8 April 2019

Artificial Intelligence and Drug Discovery

Artificial Intelligence (AI) may have some interesting consequences for the patentability of new drug substances. Virtual screening techniques, in which a computer model attempts to identify compounds which are likely to bind to a particular biological target, have for many years been part of the drug discovery process. The amount of time, money and expertise required in order to take the output of a virtual screen and provide a drug with marketing authorisation is prodigious. The extent to which AI will in future be able to reduce the considerable gap between current virtual screening techniques and a marketable pharmaceutical product is unclear, but with many big pharma companies investing in AI it seems likely that powerful new in silico models will have a role to play in shaping the drug discovery landscape.

One trend that has been seen in the past is that as automation becomes more reliable, the threshold for patentability, particularly inventive step, becomes higher. For example, European Patent Office Examiners will often refer to the skilled person’s familiarity with routine methods of salt form and polymorph screening in an attempt to deny an inventive step for inventions relating to these specific forms of drug substances. Is it possible that AI will make it harder to obtain patents for new chemical entities in the pharmaceutical field? It would certainly represent a significant shift in practice if the EPO were to view new compounds with good pharmacological activity as merely the result of routine screening methods. However if that one day becomes a reality due to advances in AI, then it does not seem wholly implausible to suggest that the bar for inventive step could be raised.

If the bar is raised, and it becomes harder to obtain patents for a new chemical entity, then what for originators? There is likely to be IP in the AI used to identify lead compounds, and exclusivity for the best AI would clearly provide a drug company with a competitive advantage. However having patents for the AI will not replace the need to have patent protection for the drug product itself. If patents for new chemical entities become easier to refuse or revoke, then the current jewels in the crown of the originator’s patent portfolio may be relegated to second-tier IP.

In this hypothetical world, patents directed to inventions which are made during the course of further investigations, such as medical uses, dosage regimes, etc., may become more important than patents for the compounds per se (at least for as long as predicting the results of those investigations remains beyond the capabilities of AI). Under those hypothetical circumstances, a drug company might even choose to forego a patent application for the compounds per se, keep the structure of the clinical candidate secret, and wait until further data are available before filing patent applications. That strategy would have the advantage of removing one of the key pieces of prior art which is currently available to cite against medical use and dosage patents, namely the patent to the compounds per se.

The potential of AI to weaken patents for innovative drug substances may mean that pharmaceutical companies, who will be essential in the development and validation of the next generation of in silico drug development models, have an incentive not to make those models too good (or at least not to disclose how good they are). If AI never gets to the point where new active compounds can be identified without inventive skill, then the innovator who develops the active compounds can be reasonably confident that they will be entitled to a patent for those compounds.

On the other hand, and looking even further ahead, if AI can take the pharmaceutical industry to the point where fewer clinical candidates fail, and the R&D cost per successful drug falls, then there could perhaps be a fundamental re-shaping of the industry and its relationship with patents.