A Robot Mapped 960 Versions of a 135-Year-Old Reaction and Found a Hidden Branch

In one corner of a four-dimensional chemical map, at high acid, high aldehyde and low urea concentrations, a reaction first published in 1891 did something it was not supposed to do.

The unexpected product was not abundant. Its yield topped out at 9 percent, which is the sort of result an optimization campaign might treat as an inconvenience on the way to something better.

This time, the inconvenience was the discovery.

Researchers at the Institute for Basic Science in South Korea used a robotic platform to run 960 versions of the Biginelli reaction, a classic method for combining three types of ingredients into a ring-shaped molecule. They were not trying to find the most efficient recipe for the familiar product. They wanted to know what else the reaction could make.

The answer, reported in Nature Synthesis, included a previously unrecognized route to an unusually elaborate bicyclic molecule. Mechanism-oriented software helped the team work out how that route might operate. Once the chemists understood it, they redesigned the synthesis, improved the yield and used the chemistry to make a family of related structures.

The experiment offers a different vision of laboratory automation. A robot does not have to be merely a tireless assistant that finds the best answer to a question humans already chose. It can make a much larger question practical: What is hiding in all the parts of the reaction map that nobody normally checks?

One reaction, 960 starting points

Pietro Biginelli first reported the reaction that bears his name in 1891. In its familiar form, an aldehyde, a beta-ketoester and urea combine under acidic conditions to produce a dihydropyrimidinone, a six-membered ring containing nitrogen.

Writing that transformation as a tidy equation creates an appealing picture. Three ingredients go in. One main product comes out.

Real mixtures are less obedient. The ingredients can react through multiple sequences, producing intermediates and side products whose prevalence changes with concentration, temperature and catalyst loading. What chemists call a reaction can therefore be a network of competing reactions.

Exploring that network point by point is brutally repetitive. The researchers used an in-house robotic platform called Robowski to do the repetition.

The robot varied four concentrations of p-methoxybenzaldehyde, four of ethyl acetoacetate, six of urea and ten of the acid catalyst. Multiply those choices and the result is a grid of 960 conditions. Every mixture was heated to 100 degrees Celsius for 24 hours.

The important difference was what happened next. Many automated chemistry systems measure how much of a desired product each experiment produced. Robowski recorded an ultraviolet-visible spectrum of every crude mixture, creating a fingerprint that also contained evidence of unexpected components.

The researchers pooled the mixtures, separated their contents by high-performance liquid chromatography and identified purified components with nuclear magnetic resonance spectroscopy. They then recorded a reference spectrum for each component and asked whether combinations of those references could reproduce all 960 fingerprints.

When the fit was poor, something was still missing.

After eight cycles of separation, identification and spectral analysis, the team had a set of 12 components that satisfactorily represented the mapped mixtures. One was compound 8a, the strange product whose measured yield reached 9 percent in the original reaction space.

The molecule that did not fit the old story

Nuclear magnetic resonance and high-resolution mass spectrometry were not enough to settle the structure of 8a. X-ray diffraction revealed a bicyclic core decorated with four substituents.

This was not simply a slightly modified version of the standard Biginelli product. It appeared to contain contributions from seven starting molecules: two molecules of ethyl acetoacetate, four of p-methoxybenzaldehyde and one of urea.

Chemists call that a pseudo-seven-component transformation. The name can sound as if seven different chemicals were thrown together, but the 'seven' counts individual molecules drawn from three substrate types.

A similar, less substituted bicyclic core had appeared in a 1970 paper, where it required a multistep synthesis. What was new here was the hidden branch of the Biginelli network and the route by which a complex substituted product emerged from the familiar ingredients.

The researchers could see the final structure and had isolated several suspected intermediates. They still needed to explain how the reaction assembled it.

For that, they used MECH, a chemical AI system built around more than 10,000 expert-encoded mechanistic transformations. These are computerized versions of the bond-making and bond-breaking moves chemists use when drawing reaction mechanisms.

MECH was not a chatbot guessing a plausible story from chemical text. It constructed networks of possible steps from the starting materials, then searched for routes compatible with the conditions and observed products.

The proposed sequence begins with the acid helping ethyl acetoacetate lose parts of its original structure. A series of condensations produces dibenzylideneacetone. That compound reacts with urea and closes into a ring. The resulting electron-rich molecule can then react with another dibenzylideneacetone through a Diels-Alder cycloaddition, building the second ring of product 8a.

That mechanism remains a proposal, not a direct observation of every molecular step. But it produced testable consequences. The team carried out targeted reactions to support key connections in the network, including a one-pot reaction between the proposed dienone intermediate and urea.

Understanding the network also suggested a more practical route. In the original mixture, several pathways competed for the same ingredients. The researchers therefore made the dienone first, then reacted it with urea. The streamlined version sacrificed the conceptual neatness of the seven-molecule process, but it raised the yield of 8a from 9 percent to 28 percent.

The team went on to synthesize nine analogues with different aromatic groups. A related extension, inspired by the initial discovery and based partly on Hantzsch reaction chemistry, produced doubly and triply annulated pyridines in yields of up to 29 percent over two steps. Those tricyclic products were not found directly in the original 960 experiments, but they show how an obscure side route can become a starting point for deliberate synthesis.

Useful molecules, or merely interesting ones?

Structural complexity alone does not make a molecule useful. The researchers tested several of the products for behavior that might justify further attention.

Some aggregated into larger assemblies in ways that changed with concentration and temperature. Several bound metal ions. One compound also changed how its mirror-image forms paired depending on which metal was present: zinc favored associations between molecules of the same handedness, while barium favored opposite handedness.

That metal-controlled chiral sorting is unusual, but its practical value remains unknown. The authors suggest possible directions including molecular sensing and responsive materials. They did not build a sensor, demonstrate a medicine or show that the compounds are ready for industrial use.

The more immediate result is methodological. The platform turned a familiar reaction into a landscape that could be surveyed rather than a recipe that could only be optimized.

It has limits. The analysis relies on products with distinguishable ultraviolet-visible absorption. Purely aliphatic products, for example, may produce no useful signal above the system's measurement threshold. A satisfactory spectral fit also does not guarantee that every trace compound has been found.

Still, the study exposes a bias built into ordinary chemical optimization. If the goal is to maximize one known product, a 9 percent stranger is a problem. If the goal is to understand the network, that same stranger may be the most valuable thing in the flask.

The robot did not replace the chemists. Automation created the map. Analytical instruments exposed the anomaly. Humans isolated and identified it. Mechanism-oriented software proposed a path, and new experiments tested that proposal.

Together, those pieces recovered a branch of chemistry that had remained unnoticed for 135 years. The intriguing question now is how many other famous reactions look complete only because nobody has yet mapped where else they can go.