The following excerpt comes from Professor Robin Feldman's forthcoming Harvard University Press book Rethinking Patent Law which will be released in 2012. The excerpt is reproduced with permission from Harvard and Professor Feldman.
Looking from the perspective of preemption can be helpful in developing an approach to many of the questions generated by the emerging field of personalized medicine. Personalized medicine is an area of applied research devoted to developing tests that operate on biological and clinical data from a patient (e.g., protein levels, genetic mutations, medical history) to provide diagnoses, prognoses, and treatment regimens specific to the patient. Cases arising in this field include LabCorp, which concerned correlating elevated homocysteine levels with certain deficiencies, and Prometheus, which concerned determining metabolite levels after administering a particular class of drugs for gastrointestinal disorders and adjusting the dosage of the drug based on the metabolite levels.
As discussed earlier, the Federal Circuit's decision in Prometheus seemed to suggest that most life science inventions would satisfy the requirements of patentable subject matter, while the PTO's application of Prometheus could lead to the rejection of numerous inventions in this arena. Neither extreme is necessary if one focuses on preemption of laws of nature and natural phenomena as the primary concern.
For example, LabCorp was a relatively simple application of personalized medicine. It involved one biomarker and a reasonably straightforward correlation for treatment. In contrast, most personalized-medicine diagnostics are developed using whole genome expression or sequencing arrays to identify hundreds or even thousands of biomarkers that can be used to diagnose a specific disease state. The machine learning algorithms used to identify these markers do not operate on statistical concepts as simple as linear correlation, which for some of us is complicated enough. Machine learning algorithms employ statistical models to identify different combinations or "patterns" of markers that correlate with a specific disease state. Usually these markers are selected and statistically modeled to compensate for human genetic and environmental variation.
Thus, most personalized-medicine programs are tremendously complex compared even to logistic regression and other simple forms of statistical analysis. They are not simply a reflection of a natural phenomenon; they are an interpretive model of nature. Nor are they analogous to or preemptive of human thought. It would be quite improbable for a physician to be able to sit down with a pen and paper and work out a diagnostic by applying a machine learning algorithm or logistic regression to hundreds of biomarker levels.i
Consider the first person who discovered that human chorionic gonadotropin (hCG) levels indicate pregnancy when they are elevated above a certain level, and assume that the inventor also created a home pregnancy test for measuring the hCG. One could think of this as a simplistic personalized diagnostic device in which a particular nongenetic marker is measured to identify a state of health. The inventor could certainly patent the kit, which would consist of the physical device with its particular components. The inventor should not, however, be allowed to patent the process of testing human urine for an elevated level of hCG and correlating that level with the state of pregnancy. The core of the invention, the fact that hCG above a particular level confirms pregnancy, is a simple reflection of nature rather than an interpretive model. To understand the difference, compare the process of measuring one marker by looking for a simple level to the average personalized invention. The modern personalized invention may utilize hundreds of biomarkers analyzed by means of statistical patterns. Even if the complex process involved in identifying the relevant biomarkers yields a limited number of biomarkers to consider for a relevant patient, that information does not relate directly to anything. For example, with a personalized medicine invention involving only a few biomarkers, each of those markers generally has confidence intervals assigned to indicate the likelihood that the presence or absence of a factor or the particular level of that marker will translate into a particular diagnosis or successful treatment. Once biomarkers are collected from a relevant patient, they must be processed by statistical modeling to determine how the various factors and confidence levels for this particular patient should be interpreted. The complexity and variability of individual humans ensures that a model like this can never be a simple reflection of nature. Rather, it is no more than an interpretation of nature, albeit one that is extremely important in the treatment of a specific disease.
In contrast, the inventor of the method of measuring hCG is looking at only one marker, and it is a marker that is elevated in the same range for pregnant women in general. Thus, it requires no complex modeling and is a direct reflection of a phenomenon of nature. As a result, it would not be patentable subject matter on the grounds that the core of the invention is no more than the discovery of a natural phenomenon.
The hCG invention should also fail on the grounds that performing the test is no more than a mental step. The method does not require complex computer analysis to interpret the data; it requires observing a particular level of a substance and reaching a conclusion from that level. Preventing human beings from looking at information and concluding something threatens to preempt simple thought.
Personalized medicine, with its marriage of biology and computer technology, provides a wonderful opportunity to understand and tease out some of the threads of patentable subject matter. These inventions demonstrate how early misconceptions about the nature of computer programs and the nature of mathematics are causing problems in modern case law.ii
Part of the difficulty can be traced to confusion between the content of something that is being expressed and the language in which it is expressed. For example, we know that laws of nature are not patentable. Some of these laws are familiar to us in the formulaic language in which we normally see them expressed. Most people, for example, would recognize one of Einstein's laws of physics expressed as E = MC2. One could express that same law in prose, however, rather than formulaic language by explaining the way in which matter and energy are interchangeable.
The choice of language is irrelevant.iii We disallow patenting of E = MC2 not because it is expressed in mathematical form but because it represents one of the building blocks of scientific exploration and endeavor. Patenting that law would preempt scientific exploration by occupying a basic concept.
In addition to laws of physics, other things can be expressed in formulaic language. Expressing something in formulaic language, however, does not mean that what is being expressed is a law of nature. Thus, the fact that computer programs are expressed in a formulaic language that looks somewhat like math to the layperson does not mean that the concepts underlying a particular program are analogous to math, let alone analogous to a law of nature. Two things expressed in the same language or the type of same language are not necessarily analogous. Comic books and the Constitution are both expressed in the English Language, but they are hardly analogous. Our focus should remain on the content of what is being expressed and on the preemptive effect that might result from patenting that type of content.
Some unfortunate language in the Supreme Court's Diehr opinion, which flows from this misperception, makes it more difficult to properly separate software claims that should be patentable from ones that should not. In searching for logic to explain why the rubber-molding invention at issue was patentable while other apparently similar inventions had failed, the Court made the following comment: "A mathematical formula as such is not accorded the protection of our patent laws, and this principle cannot be circumvented by attempting to limit the use of the formula to a particular technological environment."iv
Other courts have jumped on the language to declare that so-called field of use restrictions cannot save software patents.v Even the recent Supreme Court opinion in Bilski wandered into the same territory when referencing its earlier decision in Flook:
Flook established that limiting an abstract idea to one field of use . . . did not make the concept patentable. That is exactly what the remaining claims in petitioners' application do. These claims attempt to patent the use of the abstract idea of hedging risk in the energy market and then instruct the use of well-known random analysis techniques to help establish some of the inputs into the equation. Indeed, these claims add even less to the underlying abstract principle than the invention in Flook did, for the Flook invention was at least directed to the narrower domain of signaling dangers in operating a catalytic converter.vi
It is true, as a general matter, that if one were to claim a law of nature, the claim would not be rendered patentable by limiting use of that law of nature to a particular field. Thus, for example, an inventor could not save a claim to all uses of E = MC2 by limiting the claim to "all uses of E = MC2 in the construction field."
Sliding the analogy over to software, however, involves logical errors. The logic that appears to have been used is the following. Computer programs are mathematical formulas, and mathematical formulas are laws of nature. We know that laws of nature cannot be rendered patentable by limiting their application to a particular field of use. Computer programs, therefore, cannot be rendered patentable by limiting their application to a particular field of use.
This sequence contains a number of logical errors. Computer programs may be expressed in a language that looks like math, and some do involve calculations, but they are not necessarily analogous to mathematical formulas. Most important, they are not analogous to natural laws just because both are expressed in formulaic languages. One must look to the content of the computer program and its potentially preemptive effect to determine patentability.
As described earlier, the term "algorithm" in computer science means a series of steps performed on input data by a computer. This process may or may not raise preemption concerns. Some computer "algorithms" are based on properties inherent in types of input and output data. Such broad, generic algorithms, which can be used on a variety of types of input data, may raise threats of preemption. In other words, if an inventor asks for a patent on a software program that works with whole sets of numbers or entire types of data, such a patent would not be patentable subject matter. Particularly in light of the bargaining potential that would come with such a grant, the patent would risk tying up entire types of data rather than constituting something applied.
This does not mean that all software is unpatentable. Claims to programs that are applied to a specific type of data in the pursuit of particular types of outputs do not present the same level of preemption threat. For example, a personalized medicine algorithm (i.e., series of steps) that employs a specific type of statistical model using a fixed set of markers to produce a very specific diagnosis would not threaten to preempt other methods of performing the same diagnosis that use different markers or novel types of statistical models. Such an invention should be patentable.
Computer programs may be many things, including methods of creating useful models of the world around us, methods of providing interpretations of information, and methods of sorting information. When methods of creating a particular type of model are described at a very general level, they may threaten to preempt the broad activity of exploration. However, when claimed at the level of a specific method of sorting a particular type of information for a particular pursuit, they should constitute an applied invention. Such specificity is the hallmark of what separates unpatentable abstractions from applications of those abstractions in the useful arts in a way that is worthy of patent protection.
* * * * *
i One scholar has suggested that even LabCorp could have survived as patentable subject matter if the claims drafter had chosen a narrower description. See Patricia Dyck, "Post-Bilski Personalized Medicine: At Home on the Range" (forthcoming) (manuscript on file with author) (presenting an empirical survey of the most frequently claimed biomarkers in personalized medicine and discussing problems in allowing claims for sets of biomarkers that are small and not limited to the diagnosis of a specific disease state or condition).
ii I have discussed some of the misconceptions described here in the context of patenting genes. See Feldman, "Whose Body Is It Anyway?" 1400–1402.
iii One could easily argue that all math is invented. It is a human-made method of imposing order and structure on the natural world. For an interesting and accessible discussion of Wittgenstein's view that all mathematics is a human invention and various responses to that argument, see Ludwig Wittgenstein, Wittgenstein's Philosophy of Mathematics §3.1, http://plato.stanford.edu/entries/wittgenstein-mathematics.
iv Diamond v. Diehr, p. 191 (citation omitted).
v See, e.g., In re Bilski, p. 957 (noting that mere field-of-use limitations are generally insufficient to render an otherwise ineligible process claim patent eligible).
vi Bilski v. Kappos, p. 3231.