一个人的手的功能是什么？我们的手指的功能又是什么？一个人可以给手的功能一个不完全的列表：拿茶杯，打蚊子，等等；手指呢，也一样：夹东西，摁东西，戴指环等等。这很明显是毫无重点而且极其不完整的，然而这似乎就是我们现在的系统生物学进行功能的注释时所采用的功能的表示方法。作者就现存的功能注释的问题提出了自己的看法。
The fiction of function
Jeff Shrager
Department of Plant Biology, The Carnegie Instution of Washington, Stanford,
CA 94305, USA
Received on December 29, 2002; revised on March 24, 2003; accepted on May 9, 2003
What is the function of the human hand? What are thefunctions of our fingers? One could give a very incompletelist of functions of the hand: holding cups, swatting mosquitoes,etc., and similarly for fingers: wrapping around things,pressing on things, holding rings, etc. This clearly is a pointless exercise—incomplete in the extreme—yet it is essentially the goal of many function representation schemes in systems biology, often called ‘functional ontologies’.
There are several problems with listing the functions of biological objects. First, any object has several functions; possibly an infinity of them! The photosynthetic PSII complex, for example, serves many ‘typical’ functions: absorbing light, abstracting electrons from water, passing these to the electron transport chain, producing protons (that energize thepHgradient,
drivingATP synthase to produceATP), and producing free oxygen. One could list all of these, or try to assign them to the subunits of PSII. But independent functions cannot generally be assigned to individual proteins in a complex; like the human hand, most systems cannot be cleanly separated into subunits that have clear individual functions, nor is the composite
function a simple combination of these. Furthermore, many objects serve different functions under different circumstances. For example, in high light, PSII produces damaging reactive oxygen species (ROS). Similarly, many chaperons serve both in the folding of proteins, and in their degradation. Moreover, to try to list the complete set of functions is to exclude novel function through evolution or pathology, where objects serve new functions, often without having undergone any change in their own structure. The typical representational principle, e.g. as implemented in the Gene Ontology project (www.geneontology.org, and 2000), assigns functions to proteins (etc.) as one or a small number of decontextualized properties of objects. Let us look at the ways in which various existing databases code the function
of PSII and its component proteins (or encoding genes). The Protein Data Bank (PDB) classifies the ‘Crystal Structure [of] Photosystem II’ (Zouni et al., 2001) as a ‘Photosynthesis/
Electron Transport’ protein. Although their ‘molecule of the month’ list does not discuss PSII, it does list PSI as ‘These proteins capture individual light photons and use them
to provide power for building sugar’. The Gene Ontology offers no molecular function for PSII (or PSI), but offers a biological process as follows: ‘Gene ontology → Biological process→Physiological process→Photosynthesis’, and at this writing (12/29/2002) lists 25 proteins under this heading. SwissProt (www.expasy.org/sprot) lists many of the individual proteins in PSII. The function of psbA, for example is given as: ‘THIS IS ONE OF THE TWO REACTION CENTER
PROTEINS OF THE CHLOROPLAST PHOTOSYSTEM II’. The complex is not given an EC number of its
own (but many other protein complexes are). Indeed, as I have argued, it would be very difficult to assign a single EC number to PSII as it does several very different things, nor
can one assign EC numbers to the sub-proteins, as they are functionally related in complex ways.
To assign particular, decontextualized functions, to biological objects is an incomplete and intractable approach to biological representation. Unfortunately, function forms the basis of scientific explanations, so it is important not to lose the concept altogether. We can approach a resolution to these problems by recognizing that biological function is an interpretation and not a property of objects; it is what amounts to a local teleological analysis in the context of a particular explanation. For example, when one is trying to explain how the pH gradient arises that drivesATP synthase, PSII plays the role (i.e. has the functional interpretation) of using light energy to break up water into O2 and H+; whereas in the context of explaining how photosynthesis creates damaging ROS, PSII has a different functional interpretation. Unfortunately, when we record or computationally represent function, we usually drop the explanatory context. Even in the case of ‘simple’ enzymes that ‘simply’ catalyze reactions I argue that assigning fixed function is an inadequate representational principle. First, there are numerous multifunctional proteins. But more simply, nearly every catalyst serves at least two functions: the forward and reverse reactions.
Perhaps we can save function by recognizing the universality of the explanatory context within which functions are analyzed. To some extent, this is the approach taken by systems-level ‘pathway’ models, such as Kegg (Kanehisa et al., 2002) and BioCyc (Karp et al., 2002). This approach, however, assigns objects fixed functions, which does not permit important novelty or pathology; they are crystallized explanations, permanently fixing function by virtue of the permanence of the fixed pathways. Therefore, for the same reason that, by my argument, function cannot be fixed,explanations also cannot be fixed into pathways, for to do so is to pre-etermine the set of possible explanations, and so to exclude novelty and pathology.
I propose always dynamically assigning the functions of biological objects in the context of explanations, instead of either listing the decontextualized functions of biological objects,
or crystallizing them into pathways. How would this work? If we cannot assign fixed functions to biological objects, nor fixed pathways, what, then, are we left with? For one thing, objects have physical properties, such as molecular weight, sequence, and absorptive properties. Properties are not fixed either, but they do not change relative to explanations, but rather as a result of physical processes, such as complexing, elongation, and folding. But objects have an infinity of properties. How are we to select which are relevant? In the present theory, which specific properties are relevant is picked out by the explanation in which the objects participate. Put generally: The function of an object is a specific instantiation of an abstract process mapped onto the properties of objects as selected by the explanatory context. The proposed abstract processes are not associated with any particular objects, but are rather phenomena of physical change, such as ligation, translocation, oxidation, etc. There are a large number of these, and each requires some set of objects with particular properties to be instantiated in a particular instance. For example, DNA ligation has at least three ‘slots’ which need to be instantiated from the current explanatory environment: the two strands of DNA to be ligated, and the result. The simplest ligation process dictates constraints on these, for example that the result is composed from the ligands, and that each of these is a molecule of DNA. There may be many different versions of ligation, some of which have more or less specific
requirements, some of which are catalyzed by enzymes, etc. Given physical objects and abstract processes, the task of explanation requires finding a set of process instantiations that implement the overall phenomenon of interest (Shrager and Langley, 1990). The computational activity of instantiating processes in the context of ongoing explanations has been variously utilized in automated reasoning since the 1980s (e.g. Shrager, 1987; Forbus and Falkenhainer, 992).
This principle of contextualized function is akin to de Kleer’s (1984) ‘No Function In Structure’ (NFIS) principle, which dictates that the laws that govern the parts of a system
(i.e. their local functional representation) are not allowed to make reference to the unctioning of the whole system (i.e. the global function). But de Kleer is willing to give specific local functions for components, so long as this does not mention the global function. This limits possible evolutionarily or pathological functions of a natural system. Instead I propose a stronger constraint, called: No Function At All (NFAA), wherein there shall be no functional analysis at all permanently assigned to any object. All of the above descriptions
violate NFIS and do not even approach NFAA. Indeed, the fixed EC number (or Gene Ontology category) assignment toany protein violates NFAA. (The naming system of genes is a
horrific violation of both NFIS and NFAA, many being named after the systemic function that they appear to be involved in, e.g. ‘sevenless’, ‘high light induced’, etc. This would be a
mere oddity if the names did not hang on long after the context of their discovery has long passed into the mists of history, the names becoming permanently associated with the genes
even as the function eventually assigned to the gene have nothing to do with the given name! Under the current theory, there would be no permanent function assigned to genes, and so no
name—or, at least the name would be contextualized, just as the function is.)
Thinking of function as a property of the way in which objects fit into explanations, rather than as a property of objects themselves, has profound implications for the way in which computational biology should represent function. Recall that PSII serves a number of functions: it uses light energy to break down water into hydrogen and oxygen; it uses light energy to abstract electrons from water in order to pass them to PSI; it creates ROS; it absorbs light energy which serves to shade components that are physically behind it (relative to the light); and a large number of other potential functions. Each of these can be given in terms of abstract processes instantiated on PSII in the context of specific explanations. The abstract processes include breakdown of molecules into components, absorption of photon energy, eduction, oxidation, shading, etc. (As above, these can exist in a variety of levels of specificity, and, of course, explanation is hierarchical, so each of these could be further
subdivided into internal explanations composed of smaller processes, or composed into high level, more abstract, processes.) When trying to explain where oxygen comes from when light shines on plants, one might try to instantiate each possible process, perhaps constrained to those relevant to the context, and thence combinations of processes, based upon the slot-property requirements of the processes, and the given physical properties of PSII. In doing so, one tries to construct a set of instantiated processes that completes the explanation— that is, completes the path between current environment (e.g. light and water, but no oxygen), and the goal state (e.g.light, water, and oxygen). PSII, then, can be said to serve a particular function (or set of functions) in the context of a particular explanation. In the context of a different explanation, a different set of abstract processes would be instantiated on PSII (and every other participating object). Thus, we do not require that PSII be given a specific fixed function, or even a set of specific functions; nor do we require that a set of pathways be recorded involving PSII; instead, explanations are dynamically constituted, and PSII is ssigned a function through process instantiation in the context of these explanations.
All of this is not to say that fixed functional representations and pathway databases are wrong or useless; indeed they canbe correct, and quite useful even in their limited way. It is just to say that they are the wrong way, in my view, to understand and register function if we are to reason about it flexibly, such as in the cases of evolution and pathology. To do this, instead of being registered as a fixed property of biological objects, or in crystallized pathways, function must be determined in the context of dynamically formed explanations.
I would like to leave the reader with a conundrum.We think of protein complexes as having a hierarchy of structure: the PSII is composed of numerous proteins, each with a complex structure of domains, and so on. And one may also move upward from PSII, to the photosystem, the chloroplast, etc.—the aspects that we commonly call a ‘systems level’ analysis; systems composing into super-systems, and so on. But how did we choose PSII as an object worthy of having a function to begin with? The same can be asked of any object at any level of biological analysis! It is interesting to consider that the particular decomposition of the system under study—the very selection of components into an explanatory foreground—depends upon the explanation at hand in the very same way that I have argued defines biological function of the selected object. It may be that the selection of objects themselves interacts with the dynamic explanatory process that assigns them function. I cannot say that I can envision how this will allwork out.
ACKNOWLEDGEMENTS
I amindebted to Dafna Elrad, AndrewPohorille, Steve Bagley,Amelia Ireland and a number of reviewers for comments thatgreatly improved this commentary.
REFERENCES
de Kleer,J. (1984) How circuits work. Artificial Intell., 24, 205–280.
Forbus,K. and Falkenhainer,B. (1992) Self-explanatory simulations:
scaling up to large models. Proc 10th National Conference on
Artificial Intelligence. San Jose, CA, pp. 685–690.
The Gene Ontology Consortium (2000) Gene ontology: a tool for
the unification of biology. Nat. Genet., 25, 25–29.
Kanehisa,M., Goto,S., Kawashima,S. and Nakaya,A. (2002) The
KEGG databases at GenomeNet. Nucleic Acids Res., 30, 42–46.
Karp,P.D., Riley,M., Saier,M., Paulsen,I.T., Paley,S. and Pellegrini-
Toole,A. (2002) The ecocyc database. Nucleic Acids Res., 30, 56.
Shrager,J. (1987) Theory change via view application in instructionless
learning. Mach. Learning, 2, 247–276.
Shrager,J. and Langley,P. (1990) Computational approaches to scientific
discovery. In Shrager and Langley (eds.) Computational
Models of Scientific Discovery and Theory Formation. Morgan
Kaufmann, San Mateo, CA.
Zouni,A., Witt,H.T., Kern,J., Fromme,P., Krauss,N., Saenger,W.
and Orth,P. (2001) Crystal structure of photosystem II from
Synechococcus elongatus at 3.8A resolution. Nature, 409, 739.
