DENDRAL

What DENDRAL was

DENDRAL was an artificial intelligence project at Stanford University that began in 1964^[1]. Its main purpose was not to build a product but to study how hypotheses are formed and how discovery happens in science^[1]. The concrete task it took on was chemical: helping organic chemists work out the structure of unknown organic molecules by analysing mass spectra together with knowledge of chemistry^[1]. The Computer History Museum describes the same aim more broadly — applying the accumulated expertise of specialists to problem solving in the area of chemistry and physics^[2].

Origins

The project was carried out by Edward Feigenbaum, Bruce G. Buchanan, Joshua Lederberg and Carl Djerassi, with a team of research associates and students^[1]. The Computer History Museum's timeline names a Stanford team led by professors Ed Feigenbaum, Joshua Lederberg and Carl Djerassi^[2]. A Stanford document on expert systems dates the beginning of expert-system development at Stanford to 1965, under the direction of Joshua Lederberg, Edward A. Feigenbaum and Bruce G. Buchanan, and says the first system was called Dendral^[3].

That Stanford document also records why the project took the form it did: the goal was to aid organic chemists in determining the molecular structure of compounds, and the expert approach grew out of the failure of conventional computer methods^[3].

How it worked

DENDRAL consisted of two main programs, Heuristic Dendral and Meta-Dendral, plus several sub-programs, and it was written in Lisp^[1].

A mass spectrum, in general terms, is a record of how a molecule comes apart: the compound is broken into fragments that are sorted by mass, and the resulting pattern acts as a kind of fingerprint. Reading a structure back out of that pattern is inference rather than measurement, and that is the step DENDRAL automated. Heuristic Dendral took mass spectra or other experimental data together with a knowledge base of chemistry and produced a set of possible chemical structures^[1]. It did not search blindly over every conceivable molecule; chemical constraints were used to narrow the candidates. The Computer History Museum describes the mechanism as applying a battery of if-then rules to identify the molecular structure of organic compounds^[2].

Meta-Dendral played a different role. It was a machine learning system that proposed rules of mass spectrometry, correlating structural features with the processes that produce a spectrum^[1]. The Stanford document describes it as a program derived from Dendral that discovers new rules about the behaviour of fragmented molecules^[3]. The two halves of the project therefore ran in opposite directions: one applied existing chemical rules to data, the other drew new rules out of data.

Results and standing

By 1970 DENDRAL was performing structural interpretation at post-doc level^[1].

Wikipedia calls DENDRAL the first expert system, on the grounds that it automated the decision-making process and problem-solving behaviour of organic chemists^[1]. The Computer History Museum's timeline likewise calls it the first expert system^[2].

Influence

Many systems were derived from DENDRAL, including MYCIN, MOLGEN, PROSPECTOR, XCON and STEAMER^[1]. The pattern it established — a body of specialist knowledge encoded as rules and applied to a narrow problem, with the program explaining its conclusions rather than simply computing a number — is the pattern those later systems followed.