Computers, Artificial Intelligence, and Expert Systems in Biomedical Research

Computer science and molecular biology both developed during the two decades after World War II. Geneticists like Joshua Lederberg were the first to link the two disciplines. Francis Crick and Marshall Nirenberg, for instance, drew on information theory and the principles of computing to decipher the genetic code in the 1950s and 1960s. Lederberg himself not only borrowed concepts but entered the field of computer science itself, staking out a new area of scientific research: the acquisition, systematization, visualization, and dissemination of biomedical knowledge by computer. Computer science, information theory, and biology continue to influence one another today, as neuroscientists use computers to model the human brain, and as computer designers draw on molecular and neurobiology to devise neural networks and molecular computers.

Lederberg's goal in introducing computers into biomedical research was to aid researchers and physicians in problem-solving, decision-making, and diagnostic processes requiring analysis of a large amount of instrument and clinical data. He envisioned expert, or knowledge-based, computer systems that could emulate the inductive reasoning of scientists and doctors, as well as their ability to learn from experience, through artificial intelligence. Under Lederberg's initiative, Stanford University during the 1960s and 1970s moved to the forefront of this undertaking, which was joined by a small number of research institutions and private businesses around the country and in western Europe.

The immediate impetus for Lederberg's research into biomedical applications of computers came from his participation in the National Aeronautics and Space Administration's Mars missions from 1961 onward, for which he designed a computer-controlled mass spectrometer capable of analyzing the Martian surface for signs of life. Lederberg soon applied the theoretical principles of computerized spectrometry to experimentation in the chemical laboratory, where, in 1965, they became the foundation of DENDRAL, a prototype for expert systems and the first use of artificial intelligence in biomedical research.

DENDRAL (for Dendritic Algorithm) was a computer program devised by Lederberg, chairman of the Stanford computer science department Edward A. Feigenbaum, and chemistry professor Carl Djerassi for the elucidation of the molecular structure of unknown organic compounds taken from known groups of such compounds, such as the alkaloids and the steroids. Before the toxicological and pharmacological properties of a compound can be assessed, its molecular structure--the configuration of its atoms--has to be determined. Using the fragmentation pattern of ions produced by bombarding molecules with electrons in a mass spectrometer as entry data, DENDRAL made successive inferences about the type and arrangement of atoms in order to identify the compound from among hundreds or thousands of candidates.

By observing structural constraints within molecules which made certain combinations of atoms implausible, generating and testing hypotheses about the identity of the compound, and ruling out candidates that did not fit within the structural constraints, DENDRAL traced branches of a tree chart that contained all possible configurations of atoms, until it reached the configuration that matched the instrument data most closely. Hence its name, from "dendron," the Greek word for tree. Lederberg himself worked out the basic notational algorithm, called graph theory, to represent three-dimensional molecular structures in a form computers could understand.

In its practical utilization, DENDRAL was designed to relieve chemists of a task that was demanding, repetitive, and time-consuming: surveying a large number of molecular structures, to find those that corresponded to instrument data. Once fully operational, the program performed this task with greater speed than an expert spectrometrist, and with comparable accuracy.

The greatest significance of DENDRAL, however, lay in its theoretical and scientific contribution to the development of knowledge-based computer systems. It was the ambition of DENDRAL's creators to transfer the principles of artificial intelligence from the realm of chess and other strictly controlled settings in which they had been formulated during the 1950s, to real-world problems facing biomedical researchers and physicians. They wanted to show that computers could become experts within a concrete knowledge domain, such as mass spectrometry, where they could solve problems, explain their own conclusions, and interact with human users.

Lederberg and his colleagues believed that artificial intelligence--the use of computers for manipulating symbols, for instance the combination of words in an "if-then" inference, rather than for purely numerical calculation--could assimilate the rules of inductive reasoning and empirical judgment that guide scientists and physicians in their work, rules for which mathematical representations did not exist. Bruce Buchanan and others in Stanford's computer science department distilled these rules, which they called "heuristics," from extended interviews with Lederberg and other experts in their respective fields, and translated them into the formal code of symbolic computation.

DENDRAL ran on a computer system called ACME (Advanced Computer for Medical Research), installed at Stanford Medical School in 1965 for use by resident researchers through time-sharing, with Lederberg as principal investigator. Initially, the system performed real-time, standard numerical analysis of clinical and biomedical research data. DENDRAL was the first artificial intelligence application hosted by ACME. It was succeeded in 1973 by SUMEX-AIM (Stanford University Medical Experimental Computer--Artificial Intelligence in Medicine), a national computer resource for artificial intelligence applications in biomedicine. Users at universities and hospitals across the country were connected to SUMEX via the ARPANET, a predecessor of the Internet developed by the Pentagon in the 1960s.

By 1980, SUMEX hosted nineteen projects, including DENDRAL and its spin-offs, CONGEN and Meta-DENDRAL, programs that generated not just hypotheses for the interpretation of instrument data, but the inductive rules by which hypotheses were constructed. Other SUMEX projects included MYCIN, a program to diagnose and manage medication schedules for infectious diseases, and MOLGEN, a program under Lederberg's own supervision that aided in the planning of laboratory experiments in genetics. Among remote users of the system, researchers at the University of Pittsburgh created INTERNIST, a program that diagnosed multiple internal diseases in the same patient to assist physicians in rural health clinics and other isolated locations without access to advanced diagnostic equipment. Psychiatrists at the UCLA's Neuropsychiatric Institute simulated the thought processes of paranoid patients with a program called PARRY in order to test explanations for the causes of paranoia, and to train psychiatrists in its diagnosis.

DENDRAL and SUMEX helped define the central role computers play in biomedical research today. The projects proved that computers could carry out certain clearly defined functions in the interpretation of laboratory and clinical data. Moreover, the projects gave researchers in different locations the opportunity to experiment with new forms of communication and cooperation, such as data sharing, electronic mail, and bulletin boards. On a theoretical level, creating expert systems forced participants to explore and formalize processes of knowledge acquisition that generally remain unexamined in the course of scientific practice. Not least for this reason, Lederberg and his collaborators derived considerable intellectual satisfaction from this interdisciplinary project.

At the same time, DENDRAL project members were the first to acknowledge that the limits of expert systems were at least as apparent as their potential. Expert systems were custom-made and took years of intense labor to develop, which meant that they could not be supported outside of major research institutions like Stanford, or without generous federal funding. Requiring exact mathematical and symbolic formulation of all operational assumptions, rules, and procedures, the programming requirements of expert systems were so demanding that the systems remained linked to narrow task domains, and even these they often could not fully cover: after more than a decade of elaboration, DENDRAL was still not capable of analyzing all chemical compounds, but only those taken from specific groups of compounds. Programs dedicated to the generation rather than the application of rules (or algorithms), such as CONGEN and Meta-DENDRAL, promised to be of more general use, but for the duration of these projects they also remained tied to a single task domain, mass spectrometry.

Contrary to the speculations of participants and of the popular media, in no case did these expert systems replace the scientist or physician. At best, they could advise and assist them. Yet, Lederberg and the programs' other creators struggled to find acceptance even for this limited role of expert systems among fellow scientists and physicians, many of whom remained skeptical or uninterested. If there was an unambiguous lesson participants learned from DENDRAL and its progeny, it was that human inductive and empirical reasoning is a process of daunting complexity, and is most difficult to model in a machine.