!20151202  Drug discovery tools and KSR: the patentability risk worsens 

---    Drug discovery tools and KSR: the patentability risk worsens

      In my ongoing theme of future drug patentability being threatened by the overlap of drug discovery tools, and KSR's predictions of PHOSITAs using those tools, a recent paper published at arXiv.org illustrates this growing risk and threat to drug patentability. The article is:
Machine learning, quantum mechanics and chemical compound space
Raghunathan Ramakrishnan, O. Anatole von Lilienfeld
Department of Chemistry, University of Basel
http://arxiv.org/abs/1510.07512.
What follows are some quotes from the article. As you read, keep in mind the poorly worded, scientifically naive, unconstitutionally supported, "predictable" language of KSR. It ain't pretty.


[Conclusion, page 15, right hand column]: As such, while an atomic machine learning model will not be able to extrapolate to entirely new chemical environments that were not part of its training set, it is well capable of being used for the [KSR] prediction of an arbitrarily large diversity of macro- and supra-molecular structures, such as polymers or complex crystal phases - as long as the local chemical environments are similar.

Greg note: so you want to bust a patent on a new drug in a known class/field? Find an existing chemical environment for that drug in that class/field, and show how this machine learning program in the hands of a PHOSTIA could [KSR] predict the new drug, and thus it is not patentable under KSR.

[Problem to be solved, page 2, left hand column]: Despite a long tradition of ML methods in pharmaceutical applications [16–20], and many success stories for using them as filters in order to rapidly condense large molecular libraries [21], their overall usefulness for molecular design has been questioned [22]. In our opinion, these models fail to systematically generalize to larger and more diverse sets of molecules, compounds, or properties because they lack a rigorous link to the underlying laws of quantum mechanics and statistical mechanics.

Greg note: so even the naive methods have proven useful in automatically predicting some drugs. But couple the naive methods with the powers of quantum and statistical mechanics, and the predictions become more useful. Less patentable?

[Range of predictions]: ML models trained across chemical space are significantly less common in physical chemistry. Notable exceptions, apart from the papers we review here-within, include attempts to use ML for the modeling of chemical reactivity [35], melting points [36], solubility [37], enthalpies of formation [38], density functionals [39], basis-set effects [40], polymer properties [41], crystal properties [42, 43], or frontier orbital eigenvalues [44].

Greg note: All of these features are routine claim elements. Increasingly being predicted. Less KSR-patentable?

[Page 2, last paragraph]: Here, we refer to Machine Learning as an inductive supervised-learning approach that does not require any a priori knowledge [FROM A PHOSITA] about the functional relationship that is to be modeled, and that improves as more training data is being added.

Note: the paper goes on for 12 pages of explanation of the computational and quantum mechanical techniques to make these predictions much more feasible.

[Conclusion, page 14]: Comparing performance for molecules, nano-ribbon models, atoms, or crystals, we typically observe satisfying semi-quantitative accuracy for machine learning models trained on at least one thousand training instances. Accuracies competitive with current state-of-the-art, i.e. similar to the experimental accuracy, are typically found only after training on many thousands of examples.

[Conclusion, page 15]: Future work will show if machine learning models can also be constructed for other challenging many-body problems including vibrational frequencies, reaction barriers and rate constants, hyperpolarizabilities, solvation, free energy surfaces, etc. Open methodological questions include optimal representations and kernel functions, as well as the selection bias present in most (if not all) databases.


      There are two trends, as I discussed at a drug discovery conference one year ago. The first trend is the decreasing coherence of 103 obviousness caselaw and the decreasing compatibility of 103 caselaw with the scientific process. The second trend is the increasing power of these automated drug discovery tools to predict new, useful drugs - with the typical skilled guidance of a PHOSITA using a tool.

      At some point, these two curves intersect, and then drug patenting becomes obvious as most new drugs are viewed by the courts as the obvious predicted products of PHOSITAs using their [drug discovery tools]. And this intersection could happen in the next 20 years, which overlaps with the timeline of all new drug patents being filed now (many of which are susceptible to this KSR attack with today's drug discovery tools).

      It is surprising then how little is being written about this coming problem in drug IP circles, because it puts the entire pharmaceutical business model at risk. I can see two new insurance policies: one protecting drug companies from 103 attacks at the PTO (the policy helping to challenge the idiocy of 103 caselaw) and a second policy protecting generic companies from all future lawsuits (the policy using the idiocy of 103 caselaw and power of drug discovery tools to invalidate). Fun! Fun! Fun! with all of this risk.



IPISC - your patent insurance experts:

-- Need protection from frivolous patent infringement lawsuits?
-- Need protection from AIA attacks? From troll attacks? From IPR-troll attacks?
-- Need help asserting/leveraging your patent(s) without being forced to sell out to trolls and "monetizers"?
-- Need to borrow funding using your IP portfolio as collateral?

Call IPISC to learn about how these cost-effective, risk-minimizing, insurance-based protections are available for your quality patents in U.S., Canada and Europe.
IPISC - http://www.patentinsurance.com , 800-537-7863.
This PATNEWS sponsored by Intellectual Property Insurance Services Corp.




Greg Aharonian
Internet Patent News Service
Your "Judicial Counter-Errorism Expert"
In U.S. – 415-981-0441
For a comprehensive collection of patent quality statistics, see:
http://www.global-patent-quality.com/statistics.html
For a brochure on my patent search services, see:
http://www.global-patent-quality.com/DOCUMENTS/Brochure.pdf