Immunotherapy treatments such as antibodies targeting cytotoxic T lymphocyte-associated antigen 4 (CTLA-4), programmed cell death protein 1 (PD-1) and PD-1 ligand1 (PD-L1) have shown promise in reactivating weakened immune cells to fight cancer. While these immunotherapies have had a dramatic impact in some cancer patients, the positive results only appear in a fraction of cases. The cost of treatment and potential for immune-related adverse events make it imperative that doctors have protocols to identify patient populations with an increased likelihood of successful outcomes with immunotherapy. This has led to a search for predictive biomarkers that may allow identification of such patients. Some have turned to artificial intelligence (AI) to scour data to identify common biomarkers or other covariates in patients successfully treated with immunotherapy.
AI has the capability of reviewing a staggering amount of patient data to identify recurring patterns of shared predictive factors that would elude unaided human capacity. For example, doctors at the Institute of Cancer Research in London classified types of cancers based upon gene expression profiles collected from biopsies to identify cell types present in, and biological pathways involved with, the tumor. This has assisted in the identification of subtypes of breast cancers that appear more likely to respond to immunotherapies. Other researchers have used AI to review imaging data from pathology slides and radiological scans to identify patterns of tumor progression and interaction with the immune system. AI analysis of CT scans in particular has identified changes in those scans during clinical trials that may provide an indication of the likelihood that immunotherapy may treat patients with lung cancer. Other studies using AI have uncovered biomarkers that identify patients that will have negative results if treated with immunotherapy.
Considering the investment associated with the computational requirements to conduct AI research and the clinical and economic benefits to administering immunotherapies to patients with the greatest potential to respond, institutions will likely seek to obtain patent protection around these discoveries. Immunotherapy and disease targets are already frequently patented (as antibody composition and method of treatment claims), so patenting AI-driven improvements in patient selection and treatment optimization will necessarily focus on narrower claims. Method of treatment claims directed towards groups exhibiting the characteristics that coincide with an increased likelihood of success offer one pathway to patent protection. Such claims may describe the patient population characteristics or the dosing regimen for the specific immunotherapy. AI raises many questions from a patent perspective, including patentable subject matter, inventorship, compliance with 35 U.S.C. § 112(a), and obviousness. For the types of claims described above, inventorship and obviousness seem most pertinent.
Several patent offices have concluded that AI cannot be named as an inventor of a patent application—that remains reserved for humans. This may make it difficult for an applicant to identify the inventor. In instances of narrow application of AI, applications with specific objectives (such as human links to AI, including design of the algorithms, preparation and selection of inputs, and preparation of datasets) are readily apparent. If these inputs result in the identification of the patient population or optimized dosing regimen, the question of inventorship should not present a hurdle. But hurdles may exist in instances where researchers use third-party AI systems, or in the application of artificial general intelligence, e.g., where a machine applies knowledge and information from different contexts as opposed to focusing on a specific task and data set provided by human selection. For the former, it is key for researchers to track how their use is unique and how the researchers selected or manipulated available data sets. For the latter, human intervention may be further removed because the invention may approach or exceed human abilities. Both present thornier inventorship questions that should be addressed from the outset of the project by drafting clear contracts in the case of third-party AI systems and by tracking all human contribution to the development of the AI.
The existing standard for assessing obviousness asks whether a person of ordinary skill in the art would be motivated to combine the prior art with a responsible expectation of success. The concept of the skilled artisan may be affected by the capabilities of AI because AI presents a tool to the skilled artisan that may exceed the unaided abilities of that person. The ability of AI to analyze large amounts of data to identify patterns may be a routine task with an expected outcome from an AI perspective, while the identification remains far above what one would expect the skilled artisan to achieve. Considering that AI seeking to identify patient populations or optimized treatment with immunotherapy will rely upon correlations extracted from existing patient data, if AI becomes ubiquitous in assessing immunotherapies, inventors may confront a heightened barrier to protecting the discovered methods or optimizations.