Machine learning: new discoveries and challenges for life sciences
Machine learning originally sat on the high-tech side of the fence. Now, it has migrated into the life sciences realm of research.
Machine learning is a computer-driven strategy that uses pattern recognition to enable a learning process.
Generally, a series of data is fed in to the algorithm to train the computer. The computer, using what it has learned from the patterns it recognises in the data, makes new correlations and decisions on its own when new data is supplied.
Life sciences-related uses for machine learning include drug discovery, methods for disease diagnoses and pattern recognition in patient records.
For example, researchers have used machine learning to train computer software to recognise pathology characteristics (cell size, shape, texture, and spatial relationships between cells) to predict survival outcomes for lung cancer patients.
Machine learning has also been applied to patterns of drug side-effects and chemical structure, using the software to rapidly screen and cull the number of possibilities to put through wet-lab toxicity testing.
Drug discovery, such as predicting target structures and optimisation of hits from chemical screens, is an additional area for the application of machine learning.
The intimate involvement of the software in carrying out the methods, as well as generating the results, raises some intriguing questions for IP protection.
A stumbling block
One of the stumbling blocks in software patents has been subject matter eligibility.
Boiled down to its simplest argument, if the method is fairly broad, it may be relegated to the “abstract idea” category, and the addition of a computer to carry out the steps does not provide enough to get it past the subject matter eligibility test.
The addition of software to a method may not be sufficient to provide novelty and non-obviousness unless specific algorithms or other method steps make it distinguishable from the prior conventional processes.
For example, if comparing specific side-effects of certain drug structures is already a known process, the addition of a computer to improve the efficiency and the volume of what can be screened may be insufficient for patentability.
Patentability rejections may occur where the addition of a computer to the method does not include details that narrow the method to a specific algorithm or otherwise provide specifics to distinguish it from prior methods.
A number of life sciences-related patent applications incorporating machine learning have met with subject matter eligibility (section 101) rejections.
Applications have been rejected on the basis that the methods of machine learning are abstract, and for the reason that the biological predictions are not “real world” results that are communicated to a practitioner.
Examples include:
- The US Patent and Trademark Office (USPTO) described a submitted drug screening method claim as a combination of “obtaining data” coupled with known steps for testing the identified pharmaceutical compounds.
The method claims were rejected because under the established two-part test for subject matter eligibility, the addition of routine steps to an abstract idea did not provide the “something more” to pull the claim into the realm of patentability.
- The USPTO deemed a method for determining whether a molecule has a biological effect based on the use of training data to generate a prediction model and thereby generating a prediction to be too abstract to qualify as patent eligible subject matter. Amended claims incorporating the use of a computer program passed muster for section 101, but remained rejected for obviousness in view of methods already disclosed in the prior art.
The USPTO finally allowed the claims after the methods were limited to a machine learning method that required more than three molecular property models with the screened biological properties to be selected from antibacterial, antiviral and anticancer effects.
- One applicant successfully overcame a section 101 rejection with the addition of steps that more specifically identified the steps of the machine learning process.
The claims specified the use of an analyser of a processor to identify a first target of the drug, analyse and identify targets interconnected with that target and select a second target having the highest score associated with those interrconnections.
The applicant noted that these steps carried out by the analyser using the scoring system were the “something more” that separated the method from a more general abstract idea and that the steps were an improvement over prior methods for predicting adverse side-effects for drugs.
The integration of machine learning may provide a new hook for patentability by highlighting the ability of the software’s learning process to generate outcomes that go beyond prior methods and the predictable correlations of what could be done previously.
Only patents?
Patents are not the only vehicle for protecting inventions in this realm, and in some cases, may not be the best vehicle in view of the patentability challenges.
One potential issue with patents is the dilemma of disclosure. A patent application is generally published 18 months after filing, yet the applicant may have to settle for more narrow patent claims to address patentability rejections. Thus, competitors can get a heads-up on the idea, but then design-around the narrower issued claims and avoid infringement.
Incorporating trade secrets as an avenue of protection may help address these issues but it can have its own downsides.
Trade secrets protection requires measures to continually protect the information and vigilance in keeping the information confidential.
Misappropriation of trade secrets can also be difficult to prove when down the road a competitor appears to be using the same methods, particularly in an internal research and development setting.
A blended strategy of patents and trade secrets can provide a balance—for example, the methods for discovery may be left as trade secrets, and the outputs, such as drugs, targets, and identified bio-markers for diagnoses, can be protected by patents.
Who is the inventor?
A recent report of machine learning in the driverless car industry, “The Dark Secret at the Heart of AI” (published on the MIT Technology Reviewwebsite), pointed to the autonomy of the software in decision-making.
The choices selected by the software through the machine learning process had become sufficiently complex that the logic behind the choices was no longer obvious from the underlying algorithm.
The ability of the software now to do its own “thinking” could complicate the inventorship analysis.
The USPTO defines the conception of an invention as “a definite and permanent idea of the complete and operable invention”.
If machine learning is used as a drug discovery platform, and its decisions become autonomous in a similar manner to that described for the driverless car, when does conception occur, and who has made that conception?
Is it the author of the underlying algorithm? The individuals who came up with the concept of using machine learning to get to the drug candidate? What about the individuals who selected the dataset on which to train the software?
The identification of inventors is not just a mechanism to give individuals credit for the discovery of the invention.
Inventorship is often linked to ownership, based on the named inventors’ obligations to assign their IP to employers or to others by contractual terms.
If the creator of the algorithm or the software is one of the inventors and not an employee or contractually obligated to assign IP to the company, this could affect a company’s rights to the invention.
The company’s rights to use, license and enforce patent rights to the methods used by the software, as well as the pharmaceuticals, targets or other outputs generated, may be impacted.
What about the machine learning software itself, could it be named as an inventor?
In conventional methods, it has been generally held that the creator of a method for discovering a drug (eg, identifying a target gene and proposing a chemical screen to identify drugs that affect the target) does not on its own convey rights to the drugs that are ultimately identified with the method.
However, machine learning applications go a step further. They not only provide the screen, but also identify the ultimate drug candidates.
“The ability of the software now to do its own ‘thinking’ could complicate the inventorship analysis.”
Applying the driverless car autonomy example to a drug discovery scenario, if the machine learning software is at the point that it is selecting chemical structures for drug candidates with a logic “all its own”, such that the creators of the software themselves could not deduce the underlying selection process, would that leave the software as the one who first conceived of the “definite and permanent idea of the complete and operable invention” and thus the true inventor?
And if so, who would have ownership of the identified drug candidates? Would it be the company running the software program for the drug selection, or the providers of the software, or is there no clear assignment?
Particularly in collaborations with software providers, life sciences companies should be mindful of the difficult issues that might arise as to who owns the key discoveries and resulting products, and address the IP assignment issues early in the process.
Erica Pascal is a partner at DLA Piper. She can be contacted at: erica.pascal@dlapiper.com
