CABIOS INVITED REVIEW
Vol. 13 no. 4 1997
Pages 333-344
Linguistic approaches to biologicai
sequences
David B. SearIs
Abstract
Biologists have long made use of linguistic metaphors in
describing and naming cellular processes involving nucleic
acid and protein sequences. Indeed, it is very natural to view
the genetic 'text' and its sequential transliterations in these
terms. However, a metaphor is not a tool, and it is necessary
to ask whether the techniques used in analyzing other kinds of
languages, such as human and computer languages, can in
fact be of any use in tackling problems in molecular biology.
This paper reviews the work of the author and others in
applying the methods of computational linguistics to biological sequences.
Introduction
Only in recent years has the long-standing metaphor of DNA
as language been rigorously examined, from both theoretical
and practical perspectives. This metaphor arose early in
molecular biology, when nucleic acids were recognized as
strings of nucleotide bases comprising the famous four-letter
alphabet. The complementarity of this alphabet—with the
letter G always pairing with C across the double helix, and T
with A—was seen to permit the faithful replication of DNA
molecules. The metaphor was strengthened when the
relationship of nucleic acids to proteins was elucidated; the
so-called central dogma recognized the fundamental two-step
process that first transcribes a subsequence of DNA into
RNA, and then translates successive triplets of RNA bases to
amino acids according to the mapping called the genetic code.
Later, it was discovered that the RNAs that are translated,
called messenger or mRNAs, are further processed in higher
organisms so as to splice out intervening untranslated regions,
called introns, leaving only the exons of the ultimate transcript. These biological transformations can be seen as analogous at a number of levels to mechanisms of processing
other kinds of languages, such as natural languages and
computer languages, particularly in the approaches pioneered
by Noam Chomsky in his revolutionary work in the field of
linguistics (Chomsky, 1957).
A comprehensive review of this branch of linguistics is
well beyond the scope of this review, but a few examples may
serve to highlight some of the relevant issues. Consider the
following sentence^-——"" " ——-________^
The biologist that the linguist noticed eventually waved.
The lines above the sentence serve to connect each noun
with its corresponding verb, i.e. it was the linguist who
noticed and the biologist who waved; note that the relative
clause thus separates remote features of the sentence that are
meaningfully related to each other. This sort of 'action at a
distance' is very characteristic of natural language and
provides much of the motivation for attempts to account for
such structure using rule-based systems called grammars. It
will be seen that grammars are capable of neatly generating
just such hierarchical descriptions of the components of
sentences. They also capture another important feature of
natural languages, syntactic ambiguity, by which alternative
structural interpretations or parses are possible. In this
example, the dashed lines below the sentence highlight an
ambiguity surrounding the referent of the adverb 'eventually': was the biologist slow to wave to the linguist, or was
the linguist slow to notice the biologist?
The dependencies between nouns and verbs in the sentence
above are nested, insofar as any number of relative clauses
could theoretically be inserted within each other to create
tiers of such dependencies. Somewhat less common in
English are crossing dependencies; an artifical example
would be the following:
^^-~\
The biologist and linguist noticed and waved, respectively.
As will be seen in the next section, whether a language
entails strictly nested or crossing dependencies has significant
consequences for the types of grammars required, with
greater complexity attaching to crossing dependencies.
It is noteworthy that very subtle lexical changes (i.e. at the
level of the words of a sentence) can have drastic effects on
the parse. In the following pair of sentences, changing the
preposition 'by' to the conjunction 'as' completely rearranges
the dependencies in the underlying parses:
The biologist noticed by the linguist waved.
Bioinformatics Group, SmithKHne Beecham Pharmaceuticals, 709 Swedeland Road, PO Box 1539, King of Prussia, PA 19406, USA
© Oxford University Press
The biologist noticed as the linguist waved.
333
D.B.Searls
To appreciate the relevance of these linguistic issues to
biological sequences, consider the various kinds of structure
found in the latter. By analogy with sentences, it might be
claimed that genes have a hierarchical syntactic structure, and
indeed it is natural to draw tree-structured descriptions of
gene structure (Searls, 1993). Just as sentences exhibit distant
dependencies, e.g. between related nouns and verbs, so do
genes exhibit such dependencies, e.g. between splice donors
and acceptors. In this case, syntactic ambiguity would be a
reflection of the phenomenon of alternative splicing; it is well
known that subtle changes in 'lexical' signals in mRNA can
result in patterns remarkably like the example sentences
immediately above. Similar ambiguities can be found in the
apparatus regulating gene expression.
Another analogy to linguistics arises from the fact that
biological strings fold up in three-dimensional space, in such
a way that distant parts of those strings interact with and thus
create dependencies with each other. In RNA, the most
obvious manifestiation of such dependencies is base-pairing
interaction. As will be seen in the next section, such dependencies are very naturally expressed via grammars. Again,
the notion of syntactic ambiguity has a biological manifestation, in this case the phenomenon of alternative secondary
structure (Searls, 1993). Both nested and crossing dependencies are observed, for example in antiparallel and parallel
strands in proteins (Mamitsuka and Abe, 1994).
The author has discussed a number of other correspondences between linguistic notions and biological phenomena
(Searls, 1992, 1993), and has also built tools for patternmatching search based on linguistic parsers (Searls and
Noordewier, 1991; Searls and Dong, 1993). The latter have
extended to the problem of detection of genes in genomic
DNA sequences (Dong and Searls, 1994). These three topics
will be discussed in the following sections of this review.
Readers with interest and background in the field of formal
language theory, the bare rudiments of which are introduced
in the next section, may wish to refer to developments in that
arena that have been inspired by the biological domain
(Searls, 1989, 1995a,b, 1996); this work is also summarized
in a recent review (Searls, 1997).
Formal language theory and biological sequences
The processes of transcription and translation from strings of
one kind to strings of a different kind by processive cellular
machinery suggested to some the behavior of certain kinds
of finite state automata (FSAs). FSAs are simple models of
computation, in fact lying at the foundation of computer
science, that comprise directed graphs with labeled transitions among states. By traversing such a graph from state to
state and emitting the lexical/alphabetic labels upon each
transition, a variety of strings can be generated constituting a
334
language. Brendel and co-workers exhibited FSAs that, on
paper, modeled the processive processes implied by the
central dogma (Brendel and Busse, 1984), and in fact such
automata are implicit in numerous software packages that
perform sequence analysis. For that matter, such packages
often provide capabilities for pattern-matching search for
short substrings of interest, through the use of regular expressions (such as the UNIX grep utility); in formal language
theory, the most basic form of regular expressions corresponds to FSAs in terms of expressive power. Note that FSAs
have a dual nature: they can be seen either as generators of
languages, or as recognizers. A variation on this architecture,
called a finite state transducer, can accomplish both at the
same time; by labeling each transition with both an input and
an output, a true transformation of one string to another can
be accomplished, even more effectively modeling the process
of gene expression.
Given the felicity of this correspondence between FSAs
and biology, the question thus naturally arose as to exactly
where DNA resides on the Chomsky hierarchy of formal
languages. This hierarchy classifies the linguistic complexity
of languages (viewed purely as sets of strings) and relates
them to species of automata required to recognize or generate
them. Similarly, each level of the hierarchy corresponds to a
particular type of grammar, or rule-based system for formally
specifying languages. It happens that regular languages, those
that can be specified by FSAs and/or by regular expressions (a
kind of grammar), occupy the lowest level of the Chomsky
hierarchy. However, there are certain languages, such as the
set of palindromes (strings that read the same forward and
backward), that cannot be expressed by any FSA or pure
regular expression. Fundamentally, this is because FSAs and
regular expressions have no notion of memory that would
permit them to describe arbitrary numbers of dependencies,
other than strictly local ones.
Languages such as palindromes fall on the next level of the
Chomsky hierarchy, that of context-free languages, which
require a pushdown (stack) automaton and a more powerful
form of grammar consisting of rewrite rules. In this system,
the alphabet is augmented with a set of temporary, placeholding symbols or non-terminals, and a string is derived by
starting from some such symbol and applying the rewrite
rules to non-terminals in the developing string until they are
all eventually replaced by a terminal string of alphabetic
elements only. Thus, for example, the grammar consisting of
the rules
S—>gSc
S—>cSg
S—>aSt
S — tSa
S—> e
specifies a set of DNA molecules. The S is a non-terminal,
while the lower-case letters are terminal bases; the e stands
for the empty string, and effectively serves to erase an S. This
grammar specifies an infinite number of strings via derivations like the following, in which the S is repeatedly rewritten
Bioscquence linguistics
using the above rules:
S => gSc => gcSgc => gcaStgc => gcatSatgc => gcatatgc
The reader may confirm, by drawing lines connecting those
nucleotides that were derived by the same rule invocation,
that this grammar creates strictly nested dependencies, and in
fact nested dependencies are the only sort possible from a
context-free grammar. As noted, the resulting language lies
outside the regular languages, though any regular language
can be expressed by a grammar like the one above. Nevertheless, there are yet other languages that elude even the
context-free formalism, in essence because they entail arbitrarily many crossing dependencies. Many of these can be
captured with a context-sensitive grammar: one that can have
more than one symbol on the left-hand side of a rule, as long
as the number of symbols does not exceed that on the righthand side. Relax this latter restriction, and the resulting
unrestricted grammars are able to specify any language that
can be recognized by a Turing machine—the automaton
corresponding to this class of languages. Thus, ascending the
Chomsky hierarchy appears to require more and more computational power to accomplish general-purpose recognition
or parsing of strings, i.e. to determine their membership in a
language specified by some grammar.
Much of the author's work has been concerned with the
formal characterization of the language of nucleic acids, in
terms of its position in the Chomsky hierarchy and related
mathematical questions. For example, the grammar given
above specifies, in an idealized way, an important class of
biological sequences, those exhibiting dyad symmetry. The
strings of this language constitute, in fact, a variety of biological palindrome in which a string reads the same on one
strand of DNA as it does on the opposite strand reading
backward (the so-called reverse complement). The resulting
inverted repeats may also allow a single strand to fold up and
base pair with itself instead of with its opposite strand, in a
structure called a hairpin or, when there are a number of
unpaired bases at the turn, a stem-and-loop. Such secondary
structure is crucially important in, for example, the function
of structural RNAs in the cell. To the extent that the capacity
for secondary structure may be said to be a necessary feature
of the language of nucleic acids, we may infer that they lie
above the regular languages on the Chomsky hierarchy, since
at least a context-free grammar is required to specify such
structure in the general case. In fact, the most general
grammar of orthodox secondary structure can be shown to
consist of the grammar above, augmented with one additional
rule: S —» SS. This rule, which simply doubles the start
symbol, is sufficient to permit arbitrarily branching secondary
structures (Searls, 1989, 1993). It also has the effect of
introducing the potential for ambiguity, in that there are
sequences that can be parsed by more than one path [the
reader may wish to verify this by trying parses with this
grammar on sequences that are double inverted repeats, such
as gatcgatc; these can theoretically form hairpins, dumbbells,
or even cruciform structures, each corresponding to a different parse (Searls, 1992)]; happily, these correspond to
alternative secondary structures, providing another useful
analogy between linguistic theory and biological reality
(Searls, 1992, 1993). Stochastic forms of such grammars, i.e.
where probabilities are attached to each rule, have proven
very successful in machine-learning approaches to characterizing recurring secondary structures, as in the work of
Haussler's group (Grate et al, 1994; Sakakibara et al, 1994).
Still other biological phenomena indicate that the language
of nucleic acids may be beyond context free as well. Pseudoknots (see below) are forms of secondary structure that
require context-sensitive expression in the general case, as do
repeated sequences that are common in DNA and are arguably necessary features of the language. Closure properties
and decidability results suggest that evolution may be a
powerful force toward increasing linguistic complexity, and
that DNA may be inherently ambiguous, non-linear and nondeterministic (all formally defined properties from language
theory). These results are described in Searls (1992), and
presented in greater mathematical detail in Searls (1993,
1997); in the remainder of this review, we will address the
more pragmatic problem of recognition or parsing of such
linguistic features in DNA.
In speaking of higher-order pattern recognition in biological sequences, this linguistic point of view offers one way
to categorize the problem space and suggests established
tools for exploring it. For example, the existence of features
in the domain that are at least context free implies that simple
regular expression search will not suffice. Even in advance of
any formal linguistic analysis, however, this problem was
recognized, and a number of software packages have offered
enhanced regular expressions with 'escapes' to specify, for
example, inverted and direct repeats (Staden, 1990). Some
early programs for recognizing patterns of motifs in proteins
were based on extended regular expressions (Lathrop et al,
1987). Whether or not these offer sufficient expressive power
for a greater range of biological phenomena, the use of more
sophisticated grammars and parsers can be seen to have other
advantages. Perhaps most notable among these is the ability
to create modular, hierarchical rule sets, with detail always
presented at the appropriate level (and, of course, a wellstudied formal foundation). Parsers also typically return treestructured descriptions of the history of a derivation, called
parse trees, which are not only clear depictions of the presumed structure under study, but are also appropriate data
structures for further computational analysis. [One worker, in
fact, has studied genetic structures from the point of view of
transformational grammar, which entails operations on a
presumed canonical parse tree in order to create variations in
surface structure typical of natural language (Collado-Vides,
335
D.B.Searls
1989, 1992, 1993; Rosenblueth et al., 1996); this work, and
that of others (Bentolila, 1996), has primarily dealt with
regulatory regions.] The author has shown that the parse trees
for grammars depicting secondary structure in fact physically
resemble that structure—a trait considered desirable in
grammars purporting to model the (less literal) structures of
natural language (Searls, 1992).
The linguistic metaphor has suggested other approaches to
this domain besides one based in generative grammars and
the tradition of Chomsky (e.g. Pesole et al, 1996). A number
of workers, notably Trifonov and his colleagues (Trifonov,
1988, 1989, 1993), have borrowed in part from speech
processing and even cryptanalytic methodologies in their
analyses of biological sequences, in order, for example, to
identify and characterize 'words' or short recurring substrings
of DNA that are likely to have a functional significance. In
fact, such significant words, often recognition and/or binding
sites for other molecules which act upon the DNA, are an
important concern in systems attempting to recognize the
higher-order structures that contain them. The difficulty arises
when such words are imperfectly specified, i.e. when an exact
instance of a word is not necessary, but rather a whole array of
'synonyms' will serve, more or less, the same purpose.
Although it may be possible to describe a canonical consensus sequence for such a feature, for purposes of recognition it is necessary to allow for imperfect matching, and in
general for some notion of assessing the 'cost' of a match.
The most obvious approach is simply to allow for, and count,
mismatches at the level of individual bases (so that the cost is,
in effect, the Hamming distance). Biologically, this would
correspond to a base substitution. However, mutations involving insertions and deletions are also common; the so-called
edit distance between words (or, more often, entire strings)
accounts for the total number of such operations needed to
transform one string into another. Algorithms for efficiently
finding such a minimum edit distance alignment between two
strings (allowing gaps) have historically played an important
role in computational biology (Waterman, 1988). Again at the
level of words, a more sophisticated variation on counting
mismatches has been to compile frequency tables for the
number of times each base occurs at each position in a
number of exemplars; this frequency table is converted by a
variety of techniques to a weight matrix which is used to
assess the cost of a match over the whole word (Staden,
1984). Yet more sophisticated methods have been applied
involving hidden Markov models, connectionist techniques,
and the like, demonstrating that 'higher-order' structural analysis is important even at the level of word recognition in
biology (reviewed by Gelfand, 1995).
Thus, what we have termed the linguistic view of biological sequences involves challenges for both word recognition and syntactic analysis. We will describe practical
approaches to both these problems in succeeding sections.
336
A pattern-matching parser
Logic grammars are a well-studied set of grammar formalisms that are closely related to the logic programming
language Prolog. Most Prolog language implementations, in
fact, offer the capability to compile definite clause grammars
(DCGs) directly into executable code that constitutes a
recursive-descent parser for that grammar (Pereira and Warren,
1980). The built-in theorem prover of Prolog in effect becomes
the parser, and significant advantage can be taken of the listmanipulation and unification features that are a great strength
of logic programming.
In a DCG, a rule such as S—>gSc would be written
s—*[g] , s , [ c ] . Non-terminals are represented as Prolog
atoms and terminals are given inside square-bracketed lists.
This would be compiled to the Prolog rule:
s(S0,S):-S0=[g|Sl],s(Sl,S2),S2=[c|S].
Notice that a pair of variable parameters have been added
to the non-terminals; these difference lists represent the input
string and the remainder of the input string, respectively, after
that non-terminal is successfully parsed—in effect, the span
of the non-terminal on the input. In other words, the difference lists manage the input string, passing it from element
to element, and hiding this 'implementation detail' from the
user at the level of the DCG itself. Terminal lists in the DCG
actually consume elements from the input list, in this case by
unifying it (using the = operator) with a list having the
terminal(s) as its head and the remainder list as its tail. It can
be seen that the difference lists are arranged so that the span
of the left-hand side non-terminal is that of the entire righthand side of the rule. Thus, the rule S—>e, written in DCG
form as s —• [ ], could be translated as simply s (S, S ) . For
the overall grammar, actual top-level calls to s would succeed in forms such a s s ( [ g , g , g , c , c , c , ] , [ ] ) , signifying that the non-terminal s indeed can span the entire input
list.
DCGs also allow the embedding of arbitrary Prolog code in
rules, set off by curly braces, and the attachment of additional
parameters to non-terminals. These features raise the formal
power of DCGs well beyond context free, in fact to the top of
the Chomsky hierarchy. Moreover, with the recursive-descent
parser inherent in Prolog, it is the responsibility of the
programmer to write efficient, terminating rule sets. This is
offset by the advantages Prolog offers in rapid prototyping
and the ability easily to define new syntaxes and metalanguages especially tailored to a particular domain. The author's
syntactic pattern-recognition system for biological sequence
data, called GenLang, is a case in point (Searls and Dong,
1993).
GenLang was designed to process efficiently the huge input
strings of DNA sequence data currently available and being
produced at an exponentially increasing rate. Instead of
Biosequence linguistics
Prolog's linked lists, the input is represented in a ' C array for
efficiency which, however, entails extra bookkeeping behind
the scenes in the parser; in fact, up to a dozen 'hidden'
parameters are attached to non-terminals in order to manage
information about the input string, the parse tree and the cost
of imperfect matching.
Queries take the form <pattern>:<parse variable>=>
<input>, where the pattern generally contains the top-level
non-terminal in the grammar, the parse variable is a logic
variable to which a parse tree will be bound, and the input is a
DNA string (or file containing such a string). One other novel
infix operator is the gap, denoted by an ellipsis, which simply
consumes some length of input that may be either unbounded
(...) or bounded by a minimum and maximum extent (e.g.
3 . . .7 5). Otherwise, GenLang grammars appear very much
like ordinary DCGs, except that most objects, such as terminals,
non-terminals, and gaps, may have attached to them lists of
attributes, of the form:
<object>:[<attributei>,..., <attributen>]
where the attributes are generally operations on keywords,
and are of four types: (i) control operators, which modify the
course of the parse and the position on the input string
(departing from 'pure' logic programming, usually for the
sake of efficiency); (ii) constraint attributes, which impose
limits on quantities denoted by keywords, such as the cost
(normally, the number of mismatches) of a subtree; (iii)
specification attributes, which redefine the values of keyword
quantities with arbitrary expressions; and (iv) assignment
attributes, which bind the values of keyword quantities to
logic variables that may be carried through the parse and used
to report information at the top level. Besides the control
attributes which can modify the backtracking search, another
set of features, using the prefix operator @, can control the
position of the parse on the input string, allowing for arbitrary
translocations and additional kinds of constraints. This syntax
is demonstrated in the following (relatively complicated)
GenLang rule:
orf:[once,cost=C-S/10,
parse=[span,cost] ] —•
v
atg',...:[step=3,S=(size>30)],
@End, stop: [C=cost] ,@End.
Here, the essential framework of the rule states that the
feature o r f consists of the terminal string 'atg', followed by
a gap, followed by the feature s t o p . The control attribute
s t e p = 3 specifies that the gap is to increase in increments of
three, and the constraint attribute s i z e > 3 0 indicates that the
range 30 or less can be ignored. The latter is combined with
an assignment that binds the value of s i z e to the logic
variable S, just as the C = c o s t assigns to C the cost of the
stop. The cost of the rule, by default, is the sum of the costs of
its components: the number of mismatches in the terminal
strings plus (recursively) the costs of any non-terminals.
Here, however, the specification attribute c o s t = C - S / 1 0
redefines the cost of this rule to be an algebraic function of the
size of the gap and the cost of the s t o p feature. Another
specification attribute controls what information will appear
in the parse tree; in this case, just the numerical span of the orf
and its computed cost, without the detailed subtree that would
ordinarily be displayed.
The control attribute keyword o n c e indicates that this rule
can only succeed once in any starting position, and in this
case prevents backtracking into the interior gap. (The effect
of this keyword is similar to that of the Prolog 'cut', but it is
just one of a family of such controls on the expansion of the
parse.) This insures that orfs always end at a stop codon, and
never include one in frame. The ©Ends in the rule body refer
to position in the input string; the first one binds the position
just before the s t o p to the variable End, and the second,
since the variable is now bound, resets the input to that
position. Thus, the span reported for the orf would extend up
to, but not include, the stop codon. Again, there are a large
number of variations on the @ control syntax that allow for
arbitrary movement on the input.
Not only are gaps first-class objects in GenLang, but they
are in some ways the most important objects in terms of the
implementation. This is because gaps, which 'skip over'
input, are the search engines for individual features of interest
that they precede. Since gaps produce the majority of the nondeterminism, or backtracking behavior, in a grammar, they
not only must be made very efficient, but they are treated
specially by the grammar compiler. Gaps are typically not
executed immediately in the course of a parse, but rather are
'packaged' and passed down the parse tree to succeeding nonterminals, until they encounter some feature with which they
may combine for more efficient evaluation. For example, the
combination of such a Mazy' gap with a terminal string can be
more efficiently evaluated than by brute force matching a la
ordinary DCGs. GenLang will, at the option of the user,
create a hash table of the position of every substring of a
given length in the input string, so that a gap/string combination can simply be looked up for immediate evaluation,
instead of scanning.
Lazy gaps also permit greater variety in the search strategy,
which in the logic-based parser is ordinarily breadth-first on
the input, i.e. all applicable rules are tried at every position
before moving on to the next position. A lazy gap, however, is
passed to the first applicable rule, which can be tried in every
possible position allowed by the gap, before passing the gap
to the next alternative clause or rule. This describes depthfirst search on the input. The search strategy in GenLang can
be varied locally through the use of the rule attributes d e e p ,
w i d e , or b e s t .
337
D.B.Searls
Language
Automaton
Recursively
Enumerable
Languages
Turing Machine
Contextsensitive
Languages
Linear-Bounded
Automaton
Grammar
Unrestricted
Grammar
Recognition
Undecidable
Dependencies
Arbitrary
Unknown
Baa - »
>•>, aContext-Sensitive
Grammar
NP-Complete
Crossing
Repeats
—^H
Pseudoknots
At - » aA
Context-Free
Grammar
Context-Free
Languages
Polynomial
Nested
S-^gSc
Regular
Languages
Biosequences
Finite-State
Automaton
Regular
Expression
Orthodox
Secondary
Structure
'mill Q
Linear
Strictly Local
Central Dogma
((gla)(clt))*
Illl
Fig. 1. The Chomsky hierarchy of languages. Each class of languages subsumes the ones below it, and is associated with a certain type of automaton capable of
recognizing and/or generating such languages, as well as a form of grammar which can specify them. General-purpose recognition or parsing of languages, given
grammars that specify them, is increasingly costly as the hierarchy is ascended. On the right are the types of dependencies that can be captured, as well as
phenomena exhibited in biological sequences that appear to reside at each level of the hierarchy (see the text). Note that, while any given type of non-context-free
phenomenon may be detected in polynomial time, the recognition of the general class of such phenomena is NP-complete. Also, superposition of context-free
features may produce languages by intersection that are beyond context free.
In order more easily to specify some of the linguistically
complex features expected in this domain, the author has
developed and characterized a formalism called string variable grammar (SVG) (Searls, 1988, 1995a), as another augmentation to DCGs. SVGs allow logic variables as objects in
the grammar, standing for uninstantiated terminal strings. It is
thus possible to specify such features economically as:
tandem_repeat —• X,X.
inverted_repeat —• X, .. .,-X.
pseudoknot — X, . . . , Y,-X, . . . , ~Y.
where the tilde is a prefix operator denoting reverse complement. Like gaps, string variables can assume any length,
though in practice they are generally constrained by the
appropriate attributes. (A gap, in fact, can be seen as simply
an 'anonymous' string variable.) In the inverted repeat rule
above, the string variables constitute the two base-paired
halves of the stem, while the interior gap represents the
unpaired loop.
Note that a pseudoknot is a kind of interleaved inverted
repeat, where half of the stem of one stem-and-loop resides
within the loop of a second stem-and-loop. A number of stemand-loop structures are illustrated in Figures 2 and 3, and the
latter also contains a pseudoknot. To specify these and many
338
related features (which are often beyond context free) using
ordinary DCGs requires much more complex rules (Searls,
1989).
As has been noted, the cost of a feature is by default the
number of mismatches in terminal strings eventually contained within the feature, though this can be arbitrarily
redefined. A number of other cost functions are available as
built-ins, including weight matrices (with a variety of evaluation methods), edit distance and a facility for substitution
matrices that allow for individualized costs for every possible
alphabetic substitution. Costs combine and propagate up the
parse tree until they are constrained and/or redefined by the
appropriate attributes, so that the entire parse tree can have its
'stringency' adjusted both globally and locally over subtrees.
A number of GenLang grammars have been described
previously; here, results derived from several of them will be
reviewed. The characteristic cloverleaf secondary structure of
transfer RNA (tRNA), for example, is easily described by an
SVG (Searls, 1989). A grammar originally designed for
relatively simple bacterial tRNAs (Searls and Liebowitz,
1990) was enhanced so as to handle a greater degree of
variation, including a specific type of intron, and to detect
conserved regions in the primary (lexical) structure via
weight matrices, in addition to the secondary structure. The
resulting grammar, when used to parse the GenBank database
Biosequence linguistics
Acceptor '
Arm —
TVCArm
tRNA('Cto'): span»90470/90S83
• g g t t g t f : span=90470/90477 coSt.O
Cum: list»*tg"
D arm: span-90479/90496
stem: list«'gcc'
loop: list»'gagcggtctaa' cost«0
-seen: llst.'ggc" cost-0
base: l i s t . - g "
antlcodon arm: span-90497/90546
"cccga": span«90497/90S02
lOOp: CO9t-0
%
pre_codon: l i s t - ' t f
-codoni l i s t - ' c a a '
^ *
purinet liat-'g"
intron: span»90S08/90S40
. . . : alzo-7
•ttg": 8pan*90S15/90518 cost»0
. . . : size.22
base: lisc>'c*
-•cctga'i span-90541/90546 cost.O
, - extra arm: span-90546/90559
~V
. .. : size.13
*^ TfcsiC arm: span-90559/90576
stem: llst«'aagag"
's
loop: list-"ttcgaat" cost-6
-stem: llst»*ctctt" cost»0
s
' - • g g t t g t f : span-90576/90583 cost-0.5
Fig. 2. Transfer RNA. The generalized structure of the tRNA molecule is schematized at the left, showing both conserved features of primary sequence and
secondary structure due to folding and internal base pairing. In the center is a diagram of yeast (Saccharomyces cerevisiae) chromosome 111,315,356 base pairs in
length, with blocks showing the 10 known tRNA genes. The tRNA grammar described in the text detected all 10 in a parse that took under half an hour on a Sun
SPARCstation 2; a typical parse tree is shown at the right.
of all known DNA sequences, was able to find >95% of
known tRNAs among bacteria, invertebrates and primates (a
total of >700) with virtually no false positives (Searls and
Dong, 1993). This compared favorably with hard-coded
procedural tRNA finders developed by others, and even
achieved comparable efficiency. Figure 2 shows the results of
running this grammar against the entire length of yeast
chromosome III.
Another, more complex higher-order structure found in
certain organisms is that of the group I intron. This is a form
of self-splicing intron which is characterized, again, by both
recurring primary sequence features and by a distinctive
secondary structure, this time involving a pseudoknot (Cech,
1988). These are handled uniformly by the SVG formalism,
as shown above. In actual use, this significantly more
complex grammar was able to find the majority of such
features in the genome of a fungal mitochondrion, together
with a variety of other features which are typically pointed
out in such sequences when they are described by biologists.
These results, illustrated in Figure 3, suggest the utility of
grammars both as declarative descriptions of higher-order
structures in a consistent, well-founded syntax, and as a
means of assisting in the annotation of sequence data via
parsing. Others have created systems capable of recognizing
tRNAs, group I introns, and structures of similar complexity,
using pattern-recognition techniques that are also syntactic in
flavor, though not explicitly based on grammar formalisms
(Gautheretef a/., 1990).
Besides encouraging formal declarative descriptions of the
many complex structures recognized in sequence data, the use
of logic grammars for their procedural interpretation offers
many advantages. Prolog is an outstanding rapid prototyping
language, which carries over into grammar design, and it is
also easy to write compilers and design metalanguages for
abstract interpretation of grammars with domain-specific
syntactic features and speed-ups. The ability to embed in-line
code in grammar rules, including foreign code or even entire
expert systems, means that the grammar can be used as a
framework to house special-purpose algorithms and apply
them selectively to regions of interest determined by a higher
structural context. Particularly for computationally expensive
algorithms, such focus can lead to considerable savings.
Moreover, the ease of passing parameters up and down the
parse tree means that these algorithms, which are often highly
parameterized, can be fine-tuned to suit the context; also, the
results can be synthesized and combined under the direction
of the housing grammar framework.
These virtues of the linguistic approach to sequence analysis meet their most stringent test in the problem of detecting
genes in raw sequence data, as described in the next section.
Syntactic gene finding
With the development of large-scale DNA sequencing
technology, more and more long stretches of uncharacterized
DNA are becoming available, and a natural question is to ask
where the genes are in such sequences. Since genes represent
only a few percent of the DNA of the genome, answering the
question is not straightforward, and this has stimulated the
development of increasingly effective gene-finding algorithms.
For some time, the accepted way of approaching this
question was simply to attempt to detect protein-coding
regions, which may be expected to have certain characteristics that distinguish them from non-coding regions. Finding
339
D.B.Searls
o
1111111
"a u c a g g
-a ag a u a u a g u c c — i
p9
o
group I intron: span.3237/4660
conserved regions: spanc3237/4638
P region: span.3237/3249 cost.297
list.'aatttcaaaaac
P-Q loop: span«3249/3283
. . . : size*34
0 region: span«3283/3293 cost=230
list-"gattcgaagc"
. . . : size=135
R region: span-3428/3442 cost.258
list.'gttcaacgactaag'
R-S loop: span.3442/4626
. . . : size.82
ORF: span.3524/4508 slze>984
.. .: size.118
!
S region: span.4626/4638 cost.134
list.'aagacatagtct'
\ * secondary struture: span.3230/4660
'* *
p3: span«3230/323S list.'ttggg'
* p4: apanc3240/3246 list«*ttcaaa'
s*
p5/-p5i span=3249/3283
-p4; span.3285/3291 11st.•tttgaa'
p6/-p6i sp»n-3290/3431
,„ p7. span.3435/3440 list.-gacta"
-p3s span=3476/3481 list='cccaa'
s
. " p8/-p8i span«3481/4625
"* -p7: span«4632/4637 list«"tagtc"
•: p9/-p9: span«4640/4660
Fig. 3. Group 1 introns. The conserved primary sequence regions and characteristic secondary structure of these self-splicing introns are shown at the top. At the
lower left is a diagram of the circular mitochondrial genome from a fungus (Podospora anserina). 100 314 base pairs in length, with group I introns marked. The
grammar for this feature produced parse trees such as the one shown, and was able to find 25 out of the 33 known introns in this genome, as well as a high
percentage of the tRNA genes and open reading frames (Searls and Dong, 1993).
long open reading frames is the first heuristic in searching for
coding sequences, but it does not suffice. In vertebrates, the
average length of the coding exons (—130 bases) is not
sufficient reliably to distinguish them from open reading
frames that could be expected to exist in extensive noncoding regions by random chance. Codon usage statistics can
be employed to evaluate a putative coding sequence for its
tendency to use preferred triplets, rather than just any triplet
that is not a stop codon, and this approach has been extended
to longer 'words', with hextuple frequencies being an oftused heuristic to distinguish coding from non-coding sequence.
340
More subtle tendencies can also be exploited; for example,
the early TESTCODE algorithm (Fickett, 1982) made use of
an observed tendency in coding regions for bases to recur in
the same relative position with a cyclicity of length three, i.e.
with respect to the reading frame. Such statistical tests
produce a 'score' indicating coding potential within a window
on the input sequence, with the reliability of the score
increasing with window size. Unfortunately, once again the
average size of vertebrate exons is generally smaller than the
optimum window size, and by and large these so-called
compositional methods do not offer anything approaching
Biosequence linguistics
absolute reliability (Fickett and Tung, 1992). Nevertheless,
compositional methods have become increasingly sophisticated, and suffice to find significant portions of a number of
exons out of the many that may constitute a gene, thus
drawing attention to the appropriate region. Moreover,
approaches such as GRAIL have made use of the powerful
technique of combining evidence from several compositional
methods by way of a neural net (Uberbacher and Mural,
1991).
The compositional approach may be contrasted with the
strictly syntactic or rule-based approach to gene recognition.
The author showed that the basic 'rules' of gene expression,
including splicing and even flanking control regions, can be
conveniently encoded in a grammar that is able to recognize
DNA sequences capable of producing proteins (Searls, 1987;
Searls and Noordewier, 1991). This approach had great
appeal because of the known virtues of declarative, hierarchical specifications. However, just as syntax underspecifies natural language, allowing grammatically correct yet
nonsensical sentences, so does a gene syntax underspecify the
valid protein-encoding genes in a genome. Apart from the
deterministic machinery of protein translation, such a grammar
depends upon reliable recognizers of specific features in the
DNA, such as start sites and splice junctions. Identifying such
sites—the word recognition problem described above—turns
out to be at least as dicey as finding coding sequence. This key
distinction, between syntactic and compositional methods,
was recognized early and termed 'gene search by signal'
versus 'gene search by content' (Staden, 1984).
In fact, the problem of syntactic recognition of genes may
be thought of as the dual of the problem of compositional
recognition of coding regions, i.e. any comprehensive solution to one would effectively constitute a solution to the other.
In other words, if we could invariably identify the signals for
where genes, transcripts, splice sites, etc., occur, we would
know the coding regions, and if we could recognize coding
sequence with exactitude, we could easily infer the structure
of the genes. It soon became evident that neither approach in
isolation would suffice, and the most successful programs
today are 'hybrid' systems that combine compositional evidence with rule-based frameworks that assemble coding regions
into plausible gene structures. Thus, the experience in this
domain can be seen to have paralleled that in fields such as
character recognition and speech processing, where hybrid
statistical and syntactic pattern-recognition systems have
proven most effective (Searls and Taylor, 1993).
A GenLang gene grammar was developed in which conventional weight matrices were used to detect signal words
for translation and processing, and in some cases promoter
signals for transcription (see Figure 4). In addition to syntactic and size constraints, compositional measures contributed to the cost both locally within putative exons (based on
hextuple frequencies for coding versus non-coding regions)
and globally for entire exons. As before, sequences were
parsed left to right and a parse tree built, which also provided
a structure by which both signal and compositional costs
could be combined, with weights and thresholds determined
empirically. Thus, signals were only searched for in contexts
of the developing parse where they 'made sense'.
Other rule-based architectures had also incorporated compositional measures into frameworks for assembling genes
(Fields and Soderlund, 1990; Gelfand, 1990; Guigo el al.,
1992). A problem common to such gene prediction programs
is that of the combinatorics of exon assembly: it is not
uncommon for genes to have more than a dozen exons, and
since exons cannot be delineated with absolute certainty, an
exponential number of combinations may have to be
postulated and evaluated. Besides careful adjustment of
thresholds for exon prediction from compositional evidence,
a variety of techniques can and have been used to ameliorate
the computational expense of assembly. By far the most
successful approach to this problem has been the adoption of
dynamic programming algorithms (Gelfand and Roytberg,
1993; Snyderand Stormo, 1993). These techniques, in theory,
allow all possible locations of exons to be evaluated in the
context of an entire gene architecture, and in polynomial
time. Although parsing of context-free languages can be
performed with similar efficiency, the problem of examining
all parses, where there may be an exponential number of
them, may be intractable even so.
In GenLang, the technique of chart parsing (a linguistic
version of dynamic programming) was used to achieve reasonable efficiency, and accuracy was comparable to the best
results at the time (Dong and Searls, 1994). However, ever
more efficient and accurate systems are now outpacing what
is likely to be possible with a general-purpose parser (Burset
and Guigo, 1996; Fickett, 1996). Moreover, the function of
the gene grammar in this case is primarily to serve as a
framework on which to organize compositional sensors, which
in all provide the most effective cues. What, then, is the future
of syntactic models of gene structure?
One answer may depend on the accumulation of a more
complete catalog of genes leading to a richer set of themes
and variations (for which grammars are particularly well
suited). For example, compositional methods are known to
perform better on some classes of genes than others (perhaps
partly because of skewed training sets), and they have very
little to say about certain specialized gene structures, such as
those associated with the immune system which undergo
distinctive types of rearrangement in the genome in advance
of gene expression. Such specialized forms of genes, gene
regions, and organized gene families may prove to be more
the rule than the exception, as a more global view of genome
structure emerges. The same may be true of mechanisms of
gene expression such as splicing. Another increasingly
important theme is that of alternative forms of transcription
341
D.B.Searls
promoter
DNA
f—
1
1
1
i
exon
1
\,5'
intron
/ ^ ^
W///A\
exon
W/////M
.
\ '
UTR
i
intron
^ ^ ^ ^ ^
*
\
i
\
s
1
s
s
exon
warn.^
f
*
*
^1
S
**, '
'3'UTJ^
w
V
i
i
i start
i codon
s
*
primary i
transcript |
\
polyA
site
mRNA
, stop
• codon
Gy Ay
gene: span.19116/21079
upstream: span>19416/19S04
CCAAT-box: spanBl9416/19428 cost=64 1istc'cctgaccaatga*
...: size>46
TATA-box: span*19474/19489 cost*399 list>'gaataaaaggccaga'
...: size.IS
cranscript: span»19504/21079
CAP: span>19S02/19S10 cost-284 lisc-'gcacatat•
...: size.3
•cttccg": span-19513/19519 cost.O liet-'cttccg'
...• size.40
translation product: span»19559/20977
M: span>mS9/19S62
•atg-: span>195S9/19S62
exon: span.19559/19649
Intron: size>122
splice: span.19651/19773
donor: cost>146 listB'gtaagca'
...: size.81
acceptor: cost>3(6
branch: span«19739/19744 cost«10 list-'ctaat'
pyrimidine_rich: size-25 llst-'tttgtatctgatatggtgtcatttc'
ag: span>19770/19774 cose.338 Ust-'cagaa
axon: span.19771/19996
intron: sizQ*855
splice: span.19996/20851
donor: cost.lSl list>'gtgagtt'
...i size.808
acceptor: costa433
branch: span-20811/20816 cost*61 lists'ctcaa'
pyrlBidine_rich: size-31 list-'catgcctgtgtgttattttgtcttttgccta'
ag: span.20848/20852 cost.250 llsc'cagc1
exon: span.20851/20977
stop: llst-'tga'
...: slzo.96
polyA site: span.21073/21079 list>*aataaa'
Fig. 4. Protein-encoding genes. DNA (top) associated with genes is transcribed to RNA, from which introns are spliced out to produce mRNA. This is then
translated to protein, according to the triplet-based genetic code which includes start and stop signals. A variety of other signals are important in controlling gene
expression and processing. Genes and gene families such as sets of hemoglobin genes spanning a 73 326 base pair region (center) can be parsed using a gene
grammar, producing detailed parse trees like that shown (bottom).
for the same gene, e.g. multiple start and end sites, and
especially alternative splice sites, which add layers of combinatoric variation well suited to grammatical expression
(Searls and Noordewier, 1991). All this, and the increasingly
elaborate schemata of gene regulation currently being elucidated, will inevitably lead to more detailed and probably
more hierarchical models that are the natural domain of
syntactic description and analysis. As shown in Figure 5, such
models have potential for extending all the way from
nucleotide-level signals to global organization of gene regions.
342
The notion of modeling gene structure and the mechanisms
of gene expression, in fact, is the most appealing aspect of the
syntactic approach. Compositional methods (like neural nets
in general) may be thought of as capturing strictly empirical,
post hoc, statistical correlations, with little or no rational
model involved beyond some initial, crude structuring of the
problem. Of course, parsers such as GenLang may also be
employed strictly as opportunistic detectors using ad hoc
rules without any mechanistic forethought, but a grammar at
least provides an opportunity for rational design correlated
Biosequence linguistics
immunoglobulin_superfamily_gene_region
v _ g e n e s ( S 0 ) , . . . , d_segments(SIJ,
{recombinable(S0,Sl,S2)}
..., transcriptional_enhancer,
j_segments(S2),
..., c_segments.
v_genes(S)
> [] |
v_gene, recombination_signal(S) ,
v_gene
>
promoter,
. ., leader_peptide,
>
v_genes(S)
intron, v_segment.
recorabination_signal([H,G,N])
>
"cacagtg":(cost<2,H=sequence), 12 ... 23:[G=size],
"acaaaaacc*:[cost<2,N=sequence).
Fig. 5. A fragment of a possible immunoglobulin superfamily gene grammar.
Unlike the 'vanilla' gene grammar used for Figure 4, this specialized
grammar would capture the specific architecture of this family. The crosssection of rules illustrated span from the level of genome organization down
to specialized signals. While the lower-level rules might be useful in
recognition, the major utility of such grammars might be in the organization
of knowledge about genetic organization in computer-readable form.
with known biological structures and phenomena. Indeed,
one of the motivations for the use of generative grammars in
natural language analysis is the hope that forcing observations to conform to concise declarative formulations will lead
to clarifying generalizations, almost as a by-product; the
extreme version of this view is that the grammars discovered
may in fact reflect actual cognitive mechanisms in some
manner. The author has described a number of ways in which
grammars and parsers can be seen to model biological
mechanisms and structures (Searls, 1993); as noted above,
for example, parse trees for secondary structural grammars
physically resemble those structures, even to the extent that
ambiguity in the grammars reflects alternative secondary
structure (Searls, 1992).
It may be argued that the necessity of including compositional data to detect genes reliably compromises the validity
of the syntactic approach. In fact, the situation is no different
than for natural language, where the set of semantically
meaningful sentences is a small subset of the grammatically
acceptable sentences. Grammar may be thought of as a
framework, reflecting some core mechanism of gene expression to which protein 'utterances' are bound to conform, but
in addition to which they are subject to biochemical and
thermodynamic context and, more broadly, evolutionary
selection. In the terminology of Chomsky, a gene grammar
should reflect the competence of the cell to deal with the DNA
sequence, and not necessarily its actual performance. No
doubt, current grammars are inadequate as models of the
cellular machinery, and grammars may or may not prove to be
suitable receptacles for such knowledge as it accumulates, but
they appear promising for the moment.
For example, current thinking on the mechanism of gene
regulation (e.g. Tjian, 1995) and splicing (Niwa et ai, 1992)
points toward multiple interacting factors such as (in the case
of splicing) the combined 'strengths' of several recognition
sites, the lengths of flanking exons and/or introns, local
conformation of the RNA, and other characteristics of the
sequence neighborhood. Not only do grammars intrinsically
model such hierarchical combinations of factors, but they
provide a framework for the formal evaluation of the
semantics of the whole in terms of operations on its parts,
where those operations are functionally mapped from strictly
syntactic combinations of symbols. Such functional mappings may be as simple as combinatoric 'truth-functionals'
determining the state of some binary genetic switch, or as
sophisticated as cooperative thermodynamic binding equations, but the encouraging sign is the emerging theme of the
whole being determined by combined characteristics of the
parts. This is even more evident in the intriguing possibility of
a compositional semantics of functional motifs of proteins,
which could theoretically assemble the 'meaning', qua function, of the protein as a whole.
A subtle counterargument to this view holds that current
gene grammars tend inappropriately to combine syntactic
elements that in fact come into play at different times in
different compartments of the cell. Why, for example, should
a gene finder that purports to model gene expression take
advantage of knowledge of open reading frames (a function
of the translational apparatus) when evaluating the likely
location of splice sites (determined during RNA processing,
long before translation), and vice versa? The author has, in
fact, suggested that separate such grammars could be combined for the purposes of detection, with the understanding
that they are imperfect and incomplete in isolation, but synergistic in their interaction. [This, in fact, has interesting formal
consequences, entailing as it does the intersection of languages (Searls, 1992).] Moreover, it may eventually prove to be
presumptuous to exclude any information from such models;
recent evidence suggests that, in fact, reading frame does
have an effect on splicing, through as yet unknown mechanisms (Naeger, 1992). Clearly, it will be necessary to maintain
an intelligent balance between the use of grammars to model
current knowledge and the need to shape them in new
directions that seem to produce better results in parsing,
possibly leading to new or altered hypotheses in the process.
Acknowledgement
This work was supported by the Department of Energy. Office of Energy
Research, under grant number ER6I371.
References
Bentolila.S. (1996) A grammar describing 'biological binding operators" to
model gene regulation. Biochimie, 78. 335-350.
Brendel.V. and Busse.H.G. (1984) Genome structure described by formal
languages. Nucleic Acids Res., 12, 2561-2568.
Burset.M. and Guigo.R. (1996) Evaluation of gene structure prediction
programs. Genomics, 34, 353-367.
Chomsky.N. (1957) Syntactic Structures. Mouton, The Hague.
343
D.B.Searls
Cech.T.R. (1988) Conserved sequences and structures of group I introns:
building an active site for RNA catalysis—a review. Gene. 73, 259—271.
Collado-Vides.J. (1989) A transformational grammar approach to the study
of the regulation of gene expression. J. Theor. Biol.. 136, 403-425.
Collado-Vides.J. (1992) Grammatical model of the regulation of gene
expression. Proc. Null Acad. Sci. USA. 89. 9405-9409.
Collado-VidesJ. (1993) A linguistic representation of the regulation of
transcription initiation. BioSysiems, 29. 87-128.
Dong.S. and Searls.D.B. (1994) Gene structure prediction by linguistic
methods. Genomics. 23. 540-551.
Fickett.J.W. (1982) Recognition of protein coding regions in DNA
sequences. Nucleic Acids Res.. 10, 5303-5318.
Fickett.J.W. (1996) Finding genes by computer: the state of the art. Trends
Genet.. 12, 316-320.
Fickett.J.W. and Tung.C.S. (1992) Assessment of protein coding measures.
Nucleic Acids Res.. 20, 6441-6450.
Fields.C.A. and Soderlund.C.A. (1990) gm: a practical tool for automating
DNA sequence analysis. Comput. Applic. Biosci.. 6. 263-270.
Fields.C.A., Adams.M.D.. Kerlavage.A.R., Dubnick.M.. McCombie.W.R..
Martin-Gallardo.A.. White.O. and Venter.J.C. (1993) Identification of
genes in genomic and EST sequences. In Lim.H.A., Fickett.J.. Cantor.C.R.
and Robbins.R.J. (eds). Proceedings of the Second International
Conference on Bioinformatics. Supercomputing and Complex Genome
Analysis. World Scientific, Singapore, pp. 429-433.
Gautheret.D.. Major.F. and Cedergren.R. (1990) Pattern searching/alignment
with RNA primary structures: an effective descriptor for tRNA. Comput.
Applic. Biosci.. 6. 325-331.
Gelfand.M.S. (1990) Computer prediction of the exon-intron structure of
mammalian pre-mRNAs. Nucleic Acids Res.. 18. 5865-5869.
Gelfand.M.S. (1995) Prediction of function in DNA sequence analysis. J.
Comput. Biol. 2. 87-115.
Gelfand.M.S. and Roytberg.M.A. (1993) A dynamic programming approach
for predicting the exon-intron structure. BioSysiems. 30, 173-182.
Guigo.R.. Kundsen.S.. Drake.N. and Smith.T. (1992) Prediction of gene
structure. J. Mot. Biol.. 226. 141-157.
Grate.L., Herbster.M., Hughey.R., Mian.l.S.. Noller.H. and Haussler.D.
(1994) RNA modeling using Gibbs sampling and stochastic context-free
grammars. In Proceedings of the Second International Conference on
Intelligent Systems for Molecular Biology. AAAI Press. Menlo Park. CA.
pp. 138-146.
Lathrop.R.H.. Webster.T.A. and Smith.T.F. (1987) Ariadne: pattern-directed
inference and hierarchical abstraction in protein structure recognition.
Commun. ACM. 30. 909-921.
Mamitsuka.H. and Abe.N. (1994) Predicting location and structure of betasheet regions using stochastic tree grammars. In Proceedings of the Second
International Conference on Intelligent Systems for Molecular Biology.
AAAI Press, Menlo Park, CA. pp. 276-284.
Naeger.L.K. (1992) Nonsense mutations inhibit splicing of MVM RNA in cis
when they interrupt the reading frame of either exon of the final spliced
product. Genes Dew. 6, 1107-1119.
Niwa.M.. MacDonald.C.C. and Berget.S.M. (1992) Are vertebrate exons
scanned during splice site selection? Nature. 360, 277-280.
Pereira.F.C.N. and Warren.D.H.D. (1980) Definite Clause Grammars for
language analysis. Artificial Intelligence. 13. 231-278.
Pesole.G.. Attimonelli.M. and Saccone.C. (1996) Linguistic analysis of
nucleotide sequences: algorithms for pattern recognition and analysis of
codon strategy. Methods Enzymol.. 266. 281-294.
Rosenblueth.D.A.. Thieffry.D.. Huerta.A.M.. Salgado.H. and ColladoVides.J. (1996) Syntactic recognition of regulatory regions in Escherichia
coli. Comput. Applic. Biosci.. 12, 415—422.
Sakakibara.Y.. Brown.M., Hughey.R.. Mian.S.. Sjolander.K.. Underwoood.R. and Haussler.D. (1994) Stochastic context-free grammars for
tRNA modeling. Nucleic Acids Res., 22, 5112-5120.
Searls.D.B. (1988) Representing genetic information with formal grammars.
In Proceedings of the 7th National Conference on Artificial Intelligence.
AAAI Press, Menlo Park. CA. pp. 386-291.
Searls.D.B. (1989) Investigating the linguistics of DNA and definite clause
grammars. In Lusk.E. and Overbeek.R. (eds). Logic Programming:
Proceedings of the North American Conference on Logic Programming.
MIT Press. Cambridge, MA. pp. 189-208.
344
Searls.D.B. (1992) The linguistics of DNA. Am. Sci.. 80. 579-591.
Searls.D.B. (1993) The computational linguistics of biological sequences. In
Hunler.L. (ed.). Artificial Intelligence and Molecular Biology. AAA1/MIT
Press, Menlo Park, CA. pp. 47-120.
Searls.D.B. (1995a) String Variable Grammar: a logic grammar formalism
for DNA sequences. J. Logic Programm.. 24. 73-102.
Searls.D.B. (1995b) Formal grammars for intermolecular structure. In
Proceedings of the First International IEEE Symposium on Intelligence
in Neural and Biological Systems. IEEE Computer Society Press. Los
Alamitos. CA. pp. 30-37.
Searls.D.B. (1996) Sequence alignment through pictures. Trends Genet.. 12.
35-37.
Searls.D.B. (1997) Formal language theory and macromolecular structure. In
Farach.M., Roberts.F.S. and Waterman.M. (eds), DIMACS Special Year
Volume: Mathematical Support for Molecular Biology. American
Mathematical Society. Providence. Rl, in press.
Searls.D.B. and Dong.S. (1993) A syntactic pattern recognition system for
DNA sequences. In Lim.H.A.. Ficketl.J.. Cantor.C.R. and Robbins.R.J.
(eds). Proceedings of the Second International Conference in Bioinformatics. Supercomputing. and Complex Genome Analysis. World Scientific.
Singapore, pp. 89-101.
Searls.D.B. and Liebowitz.S.A. (1990) Logic grammars as a vehicle for
syntactic pattern recognition. In Proceedings of the Workshop on Syntactic
and Structural Pattern Recognition. International Association for Pattern
Recognition, Surrey. UK, pp. 402-422.
Searls.D.B. and Noordewier.M.O. (1991) Pattern-matching search of DNA
sequences using logic grammars. In Proceedings of the 7th Conference on
Artificial Intelligence Applications. IEEE Computer Society Press, Los
Alamitos, CA. pp. 3-9.
Searls.D.B. and Taylor.S.L. (1993) Logic grammars for document image
analysis. In Baird.H.. Bunke.H. and Yamamoto.K. (eds). Structured
Document Image Analysis. Springer-Verlag. Berlin, pp. 520-545.
Snyder.E.E. and Stormo.G.D. (1993) Identification of coding regions in
genomic DNA sequences: an application of dynamic programming and
neural networks. Nucleic Acids Res.. 21. 607-613.
Staden. R. (1984) Computer methods to locate signals in nucleic acid
sequences. Nucleic Acids Res.. 12. 505-519.
Staden,R. (1990) Searching for patterns in protein and nucleic acid
sequences. Methods Enzymol., 183. 193-211.
Tjian.R. (1995) Molecular machines that control genes. So. American. Feb.
54-61.
Trifonov.E.N. (1988) Nucleotide sequences as a language: morphological
classes of words. In Bock.H.H. (ed.). Classification and Related Methods
of Data Analysis. Elsevier. Amsterdam, pp. 57-64.
Trifonov.E.N. (1989) The multiple codes of nucleotide sequences. Bull.
Math. Sro/., 51. 417-432.
Trifonov.E.N. (1993) DNA as a language. In Lim.H.A., Fickett.J..
Cantor.C.R. and Robbins.R.J. (eds). Proceedings of the Second International Conference in Bioinformatics, Supercomputing, and Complex
Genome Analysis. World Scientific. Singapore, pp. 103-110.
Uberbacher.E.C. and Mural,R.J. (1991) Locating protein-coding regions in
human DNA sequences by a multiple sensor-neural network approach.
Proc. NatlAcad. Sci. USA. 88. 11261-11265.
Waterman,M.S. (1988) Computer analysis of nucleic acid sequences.
Methods Enzymol.. 164, 765-793.
Received on November 28, 1996: revised and accepted on February 2, 1997