1. Using Open Data and Tools to Predict Which cancer Patents Would Be FDA Approved？
The Institute of Scientific and Technical Information of China
Yang Guan-Can Oct 26, 2016

2. 1. Predicting Successful FDA Approval
Patent grant and FDA approval are both important to a pharmaceutical production enterprises.
approval processes expensive and time-consuming, uncertainty.
Inability to predict which patents could be FDA approved.
So triaging potential patents based on their likelihood of accepted FDA approval would be important.
Patent grant and FDA approval are both important to a pharmaceutical production enterprises. Patents are crucial to protect a company’s ideas while FDA approval is necessary to legally market their products. However the PTO and FDA approval processes are expensive and time-consuming, with full of uncertainty.
Till now, we cannot clearly identify which granted patents would continue to apply a FDA approval. Which slow down the speed of patent prosecution and increase the risk and cost of enterprises.
So triaging those potential product-ready patents based on their likelihood of FDA approval would be very important.

3. 2. Opportunities of Open Data & Tools
It is really not easy to solve this problem, because the path from patent grant to FDA approval is complicated and uncertain, and accompanied with enterprises’ multiple strategies, such as “Metabolite Defense”.
Fortunately, more and more institutions and individuals joined the open data and tools list recently, providing us with excellent resource, which is new opportunity for us exploring to resolve the tricky questions. PatentsView just is an example, and it provided API interface and bulk downloads based on further data processing and inventor disambiguation, Recently, USPTO has provided a cancer moonshot patent data, which linked NIH funding data and FDA approval data, which externally expands the dimensions of our predictive analytics. At the same time, some commercial vendors contributes some free tools, just like KNIME, RSTUDIO, with rich machine learning algorithms and interactive visualization functions. So it is time for us to integrate all these tools helping innovators have more in-depth insights on the messy world.

4. 3. Predictive Workflow on Patent2FDA
KNIME analytic platform is a free tool;
Supporting large data; integrating various projects, e.g., text mining, deep learning.
We build the whole predictive project on KNIME
CRISP-DM rule

5. 4. Results and Model Evaluation
Comparing several models and the performance of Random Forest is best;
However, the result is only “forward looking statements”, need more validation.
Model Name
Accuracy
Cohen's Kappa
Area Under Curve
Random Forest
0.9624
Multi_layer Perceptron
0.9
0.8
0.8041
Decision Tree
0.933
Naive Bayes
0.9363
Logistical Regression
0.9510
Gradient Boosted Trees
0.9408

6. 5. Model Interpretation
The applicant’s historical experience on FDA approval
Similarly, the assignee’s experience is also important.
Lawyer and Agency experiences;
CPC group: A61K and C07K concurrence;
Patent average process ratio
The state of assignee’s

7. 6. Thoughts
The path from patent grant to FDA approval is complicated and uncertain.
Predictive analytics may provide us some insights.
In the future, we wish more and more open data could be added into model, which would facilitate the performance of predictive model, such as text mining, network analysis.

8. Thanks