CHEMICAL CPU’s (Chemical Computational Processing via Ugi Reactions)


For decades, computers have been used to explore chemistry, but can chemistry be used to compute?  Over the past half century, the digital electronics revolution has completely transformed our everyday lives and fueled the whirlwind advances in computing that now enable scientists to routinely model complex biophysical and quantum chemical phenomena using no more than their desktops. Many of these advances are a consequence of Moore’s Law, the empirical observation that the number of transistors on a chip doubles every 18 months, but many indicators suggest that Moore’s Law may soon be coming to its end as transistor sizes approach their fundamental physical limits. While this poses a grave challenge for conventional electronics, it also represents a revolutionary opportunity for chemistry, which naturally operates at such small scales.

Our Brown DARPA Molecular Informatics team is rising to this challenge by developing new  ways of storing information in and computing using small organic molecules. While researchers have long pursued biomolecular computing, small molecules offer a number of key advantages, including their ubiquity, low cost, and multitude of properties that can be designed and manipulated using spectroscopic/spectrometric tools. We have developed new techniques for storing megabytes of information in the unique masses of molecules produced via multicomponent reactions and reading that information with minimal error via high resolution mass spectrometry. We have moreover developed multilayer molecular perceptrons based upon chemical multiply accumulate and thresholding operations that may be implemented and harnessed to recognize images, filter time series data, and sense molecules. Our work paves the way toward using everyday organic molecules for extremely fast, massively parallel classical computation, while also shedding light on how biomolecules perform the computations essential to life.



In general, information may be stored by selecting certain states from a much larger possible space of states. Thus, information may be stored within molecules by synthesizing specific combinations of molecules from a much larger possible set of molecules. Bits can be mapped to the presence or absence of multicomponent reaction products with unique masses in solution and read out via mass spectrometry.

Using state-of-the-art liquid handling robotics to facilitate synthesis and high resolution FT-ICR mass spectrometry, we are able to store and access upwards of megabytes of chemical information. We have moreover developed a full analytical theory of information storage in molecular mixtures, including bounds on information capacity and comparisons with biomolecular techniques.



One of the simplest types of neural networks that undergirds many modern machine learning algorithms is a perceptron, a linear classifier that takes in a set of inputs and outputs whether those inputs correspond to a certain category. We have developed a variety of molecular perceptron implementations, including volumetric (see Figure 1) and fully chemical autocatalytic versions, that recognize images and filter time series data represented in molecular form. We are also designing molecular perceptrons that can be used as in situ chemical sensors that learn when to trigger biological or inorganic catalysts that perform desired functions. As many biochemical reaction networks may be viewed as being able to learn, the development of entirely synthetic “intelligent” reactions will shed light on how biology makes decisions.



  1. Ray, A., et al. Computing with Chemicals: Perceptrons Using Mixtures of Small Molecules. IEEE International Symposium on Information Theory (2018).
  2. Arcadia, C., et al. Parallelized Linear Classification with Volumetric Chemical Perceptrons. IEEE International Conference on Rebooting Computing (2018).
  3. Rosenstein, J. K., et al. Theory of Information Storage in Molecular Mixtures. Submitted (2018).

Figure 1 (Molecular Perceptron)

A schematic of a molecular perceptron. M MNIST images are stored in parallel within a single 96 well plate using M different phenols. At the same time, perceptron weights are trained on hundreds of images in silico. All wells with positive weights are transferred to a positive pool and all with negative weights are transferred to a negative pool. The volumes of the analytes transferred are proportional to their weights. Thresholding is performed by measuring the analyte concentrations in each of the wells and comparing. If the final concentration is positive, the image has been matched.