerance to errors in chemical reactions, and thus is within the reach of current DNA technology. The learning algorithm evolving the hypernetworks is based on the primitive molecular operations of variation (through combinatorial chemistry), selection (gel electrophoresis), and amplification (polymerase chain reaction). The size of the DNA library is typically of O(1015). Implementation of the hypernetworks in wet DNA computing is in progress.
   As a theoretical model, the evolving molecular hypernetworks offer interesting new perspectives for both molecular computation and evolutionary computation. For example, the molecular evolutionary algorithms are a novel class of evolutionary algorithms which makes use of simple evolutionary operators and a huge population size, opening up new theoretical issues and applications of evolutionary computation. Also, the hyperedges can be interpreted as memory fragments or chunks and the whole network of hyperedges represents an associative memory. This suggests the hypernetworks have a potential to simulate learning and memory mechanisms of a huge, randomized, neuromolecular networks in synapses and cell assemblies (FOCI-2007).

We have recently performed extensive simulations of the molecular hypernetworks to characterize its strengths and potential limitations in practical applications. As a machine learning model, the hypernetwork has several interesting properties. It can capture the higher-order interactions between features, which is useful for knowledge discovery. The hypergraph structure can be converted into association rules and thus comprehensible to domain experts. In biological pattern discovery, for example, the hyperedges allow for biologists to discover building blocks such as network motifs and modules. In addition, a hypernetwork achieves high accuracy and generalization performance since it builds a weighted ensemble structure of a large number of hyperedges.
   Applied to several real-life problems such as biological data mining and text mining, the hypernetworks obtained very competitive results compared to leading-edge machine learning methods, including decision trees, perceptrons, support vector machines, naïve Bayes, and Bayesian networks. The problems we tested include the identification and target prediction of microRNAs, cancer diagnosis based on aptamer-chips, microRNA-based molecular diagnosis, and spam filtering from a stream of incoming e-mail messages.
   Work is going on to see the plausibility of hypernetworks as a model of lifetime personal memory that can eventually deal with a personal database for everything, such as personal blogs or lifelogs like MyLifeBits. As an initial step to this direction of research, we experiment on a large corpus of captions collected from movie DVDs. The interim goal is to learn from the corpus to generate the sentences of a similar style or to mimic the personal experience of language. This system can be used, for example, for predicting user¡¯s cognitive behavior and for delivering personalized Internet services.

The promising results of computer simulations of molecular evolutionary learning of hypernetworks and the difficulty of large-scale simulations lead us to build a customized silicon chip to accelerate the simulation. We designed and fabricated a hypernetwork model in FPGA in collaboration with Inha University. The chip¡¯s performance is being benchmarked on image pattern recognition and cancer diagnosis data. We found the hypernetwork model suitable for hardware implementation in memory chip technology since it ¡°learns¡± by primitive memory access and matching processes, such as ¡°molecular¡± operators of hybridization, selection, and amplification, thus not requiring high-precision numerical calculations. We are interested to know how far this technology can be pushed to solve large-scale problems, especially in biological data mining, pattern classification, multimedia information retrieval, and web data mining.
   Experiments are in progress to implement the hypernetwork models in vitro using DNA computing technology. This research aims to demonstrate that ¡°molecular evolutionary learning¡± is feasible using wet molecular materials and biochemical tools. This might lead to revolution in molecular diagnosis devices since the diagnosis can be directly made in vitro without the help of in silico technology.