Neural networks predict reactions.

Neural networks predict reactions

When chemists want to synthesize a chemical, they start with commercially available ingredients and then design a sequence of chemical reactions, which they hope will rearrange the starting materials into the desired product. So the more accurately chemists can predict reaction outcomes, the more efficient their syntheses.

A growing branch of artificial intelligence known as neural networks may help chemists make better predictions, according to David W. Elrod, a computational chemist with the Upjohn Co. in Kalamzoo, Mich. He and associates at both Upjohn and Western Michigan Universtiy in Kalamazoo report using a neural network, which they simulate and run on a personal computer, to predict outcomes of simple chemical reactions. The network's accuracy matches that of human chemists and surpasses existing computer expert systems, the researchers say.

As a testbed for the neural network approach to predicting reaction products, the scientists chose a well-studied family of reactions known as electrophilic aromatic substitution. These reactions involve chemically displacing one of six hydrogen atoms, each attached to one of the six interbonded carbon atoms forming the hexagonal benzene ring. When a hydrogen atom gives up its spot on the ring, it leaves behind an electron, an attractive target for an electron-seeking, or electrophilic, atom or chemical group.

Often, one of the benzene's six carbons already hosts a "substituent" that governs further substitutions by biasing where on the ring an electron will be available for bonding. One class of substituents -- called ortho-para directors -- favors further substitutions mostly at the adjacent "ortho" carbon and at the "para" ring-site three carbons away. The other class -- called meta directors -- favors subsequent substitutions two carbons away at the "meta" ring position.

The best reaction-predicting neural network the researchers have tried so far consists of three layers of units interconnected somewhat like neurons. The 25 units of the input layer contain chemical information about the first-ring substituent. Through synapse-like connections that can either strengthen or weakne, these units funnel into, and collectively either excite or inhibit, each of the middle layer's five members. In a similar way, these units then turn on or off the two output units, which respectively represent ortho-para-directing or meta-directing first substituents.

The researchers "train" the network by plugging chemical information on 32 first substituents into the input layer. The network "learns" by changing the connection strengths between the units of three layers until the output predictions coincide closely with experimentally determined results of actual reactions. When tested with 13 ring substituents not included in the training set, the network predicted ortho, meta and para product proportions within 20 percent of actual values in 10 cases. That equals the performance by a small set of human chemists and beats out by three an existing conventional computer expert system for predicting reaction outcomes.

Neural networks can help chemists predict reaction outcomes, the researchers conclude. But Elrod says more elaborate networks will have to be developed for use with complicated reactions.