1. Data Mining Consultant GlaxoSmithKline: US Pharma IT
Integrating Discovery, Development, and Commercial Data into Data Mining
Jennifer Sloan
Data Mining Consultant
GlaxoSmithKline: US Pharma IT
15 September 2004

2. Data Mining Definition
Data Mining is a process that uses a variety of data analysis tools to discover patterns and relationships in data that may be used to make valid and accurate predictions.

3. Data Mining is a tool that allows us to
Identify problematic areas
Control process variability
Make concrete decisions on business needs
Develop a model which can aid in future business decisions

4. Analyzing Multivariate Data Managing Data Usage Model Building
Commercial Data
Analyzing Multivariate Data
Managing Data Usage
Model Building

5. Multivariate Data Sets
Data are multivariate in nature
Large data sets containing multiple criteria within each observation
Comparing multiple vectors is nearly impossible without reducing to a single point

6. Here we view 5-dimensional information on one observation
Here we view 5-dimensional information on one observation. Each point represents a prescriber and the color represents a Market Share increase or decrease. Overlapping distributions make this difficult to interpret and further analysis is required. Over 200K observations are represented in this graph.

7. The same observations are observed but now two-way interactions between the variables help us determine which variables are affecting market shifts and lead to constructing models which will predict prescriber behavior.

8. Drug Development

9. Drug Development Issues
Adverse Event Reporting System (AERS)
Over 2 million AE reports and approximately 2000 drugs and biologics submitted to the FDA since 1968
Creates Extremely Complicated Matrix of Data
Recently, Data Mining methods have helped address this issue with the development of a method used to examine large databases for associations between drugs and AEs

10. Data Mining Algorithm
Multi-Item Gamma Poisson Shrinker (MGPS) Developed by William DuMochel (AT&T)
Through statistical modeling, this Empirical Bayesian method identifies higher-than-expected reporting relationships of drug-event combinations
Automated, web-based system with rapid drill-down capability
MGPS runs using all event terms and drugs in the AERS database and produces results for all drug-event combinations

11. MGPS: Significance Handles Complex Stratification
(age, gender, year of report > 945 categories)
Performs complex computations in minimal amount of time: Much MORE EFFICIENT
Real World Example:

13. Membership: PhRMA-FDA Working Group
Chair: June Almenoff (GSK)
FDA Involvement
Involved PhRMA companies: Abbott, Allergan, AstraZeneca, Bristol-Myers Squibb, GlaxoSmithKline, Johnson & Johnson, Lilly, Merck, Novartis, Schering-Plough, Pfizer, Roche, Wyeth

14. Drug Discovery

15. SCAM—Statistical Classification of Activities of Molecules
Recursive partitioning customized for chemistry
Creates a structure activity relationship (SAR) mode7l
Handles large numbers of descriptors (> 1 million)

16. SCAM : Data Structure
Biological
Activities Y1 Y2 Y3 Y4
...
... Yn
> 2 million
>100K

17. SCAM’s Recursive Partitioning
Ave = 0.34
SD = 0.81
Feature
n = 1614
ave = 0.29
sd = 0.73
n = 36
ave = 2.60
sd = 0.9
rP = 2.03E-70
aP = 1.30E-66
AID, Automatic Interaction Detection, is a simple statistical method for finding interactions in large data sets. It is recursive. One variable and a split point is selected to divide the data set into two or more groups. Each group is divided in turn. When the groups become too small or are homogeneous, the splitting is stopped. A tree results and the sub-grouping rules are read off the tree. NB: It is possible to separate off different groups that are following different mechanisms. There are serious multiple testing questions. Also at issue is the statistic to use for splitting.
Signal
t = = = Noise
1 1 +

18. SCAM Tree

19. Advantages of SCAM
Works for complex situations, mixtures and interactions.
Output is easy to understand and explain
High statistical power
Produces a valid answer

20. SCAM Drawbacks Data greedy Only one view of the data
Binary descriptors may be too “crude”
Disposition of outliers is difficult
Highly correlated variables may be obscured
Higher order interactions may be masked
The are of course disadvantages to any algorithmic method used to analyze large data sets.

21. Concluding Remarks
Data Mining enables us to efficiently handle LARGE amounts of data
Data Mining allows us to perform analyses IN REAL TIME
Data Mining covers a wide array of topics in drug industry and its benefits are plentiful

22. References
Almenoff, June S, et al. “Disproportionality Analysis Using Empirical Bayes Data Mining: A tool for the Evaluation of Drug Interactions in the Post-Marketing Setting.” Pharmacoepidemiology and Drug Safety,12, (2003).
Donahue, Rafe. “An Overview of Data Mining in Drug Development and Marketing.” May 2003.
Hawkins, D.M. and G.V. Kass, “Automatic Interaction Detection.” Topics in Applied Multivariate Analysis, ed. Hawkins, (1982).
Hawkins, D.M., S.S. Young and A. Rusinko. “Analysis of a Large Structure-Activity Data Set Using Recursive Partitioning.” QSAR, 16, (1997).