Azimuth 2.0 release

Changelog:

1. added random_state to GBRT call to ensure reproducibility of experiments.
2. fixed continuous integration tests (travis-ci).
3. increased number of guides tested in unittest to 1000 (up from 3).
4. enforcing dependency on scikit-learn >= 0.17 because earlier versions produced different results due to stochasticity.
5. re-generated the model pickle files to fix a bug introduced in June 2016 (340ab04), when we updated the feature names without re-generating the pickle files.

This release has been pushed to pypi, but has not been pushed to the Azure ML server. Thus, there are differences in prediction values between the web service and the Python code (see #1).

Nature Biotech. 2016 release

This code corresponds to the experiments, results and models that accompany our 2016 Nat. Biotech paper: doi:10.1038/nbt.3437.

After publication, a small indexing bug (#1) was found that had no substantial effect on the model or reported results (all Spearman correlations reported in the paper remain the same), but has been corrected in future versions to avoid confusion. This bug was fixed starting in commit 8d11972.

On 6/11/2016 we became aware that there are small differences in the exact values between the bug-fixed version and the original version even though the Spearman correlations (and therefore quality of the model) remain the same.