1 Introduction

The general goal of the research proposed here is to develop a
computer program (termed MOLGEN) to assist a motecular geneticist in
planning laboratory experiments. It is to be an interactive systen,
drawing on both the expertise of the hunan geneticist and the expert
knowledge stored in the data base. The systen nust be convenient and
confortable for the geneticist, and will need to be sophisticated and
powerful from the point of view of computer science. The program will
itself be an experiment in the development of intelligent systens.

The need for such a. program is suggested by the recent growth
of technology in molecular genetics. In the past three years many new
site-specific restriction enzymes have been discovered (Nathans 1975).
These enzymes have been applied to physical mapping of chronosomes,
nucleotide sequence analysis of DMNA, isolation of genes, and
restructuring of DNA molecules. Such rapid growth extends throughout
the field, with continuing advances in separative and physical
techniques. With the many new procedures, there are possibilities for
novel combinations of techniques which broaden the horizon of possible
experiments. At the same time, it has become increasingly difficult
for any one scientist to keep track of what is available, as well as
the limitations and use of the newer techniques. Many "ingenious"
discoveries, in fact, can be viewed as a judiciously selected
combination of well established unit processes. Hence an intelligent

computer system is likely to be pragmatically useful in this area.
An immediate apvlication of the proposed svsten will be the
simulation of reconbination of randon segments of bacterial DNA cut by
restriction endonucleases. In experiments currently underway in
Professor Lederberg’s laboratory, geneticists would like to have
quantitative predictions of reaction component concentrations at any
particular time, given initial concentration of Bacillus subtilis DNA
and site-specific restriction enzymes. The conception of the eueyne
simulation program (see part III) in the MOLGEN systen was directed
precisely toward this kind of need. Part of this effort will involve
the creation of a knowledge base of restriction enzycves with associated
kinetic data, recognition site information, and characterization of
availability and degree of purity. Convenient access to a
conprehensive source of this information would itself be of
considerable value.

A subproject which should also prove very useful to the
geneticist will be the sunnarization of all enpirically established
sequences of ‘DNA and RNA, along with restriction enzyne site
specificity. Such information might play an important role in the
current effort of Professor Lederberg’s laboratory in dissecting the
genetic structure of Bacillus subtilis. Recognition sites in a wide
variety of naturally occuring nucleic acid species has been sunnarized
in a preliminary way by Sobell (Sobell 1973).

An important factor in selecting this domain was the

availability of a great many tangible problems covering a wide range of
difficulty. This means that it will be possible to develop the svysten
incrementally, and still be able to perform useful tasks even at the
outset.

That branch of computer science known as artificial
intelligence (AI) has already developed techniques for conputer-
assisted problem solving in the experimental sciences. Exanples of
successful projects are: the DENDRAL project at Stanford in the field
of mass spectrometry; the organic synthesis systems developed by Vipke,
Corey, and others; the MYCIN project at Stanford in the field of
chemotherapeutic advice; and Pople and Myer’s DIALOG system for nedical
diagnosis (Pople 1975). The joint effort of computer scientists and
experts in the experimental sciences has allowed the creation of
excellent computer systems. These systems are presently being used to
facilitate research in the experimental sciences. The ideas and
technology behind the projects are now available for application to an
experimental planning program for molecular genetics.

There appear to be two factors common to those task donains to
which computer-assisted problem solving techniques have been applied
successfully:

1) the development of systematic and logical methods of
analysis of problems and application of tools by the
experimental scientists themselves,

2) the possibility of solving problens in the donain
using a strongly constrained and consistent body of

knowledge.
In the current state of the art, heuristic problen solving prograns
containing a few thousand facts and heuristics about a task domain can
be built and used effectively. The existence of these two factors in a
domain suggests that the computer scientist can build an intellectual
bridge into the domain in a period of time that might be measured in
months rather than years.

Some problems of experimental molecular genetics currently
share both of these factors. For example, the growth of knowledge about
restriction enzymes has been important to genetics because of their
many applications to manipulation and separation of DNA structures.
Despite the large number of enzymes, each can be represented by a small
set of descriptors. This makes the task of representing the knowledge
far more limited than it might first appear.

One major issue of artificial intelligence that relates to this
research is how to plan experiments involving the choice of several
steps in a large space of possibilities. This includes such
fundamental points as when to design an entire experimental plan first
and then fill in details later (top-down), and when to design a
conplete step at a time (bottom-up). A second major issue will be the
development and management of large knowledge bases. The MOLGEN system
will need extensive files of knowledge about DNA structures, enzymes,
separative techniques, and physical processes. For exanple some of
this knowledge will involve mathematical models of kinetics; some will

be detailed sequences of base pairs; some will be "fuzzy" and inexact
descriptions of a particular process or structure. The knowledge bases
must be designed to allow as much of the data as possible to be handled
in a uniform manner. A third major computer science problem will be
the creation of a simulation program which can simulate the action of
enzymes on DNA structures. Many of the current techniques developed
for simulation in other domains may prove useful here. The NOLGEN
system will also have to deal with cases of incomplete and inexact
knowledge in the planning and the simulation stages. Molecular
genetics, as a task domain, offers problems which are limited and
structured enough to be feasible with available techniques, as well as
those which are challenging enough to advance the art.

The overall purpose, then, of providing computer assistance in
the planning process is to make it possible for an investigator to
explore a wide range of possible experiments. Calculations of the
effectiveness of computer assisted planning can only be approximate,
but the following is offered for a rough estimate of the potential.
Professor Lederberg has suggested that the knowledge actively used in
the planning of experiments in molecular genetics could be encoded as
approximately a thousand computer "rules". (This fits confortably in
the range used by current knowledge-based intelligent systems.) It has
also been observed that most experiments in molecular genetics involve
on the order of seven levels of formal decision once the context of a
particular goal has been set. We believe that fairly crude heuristics

can reduce the likely choices to perhaps ten. Ten major alternatives at
each level of planning yields a total "search space” of about ten
million alternatives. A search space of this size fits within the
limits of current AI programs but still demands a definition of more
powerful search-reduction heuristics for high performance. The purpose
of these estimates is not so much to give a true measure of program

performance, as to illustrate the notion that the nunber of

possible
experiments is indeed much greater than could actually be performed.
Prof. Lederberg has suggested that a nunber of important and perhaps
novel experiments are not being considered because of the difficulties

in keeping current in a rapidly expanding discipline. Meeting this

need is precisely the goal of the MOLGEN project.
2 Background and Related Work

2.1 Relationship to Other Research in Artificial
Intelligence
There are two primary reasons for proposing the MOLGEN project:
(1) to develop a high performance conputer progran
augmenting human intelligence in important research
areas in molecular genetics,
(2) to provide a context for testing and extending ideas in
the design of intelligent systens.
This section will discuss the existing scientific and technological
base and some proposed extensions and show the relationships between

MOLGEN and other AI programs.

2.1.1 Planning and Heuristic Search

The problem of planning an experiment fits into the paradiga of
heuristic search, a well established framework for conputer based
protiem solving. Problems in this framework can be represented as a
tree of subproblems, with solutions existing at unknown depths along
unknown paths. Judgmental rules and procedures, called heuristics, are
applied to direct the search. When we speak of the expertise or good
judgment of an expert, we are often referring to the heuristics he has
developed to search effectively.

There are typically a number of alternative procedures to

choose fron at each step in the course of an experiment in molecular
genetics. These alternate steps can be viewed as a family of

possible
experiments. This family may then be represented as a tree structure
so that the planning of a particular experinent is equivalent to
choosing a path in the tree. Each node represents a certain “state” in
the experiment where the physical entities being used by the
experimenter are in a particular condition and certain quantities are
known. For example, the physical entities may be a mixture of
oligonucleotides whose terminal base sequences are known. The activity
in the planning process centers around selecting an operation to carry
the experiment to the next state. These operations are chosen fron the
program’s knowledge base and are naturally expressed as transformations
from one state to another. For exanple, in the the context of
molecular genetics, a chronatography step could be represented as a
transformation which "transforms" a mixture (initial state) iato its
separated conponent parts (final state). Each particular type of
chromatography would be more specifically described by the parameters
of the physical entities which are active in the separation, For
example, gel electrophoresis would be described in terms of its ability
to separate oligonucleotides according to length. Part of the
knowledge embedded in the transformation would be a rule for estimating
the effectiveness of gel electrophoresis on any particular mixture of
oligonucleotides for the various gels which are available. Finally, a
completed plan is a sequence of transformations which when applied

successively to the initial state of the experinent transform it to the

desired goal or final state.
This model of problem solving based on the selection of a path
according to heuristic knowledge is often called the heuristic search
paradigm. It has been used as a model for a variety of problems since
such early projects in AI as the "Logic Theory “fachine"™ (1956) by
Newell and Simon. One dimension of progress in AI has been the

£

development of techniques for the creation of plans within the

heuristic search paradigm.

2.1.2 Strategies for Planning

Several algorithms have been developed to assist in choosing a "nininun
cost" path in a heuristic search tree (Nilsson 1971), where an
estimating function is available which can measure how "close" any
intermediate point in the experiment is to a final state. In order to
guarantee that the algorithms will find a ninimal path, the estimating
function must never overestimate the distance to the goal. For conplex
problems, a practical difficulty continues to appear in many contexts.
Simply stated, it is sonetimes best to retreat fron a goal in order to
get closer to it. In mathematical theorems, this arises in those cases
where it is easier to solve a more general theorem than a specific one.
In organic synthesis, it is sometimes better to build up a rather
complicated structure which seems "farther" from a target compound than
some current step in the synthesis, but from which a direct and often
elegant reaction will transform the complicated structure almost
directly to the desired product. |

In complex scientific domains, the distance estimating
functions are almost impossible to find, Even if such a metric can be
found, it is unlikely we can guarantee that! it never “overestimates” the
distance to a goal state. Nevertheless, in special cases, quantitative
distance measures can sometimes be found when discrinination among the
products of the alternative transformations is statistically testable.
Otherwise, planning programs must rely on domain specific or even goal
specific knowledge to guide the selection of the next step.

Some more general techniques for planning have been developed
for robot problem solvers. For example, the STRIPS systen (Fikes
1972), in the course of building long plans, attempts to assemble short
sequences of operators which may later prove useful. The idea is that
it is more efficient to assess the effect of the smaller plans and
reuse bike iuan Lo generate them again from scratcn tater. ine
aiyoritnms inctude metnodas ror disentangiing tne Local errectS or Ctnat
piece or a pian trom its immediate context and predicting its etrtrect in
a naw context.

Another planning idea of general application has been developed
in Sacerdoti’s work on hierarchical planning (Sacerdoti 1973). This
can be seen as a_ formalization of the common sense notion of planning
major steps first, and attending to details later. In our terms, it
means recognizing that some preconditions for use of a transformation
are more important than others. This ordering allows the nost
important criteria to be considered first. In contrast to earlier

methods for generating plans, which proceeded a step at a time toward a

10
goal, hierarchical planning involves sketching out a plan in several
passes of refinement of detail. This approach is applicable where it is
Poessibie tu andicatce an uruering ror tne criteria used in choosing tne
next step of a plan.

Our examination of several experiments in molecular genetics
suggests that this idea has a great deal of relevance. For example, in
dealing with enzyme data, it is far more important for most purposes
that an enzyme be applicable, i.e. that it have the desired effect on a
substrate, than that its most active pH be in a certain range. In
terms of planning, this suggests that "“applicability" is a more
important piece of planning information than is "pH range". Hence, the
adjustment of pH in an experiment is a detail which can be considered
during a later pass in the planning process. This ordering of
importance need not be assumed constant for all experiments, but may be
considered as part of the planning heuristics. The application and
development of these planning issues will be part of the interesting
computer science questions to be investigated in the MOLGEN project.

The problem of experiment planning can be compared to the
automatic generation of computer programs. The planner’s task is to
find a sequence of instructions that will transform the input into the
desired output. However, the tools that. the experiment planning
program will work with are conceptually simpler and the range of
possible "programs" or experiments is more restricted than those

utilized in the general problem. Although there are many unit

11
processes, the planning will be facilitated by needing little or no
recursion or looping, limited "data types", few steps, and in addition
will have a detailed knowledge of the subject matter. Thus the

heuristics necessary to plan experiments are more manageable.

2.1.3 Acquisition of Knowledge

One trend in the development of intelligent systems has been an
increase in the size of the knowledge base used by _ the prograns.
Domain specific knowledge is used to guide and hence speed their
search. The process by which a system acquires this knowledge is one
of the central design issues for an intelligent systen. The DENDRAL
programs, for instance, which are used in the analysis of mass spectra
of organic compounds (Buchanan 1969), acquired their chemical and nass
spectral knowledge by a painstaking process in which an AI scientist
worked with a chemist, eliciting from the chemist the theories, facts,
rules, and heuristics applicable to reasoning from mass spectra. This
same approach has been used successfully in protein structure
determination using x-ray crystallography and in diagnosis of glaucoma
eye disease (Kulikowski 1973).

More recent versions of the DENDRAL program have special built-
in editor programs for acquiring knowledge and data from the chemical
domain. For exanple, a chemical structure editor is used ‘in entering
Fragments of Chemical scetuccures, aud other editors speciairze in
acquiring chemical structure constraints and fragmentation rules.

Another approach, typified by the SECS progran for organic chenical
synthesis (Wipke 1975), involved use of a special Language (ALCHEM) for
describing organic reactions. For each of the areas of organic
synthesis expertise, a chemist specialist was responsible for encoding
in ALCHEN the particulars about the useful reactions in his specialty.
The idea was that a user of the SECS program could then draw upon the
combined knowledge from all the experts who had participated. This
process not only contributed to the growth of the knowledge base, but
also resulted in changes and additions to the ALCHEM language which
made it both more powerful and more convenient.

The MYCIN program, used for diagnosis and treatment planning of
infectious disease using antibiotics (Shortliffe 1973, 1975) uses a
more elaborate method for the acquisition of knowledge. In this systen,
knowledge is embedded in a collection of some 309 decision rules. New
rules can be added via a rule acquisition progran. This program
accepts rules in a stylized form of English, converts then to an
internal form, and then prints them out again in English to indicate
its understanding of the rule. The system also has a model of its
knowledge, in the form of abstracted descriptions of classes of rules.
It uses these both to help interpret the rule the physician has typed
in, and as acheck on its content. If, for exanple, the new rule
differs from most others in that class, the physician is warned, and
can then alter the rule if he wishes.

One of the more interesting components of the MYCIN program is

its explanation system. This concept was introduced earlier in

13
Vinograd’s SHRDLU progran (Winograd 1972), and refers to the ability of
a program to explain its activities to its user. For example, NHYCIN
may ask the physician a particular question about the results of a
culture. If the physician wants to know why the question was asked, he
can ask the system. It will then examine its reasoning, and indicate
the purpose of the question by displaying the relevant rule. The same
rules which guide the consultation are thus being used to explain the

program’s behavior.

2.1.4 Applications to MOLGEN

It is our goal in the development of the MOLGEN system to draw
upon and extend some of the ideas and techniques mentioned above. With
regard to planning, we will be experimenting with ways to relate
hierarchical notions of planning to the concepts of molecular genetics.
Furthermore, we will investigate techniques to make the NOLGEN programs
handle the diversity of goals in molecular genetics experiments, Note
that when a physician starts to use MYCIN, the program does not have to
ask him what he wants to do. The physician always wants to diagnose a
disease and prescribe treatment. For MOLGEN, however, the experimental
goals are as diverse as the experiments in molecular genetics.

We expect MOLGEN to draw upon the same knowledge base in its
efforts to work on a variety of experimental planning problems . In
spite of the diversity of experiments in molecular genetics, we have
categorized three main types of experiments as follows:

l. Given a structural hypothesis, plan experinents and

14
procedures which would verify the hypothesis. (This is
an example of structural analysis.)

2. Given starting and target DNA structures, plan
experiments and procedures which will transform the
initial structure to the final structure. (This is an
example of structural synthesis.)

3. Given an initial DNA structure, hypothesized final
structure, and a set of steps in an experiment to be
Carried out, Verify tite prouvabie currectness of ctne
final structure or show any other possible final
structures. (This is an example of a “proof” checker.)

An initial goal towards meeting the first two objectives is the
development of a planning program which will allow a user to “sketch”
out an experiment, and allow MOLGEN to fill in the details . An example
showing how this might proceed is given in section 3 of this proposal.

A second area for developing on the themes mentioned above is
in the atea vt Kuowleage acquisition and representation. ine nain
extension to previous ideas has to do with the handling of a diversity
of data types. MOLGEN must deal with a broad range of types of
knowledge --~ including DNA structures, enzymes, chromatographic
techniques and various physical processes. For each of these types of
data, MOLGEN is expected to assist the user in entering and modifying
the data and knowledge base. This should be done in a form convenient

to the user’s view of each particular type and yet at the same time in

15
a way which keeps” the program’s diverse collection of knowledge and
data types fron becoming too complex to be manageable. Similarly, the
explanation system must be implemented in a way that makes the
reasoning of the program accessible to the user, through all levels of
the knowledge base. A successful treatment for a broad spectrun of

diverse knowledge will be a major design goal for the MNOLGEN project.

2.2 Related Work in Computers Applied to Genetic
Biochemistry

We propose to provide in this section a summary limited to work
directly related to the simulation and planning aspects of the MOLGEN
project.

Some of the most advanced prior work in simulation has been
done in the area of enzyme kinetics. The basic approach has been to
set up the differential equations describing the chemical reactions ,
and then solve those equations on either a digital (Chance and Shephard
1969, Bates 1973, and Garfinkel 1968) or analog (Chance 1967) conputer
system. It is interesting that attempts have been made to facilitate
use of the systems by working geneticists, either by means of
interactive display mechanisms (Bates 1973), or an English-like
language for describing the chemical systems (Garfinkel 1968). We
intend to expand upon these user-oriented techniques during our system
development.

We also know of a system (Green 1970) which uses matrix

16
representations of the steady-state equations describing enzymes to
permit rapid calculation of the distribution of all enzymatic forms at
any given moment. We feel that the accurate modelling of this steady-
state behavior will prove vital to our enzymatic simulation systen,
although we still have to investigate the general applicability of the
matrix representation technique for multi-enzyme systems.

Crothers (1968, 1971) has developed several computer programs
to simulate various aspects of the process of de/renaturation. The
detailed mathematical models available on conformational changes
(Bloomfield 1974) will facilitate the computer simulation of these
processes,

The problem of representing the often “fuzzy" knowledge about
DNA structures has been considered in a model for the heterogeneity of
base composition (Elton 1974). The point is that the exact conposition
is rarely known for more than a few hundred bases. Elton’s model
classifies the DNA structures into "segments" with different underlying
base compositions, each segment categorized by an estimate of Length
and G-C content. Our representation model for DNA structures will
allow such classification of portions of DNA, as well as many other
factors like the distribution of important features (nicks, gaps, etc.)
along parts of the DNA strands.

Finally, we find of direct relevance to our design of
experimental plans the use by Khorana’s group (Powers 1975) of conputer

programs to help plan the detailed steps involved in gene syntheses.

17
The genes are synthesized in small segments and then joined by some
chemical or enzymatic means. The program, using knowledge of the
chemical environment, determines optimal segment lengths and order of
joining. This will be one of many experimental paradigms we hope the

MOLGEN system will encompass.

2.3 Supporting Research at the Stanford Heuristic
Programming Project

The Heuristic Programming Project at Stanford, of which the
MOLGEN research will be a part, is one of the few AI research groups
that emphasizes development of intelligent systems for use in the
sciences. This section discusses some of these intelligent systems
which are being developed and actively used. These interdisciplinary
projects serve to illustrate the range of experience that the MOLGEN

research can draw on.

26501 The DENDRAL Project

The Heuristic DENDRAL project was started by Professors
Lederberg and Feigenbaum in 1965, and has diversified and natured
considerably since that time. The DENDRAL program (Buchanan 1975 and
Appendix A) is an application of artificial intelligence to bionedical
molecular structure determination problems. The program is designed to
explain enpirical data, in particular, to interpret experimental data
from amass spectrometer, an analytical instrument used in organic

chemistry. It uses heuristic search techniques to explore plausible

18
candidate explanations of the empirical data in much the same way that
other artificial intelligence programs explore spaces of hypotheses.

A major contribution has been demonstrating that a large amount
of scientific knowledge can be managed and used by a conplex problen-
solving program. Another contribution has been showing that ATL
techniques can be successfully applied to scientific problems.

The DENDRAL molecular structure building program, CONGEN, is
currently used by chemists as a tool for building molecular structures

froa stcvucturai pieces wnicen wne cnemist uas anfercea Crom wnatever

ur

datd ate avaliabie. Tne DENDRAL FuaAnnen (Smatn L943) uses a Large
amount of knowledge specific to mass spectrometry to infer structural
constraints directly from empirical data. The DENDRAL PREDICTOR is a
simulation model of the mass spectrometer that allows the program to
make testable predictions of candidate hypotheses. META-DENDRAL is a
program for discovering new inference rules in the field of mass
spectrometry.

From the start, major emphasis has been placed on avoiding the
"outsider-insider" distinction by having trained chemists work on the
programs and make them useful to working scientists. This emphasis is

"problems conducted without

in contrast to other research on "toy
reference to the needs of working scientists. Since May 1974, the
project has moved to the interactive computing environment of the NIH-

funded SUMEX-AIM facility (Carhart 1975). Because of this, many

scientists outside this university have been able to use the DENDRAL

19
conputer programs to further their own research. These programs are
currently receiving heavy use from local users and outside users who
are investigating mass spectrometry problems for a variety of different

compound classes.

2.3.2 The MYCIN Project

The MYCIN project (Davis 1975 ) at Stanford is another
interdisciplinary research group including members of the Heuristic
Programming Project, whose goal is to apply artificial intelligence
reasoning programs to a scientific problem. The objective of this
MYCIN research is to develop ‘a computer-based system capable of
using the judgmental knowledge of experts to assist physicians in
selecting appropriate antimicrobial therapy for patients with
infectious diseases. The work concentrates initially on the use of
antimicrobial agents in the treatment of bacteremias.

The current MYCIN program utilizes data available from the
microbiology and clinical chemistry laboratories, plus the physician’s
response to computer-generated questions, to provide physicians with
consultative advice on antimicrobial therupy. Therapy selection takes
into account both the patient’s infections and the range of possible
identities of the organisms causing these infections. One contribution
of the MYCIN system has been developing general techniques’ for
explaining its reasoning steps to permit the program to explain all of
its actions and reasoning, including, for example, the deduction of the

identities of pathologic agents. Another important part of the MYCIN

20
systen is a rule-acquisition program for computer acquisition of
judgmental knowledge about those concepts which the program uses in
making deductions. This permits experts in the field of infectious
disease therapy to teach the system those therapeutic decision rules
which they use in their clinical practice. Making rules, exploring
MYCIN’s reasoning, or adding new knowledge to the system is done in a

stylized subset of English.

2.3.3 Protein Structure Modeling Project

The Protein crystallography project involves scientists at the
University of California at San Diego as well as members of the
Heuristic Programming Project. The general objective of the project is
to apply problem solving techniques in artificial intelligence to the
"phase problen" of x-ray crystallography, in order to determine the
three~dimensional structures of proteins. Specifically, the task is to
develop a computational system which can infer the tertiary structure
of a protein molecule in the absence of phase information normally
obtained fron nultiple isomorphous replacement procedures.

In the context of artificial intelligence, project objectives
center around knowledge acquisition and program organization. The
knowledge acquisition task involves formalizing the knowledge and
heuristics used by expert protein crystallographers to infer the
tertiary structure of proteins from x-ray crystallographic data, and to
formalize this expertise in appropriate data structures and heuristic

procedures. The program organization plans are variations on the

21
HEARSAY (Reddy 1973) model of cooperating "expert" modules in

program.

22
3 Detailed Plans for Work During the Proposal Period

3.1 Introduction

This section presents our overall view for the molecular
penetics system as well as a detailed list of specific goals for the
first year covered by this proposal. The systen will be conposed of
three major, interacting parts: an experiment planning systen, an
enzymatic action simulation program, anda collection of knowledge
bases containing the rules and heuristics of molecular genetics.

The experiment planning program will collect information about
a problem from the user, select an appropriate methodology (information
retrieval, simulation, hierarchical planning, or some conbination) and
then work interactively with him to solve the problen. Some questions
can be answered by information retrieval from a knowledge base, such as
a question about the enzymes” that function under certain pH or salt
concentrations. On the other hand, if a geneticist asked for a
prediction of the ratio of linear to circular DNA in a test tube after
10 minutes application of ligase, then a straightforward simulation
would be in order, In verifying that a given set of steps in an
experiment produces the final result, a combination of sinulation and
retrieval would be used by the planning progran. In more difficult,
higher order problems, such as structural analysis or synthesis,

hierarchical planning is needed, perhaps involving both retrieval and

simulation as well.

23
A simulation program will provide detailed nodelling of the
action of enzymes on DNA structures. Our goal is a program which will
be able to produce accurate enough results to answer questions like the
one posed above, i.e. what will happen after application of ligase for
10 minutes, 2 hours, etc.

The knowledge bases will be composed of collections of the
rules and heuristics used by geneticists, as well as facts about
enzymes, experimental methods, and physical processes like
renaturation. They must allow access in retrieval, simulation, and
planning modes, so provision for the representation of many types of
knowledge must be supplied.

Along with the major system components discussed above, certain
themes will remain dominant during all phases of system design.
Primary consideration will be given to making the system an easy and
natural tool for the geneticist to use. DNA structure entry, editing,
and display will be by way of an interactive progran with nunerous
user~oriented features. Explanation facilities (in the model of the
MYCIN system) will be provided whenever possible, and all knowledge
bases will be made easily extendable and modifiable. Building the
trust of the user is considered necessary to the continued developnent
of the system. The best way to provide for this will be to insure that
it is used vy geneticists, ana that a Strung interaction petween
computer scientists and genetics experts continues.

Tic Vetiainder OF tCuas patt of vur proposat wliit aiscuss in

24
decait our plan ror system devetopmenc in tne rirst year. preeiiy
sunmarized., our four major goals are:

1. Begin the construction of a rudimentary experiment planning
program and an associated planning knowledge base.

2. Build a simulation program which will apply basic enzymatic
actions to DNA structures and summarize the resulting
Structures.

3. Design the knowledge base containing rules and heuristics
used by the geneticists, enzyme knowledge, and knowledge of
experimental methods and physical processes.

4. Complete the construction of an interactive DNA structure

entry and editting systen.

25