Ex Parte Jones

Board of Patent Appeals and Interferences

Aug 28, 2012

11217205 (B.P.A.I. Aug. 28, 2012)

UNITED STATES PATENT AND TRADEMARK OFFICE
__________

BEFORE THE BOARD OF PATENT APPEALS
AND INTERFERENCES
__________

Ex parte JOHN E. JONES III
__________

Appeal 2011-007404
Application 11/217,205
Technology Center 1600
__________

Before DEMETRA J. MILLS, ERIC GRIMES, and MELANIE L.
McCOLLUM, Administrative Patent Judges.

GRIMES, Administrative Patent Judge.

DECISION ON APPEAL
This is an appeal under 35 U.S.C. § 134 involving claims to an
automated method of sorting plant embryos. The Examiner has rejected the
claims as obvious. We have jurisdiction under 35 U.S.C. § 6(b). We
reverse.
STATEMENT OF THE CASE
“Reproduction of selected plant varieties by tissue culture has been a
commercial success for many years” (Spec. 1: 13-14) but “[o]ne of the more
Appeal 2011-007404
Application 11/217,205

2
labor intensive and subjective steps in the embryogenesis procedure is the
selective harvesting from the development medium of individual embryos
suitable for germination” (id. at 2: 23-25). “A skilled technician evaluates
the morphological features of each embryo . . . and selects those embryos
that exhibit desirable morphological characteristics. This is a highly skilled
yet tedious job that is time consuming and expensive.” (Id. at 2: 28 to 3: 1.)
The Specification discloses “classification of plant embryos by the
application of classification algorithms to digitized images and/or data
relating to or based on the absorption, transmittance, reflectance, or
excitation spectra of the embryos” (id. at 4: 2-4). The classification model is
developed using penalized logistic regression analysis (id. at 5: 1-5). “The
most common application of logistic regression is for two classes” (id. at
10: 4-5) and “in one embodiment, a method of the present invention may be
used to classify embryos into generally two classes of relatively high-
quality, acceptable embryos and relatively low-quality, unacceptable
embryos” (id. at 11: 5-7).
Claims 1-7, 14, and 16-20 are on appeal. Claims 1-7 are directed to a
computer-based method of classifying plant embryos based on their
germination potential. Claim 1 is the only independent method claim, and
requires acquiring sets of images or spectral data from embryos of known
germination potential; the sets are each then associated with a class label
(“e.g., high quality [or] low quality,” Spec. 5: 11-12) and a set of metrics is
calculated based on the sets. The claimed method then requires, “using the
computer, applying a penalized logistic regression (PLR) analysis to the sets
of metrics and their corresponding class labels to develop a PLR-based
Appeal 2011-007404
Application 11/217,205

3
classification model” (claim 1, step (d)). The classification model is then
applied to sets of images or spectral data from embryos of unknown
germination potential, and the computer outputs the results in a specified
format. The complete text of claim 1 is reproduced in the Claims Appendix
of the Appeal Brief.
Claims 14 and 16-20 are directed to a computer including computer-
executable instructions that perform steps corresponding to the steps of
claim 1, including “applying a penalized logistic regression (PLR) analysis
to . . . sets of metrics and their corresponding class labels to develop a PLR-
based classification model” (claim 14, step (b)).
The Examiner has rejected all of the claims on appeal under 35 U.S.C.
§ 103(a) as obvious based on Timmis,
1
Dettling,
2
Holmes,
3
Pearce,
4
and
Dorling
5
(Answer 4). The Examiner finds that Timmis discloses an embryo-
classifying method that includes most of the steps of claim 1 (id. at 5-6) but
“does not show a statistical method that comprises penalized logistic[ ]
regression” (id. at 6).

1
Timmis et al., WO 99/63057, published December 9, 1999.
2
Dettling et al., Supervised Gene Clustering with Penalized Logistic
Regression, Seminar für Statistik, Eidgenössiche Technische Hochschule,
Research Report No. 115 (2003).
3
Holmes et al., A Representation for Accuracy-Based Assessment of
Classifier System Prediction Performance, in Lanzi et al. (Eds.), IWLCS
2001, LNAI 2321, pp. 43-56 (2002).
4
Pearce et al., Evaluating the predictive performance of habitat models
developed using logistic regression, 133 Ecological Modelling 224-245
(2000).
5
Dorling et al., Maximum likelihood cost functions for neural network
models of air quality data, 37 Atmospheric Environment 3435-3443 (2003).
Appeal 2011-007404
Application 11/217,205

4
The Examiner finds that “Timmis emphasizes that data analysis
techniques can be put together in an almost infinite number of combinations
to achieve the desired results” (id.) and that “Dettling shows statistical
analysis using penalized logistic regression produces superior results
compared to other statistical methods” (id.). The Examiner concludes that it
would have been obvious “to modify the method of classifying plant
embryos of Timmis et al. with the penalized logistic[ ] regression of Dettling
et al. because Dettling et al. shows statistical analysis using penalized
logistic regression produces superior results compared to other statistical
methods” (id. at 8).
6

Appellant argues, among other things, that “there is no apparent
reason to combine[ ] Dettling et al. with Timmis et al. to arrive at the
claimed invention” (Appeal Br. 13). Appellant argues that Dettling
describes its penalized logistic regression algorithm, which it calls Pelora, as
useful for “data that are subject to the „large p, small n‟ problem” (id. at 14);
i.e., “where there are relatively few number of samples, n, (i.e.,
observations) and p is the dimension of the data” (id.). Appellant argues that
“Timmis et al. does not fit the data pattern of the „large p, small n‟ problem
of Dettling et al. and so would not have commended itself to a skilled
artisan‟s attention” (id.).
We agree with Appellant that the Examiner has not persuasively
shown that a skilled worker who was familiar with the disclosures of

6
The Examiner relies on Holmes, Pearce, and Dorling to establish the
obviousness of the specific type of output required by claim 1. See Answer
6-8.
Appeal 2011-007404
Application 11/217,205

5
Timmis and Dettling would have considered it obvious to modify Timmis‟
method by using the penalized logistic regression analysis disclosed by
Dettling.
Timmis discloses “classification of plant embryos by the application
of classification algorithms” to images or spectral data (Timmis 3: 28-30).
Timmis discloses several types of algorithms that can be used to develop a
classification model (see id. at 8: 17 to 9: 7), including “Logistic and Probit
Regression” (id. at 9: 7) but does not specifically suggest using penalized
logistic regression. Timmis states that
[t]here are many data analysis methods that can be applied to
develop and use classification models that allow plant embryos
to be classified by quality. The above described mathematical
methods are a sampling of some of the major techniques.
However, it should be emphasized that data analysis techniques
can be put together in an almost infinite number of
combinations to achieve the desired results.
(Id. at 13: 9-13.)
Dettling states that “[m]icroarray experiments generate large datasets
with expression values for thousands of genes, but not more than a few
dozens of samples” (Dettling, abstract). Dettling states that
[a]n important task is to reveal groups of genes which act
together, for example in pathways, and whose collective
expression is optimally predictive for a certain response
variable y. . . . However, finding the groups is difficult: we are
facing computational problems due to the sheer amount of
predictor variables (genes), and statistical difficulties due to the
“small n, large p” phenomenon.
(Id. at 1.)
Dettling discloses “Pelora, a novel approach to supervised clustering,
using an objective function based on penalized logistic regression analysis”
Appeal 2011-007404
Application 11/217,205

6
(id. at 2). Dettling applies the Pelora algorithm to gene expression datasets
from several types of cancer, and compares those results to the results using
other algorithms (id. at 14). Dettling concludes that “Pelora is best, before
Wilma, the support vector machine, diagonal linear discriminant analysis and
the nearest-neighbor rule. Pelora is the best classifier on estrogen, nodal
and prostate data, and second best on colon and lymphoma data. It performs
worse on leukemia data, however.” (Id.) Dettling states that Pelora
“group[s] thousands of genes into a few small gene clusters, that are very
informative for explaining the outcome y” (id. at 14-15).
Dettling describes Pelora as a “methodology for supervised clustering
of genes from microarray experiments which is potentially beneficial in
medical diagnostics and prognostics, as it identifies clusters of interacting
genes whose expression centroids have high explanatory power for the
response variable” (id. at 16). Dettling states that “[a]lthough Pelora was
specifically developed for the analysis of microarray data, it may be useful
for other data that are subject to the „large p, small n‟ problem and where a
few underlying clusters are expected to determine most of the outcome
variation” (id. at 17).
We agree with Appellant that Dettling describes its penalized logistic
regression analysis as being suited to analyzing a specific type of data,
similar to the data generated by a microarray experiment, in which a
relatively small number of samples (“small n”) each provides a large number
of data points (“large p”). The Examiner has not provided persuasive
evidence or technical reasoning to show that classifying images or spectral
data derived from plant embryos, as taught by Timmis, has characteristics
Appeal 2011-007404
Application 11/217,205

7
similar to those described by Dettling for the process of identifying clusters
of coexpressed genes in microarray data, which the Pelora algorithm was
designed to address.
Thus, although Timmis states that data analysis methods could be
used in “an almost infinite number of combinations” (Timmis 13: 9-13) to
develop classification models for plant embryo image or spectral data, the
Examiner has not pointed to evidence that a penalized logistic regression
analysis, specifically, would have been an obvious choice to a person of
ordinary skill in the art, based on Timmis and Dettling.
In the Answer, the Examiner argues that “[t]he passage of Dettling et
al. that appellant refers [to] is a suggestion to apply the PELORA algorithm
to other problems such as the “small n, large p” or problems where a few
underlying clusters are expected to determine most of the outcome variation.
Timmis et al. fit the second criterion.” (Answer 14, emphasis added.)
The passage referred to in Dettling, however, refers to problems that
meet both of the criteria cited by the Examiner, not either one alone. See
Dettling 17: “Pelora . . . may be useful for other data that are subject to the
„large p, small n‟ problem and where a few underlying clusters are expected
to determine most of the outcome variation” (emphasis added). Compare to
Dettling, abstract: “Microarray experiments generate large datasets with
expression values for thousands of genes [“large p”], but not more than a
few dozens of samples [“small n”]. A challenging task with these data is to
reveal groups of co-regulated genes [“a few underlying clusters”] whose
collective expression is strongly correlated with [“are expected to determine
most of”] an outcome variable of interest [“the outcome variation”].” The
Appeal 2011-007404
Application 11/217,205

8
Examiner‟s apparent position that Dettling suggested using Pelora for
problems that did not present a “large p, small n” problem is therefore not
supported by the reference.
The Examiner also argues that “Timmis et al. shows that highly
correlated metrics may be combined to reduce the total number of metrics
(p. 22, lines 10-17). This type of reduction in metrics is what Dettling is
describing in the Pelora algorithm (p. 6, section 2.4.2).” (Answer 14.)
This reasoning is also unpersuasive. First, the Examiner has provided
no reasoned explanation of his position that the Pelora algorithm amounts to
a method of combining highly correlated metrics, and no basis for that
equivalence is apparent to us based on the cited section of Dettling. Second,
even assuming the Examiner‟s statement is accurate, the Examiner has not
explained why it would have been obvious to use an algorithm that reduces
the number of metrics by combining highly correlated metrics as a basis for
developing a classification model that classifies plant embryos based on
image or spectral data.
SUMMARY
We reverse the rejection of claims 1-7, 14, and 16-20 under 35 U.S.C.
§ 103(a) because the Examiner has not persuasively shown that it would
have been obvious to practice Timmis‟ embryo-classifying method using the
penalized logistic regression analysis disclosed by Dettling.

REVERSED

alw