Key decisions in European patent disputes from the EPO and UPC

Part 1 – T1191/19 and T0161/18 - sufficient disclosure in AI inventions

Relevance of these decisions

The European Patent Convention (EPC) requires that a patent discloses an invention “in a manner sufficiently clear and complete for it to be carried out by a person skilled in the art.” (Article 83 EPC). This requirement has predominantly been relevant to the chemistry and biotechnology fields where it is not always clear how or if a claimed invention will actually work. In contrast, in the mechanical and electronic fields, it is usually much more apparent that an invention will work, and how to implement it. It has therefore been rare for patents in those technical fields to lack sufficient disclosure.

This general distinction has been eroded by the advent of artificial intelligence (AI), since it is not always apparent how a claimed AI invention operates, or even if it will work. A particular question arises in AI systems which rely on training data to develop their “skills”, and whether there is a need to disclose that training data in an application. In this way, the requirement of sufficient disclosure in AI inventions is giving rise to issues more similar to chemistry and pharmaceutical technologies than mechanics and electronics. Recognising the developing field of AI, the latest edition of the EPO’s guidelines for examination now include a comment specifically on the requirements in this area.

General Principles

The general principle of sufficient disclosure, which applies in all technical fields, is that a patent application must include sufficient information for the Skilled Person (armed with their Common General Knowledge) to implement the claimed invention across its full scope. It is important to note that it is the claimed invention that is of interest, and that the requirement applies across the full scope of the claim. That is, if the claim extends to embodiments which could not be implemented the requirement of sufficient disclosure is not met.

Where the claims are directed to an invention which relies on, for example, a neural network, the application must disclose sufficient information for the Skilled Person to implement that neural network for the claimed purpose. It is not enough to disclose generally how to implement a neural network, but it must be disclosed how to implement the neural network that performs the invention (or at least one neural network that can do so). Therefore, recognising this is a gross over-simplification, if performance of the invention relies on specific weights for nodes of the neural network, those weights, or instructions how to arrive at them, would need to be disclosed. Without knowing those weights or how to get them, the Skilled Person could not implement the required neural network.

To continue the simplified example, usually in neural network cases the weights are determined by a training process based on training data. One way of meeting the sufficiency requirement would be to include the training data, but that is usually very undesirable (i.e. to hold back what is likely to be a trade secret and/or confidential). The latest EPO Guidelines have explicitly stated that it is not necessary to include the training data, but the “characteristics that required to reproduce the technical effect [of the claimed invention] must be disclosed unless the skilled person can determine them without undue burden using common general knowledge”. It is left open what “characteristics” means, but generally this is taken to mean a description of the training data, the set of suitable subjects, and how it is annotated to allow it to be used to train the neural network.

T1191/19 and T0161/18

Each of these decisions considered applications for inventions with an AI element.

In T0161/18 the claimed invention concerned a medical apparatus which used a neural network to calculate medical parameters. The claim specifically stated that the weighting values of the neural network were determined by training. To meet the requirements of sufficient disclosure it was therefore required that the Skilled Person could training a neural network to arrive at one which calculated the required parameters.

The application did not include an example data set but did include a statement that “input data should cover a wide range of patients of different ages, genders, constitution types, health conditions and the like”. The Board of Appeal (TBA) found that this was insufficient because it does not describe the characteristics of the data which result in correct training. In fact, the statement doesn’t narrow the data set at all and simply refers to data covering a very wide range of sources. Since the application did not specify the characteristics of the training data which would allow correct training, the TBA found the Skilled Person could not reproduce the invention without undue burden to generate a suitable training data set.

In T1191/19 the claims were to a computer-implemented method which relied on “training data” and “validation data”, but no example data sets were provided, nor were the characteristics of those data sets explained. In addition the application gave no details of the types of neural networks, functions, or leaning mechanisms that might be suitable. The TBA described the application as “more like an invitation to a research programme”, and hence found the Skilled Person would have to bear an undue burden.

Take Home Messages

It is clear that a patent application for an AI invention must include details of a training data set such as the relevant set of relevant sources, and the features of the relevant data, such that a suitable set can be created to allow training of the neural network. Similarly sufficient information must be provided to enable selection, implementation, and training of a suitable neural network. It will be interesting to see how this caselaw develops in the AI field, in particular if applicants become willing to provide some data, and if so whether that can render sufficient a broader claim (again an issue that has more prominently featured in chemistry and biotechnology fields, in particular in and following the recent G2/21 Enlarged Board of Appeal decision).

The issues discussed above also have a strong link to inventive step. Since neural networks and their advantages are now well known, there can be no inventive step in simply using a neural network for a known purposes. Instead the inventive step must lie in some feature of the neural network which provides a technical effect. Even if sufficiently disclosed, the general principle of training a neural network on selected data is unlikely to provide an inventive step, unless it can be shown that there is a technical effect arising from the training data beyond the known process of training. Again, it will be interesting to see how caselaw develops in this respect, again as to what degree of technical effect will be able to support an inventive step for broader claims.

Part 2 – How the UPC manages parallel/bifurcated proceedings, plus insights into conducting trials

Relevance of the Decisions

Since the race to file actions on Day 1 of the UPC, following years of high profile global patent litigation, Amgen v Sanofi/Regeneron case has been the avidly watched and significant pharma case shaping early caselaw at the UPC. Following the ‘bifurcation with a twist’ (see our article published on JUVE patent here), all eyes have been on how the parallel cases will proceed before the Local and Central Divisions (here, both in Munich). During February, we got some answers, via an interim decision from each division.

The outcome is a central validity-only trial proceeding first, ahead of a local infringement-only trial; with the latter slower but not stayed. As the parties near the central validity trial, there are also some interesting practice points concerning the conduct of trial, in particular in a technically complex pharma case.

Decisions

We consider each decision in turn:

• the first order came from the Munich Local Division (MLD) on 2 February 2024, issued by the full panel of the Judge-rapporteur (Zigann), alongside a further 2 legally qualified judges (LQJ) (Pichlmaier and Granata) plus 1 technically qualified judge (TQJ) (Dorland-Galliot); and then
• the order from the Munich Central Division (MCD) on 27 February 2024, issued by the Judge-rapporteur (Kupecz), who in the action sits in a panel with alongside a further legally qualified judge (LQJ) (Voß) with a technically qualified judge (TQJ) (Struve).

The MLD fully bifurcates validity from infringement, and goes slower

To recap, Sanofi and Regeneron took a split approach in defending Amgen’s local infringement action: only Regeneron filed a counterclaim for revocation and Sanofi did not (to avoid conflict with its existing central revocation action).

There is no reference to hearing the parties in this order, albeit decisions under Art.33(3) (in this choosing option b) to ‘bifurcate’) require the parties to be heard. It is also not the interim conference, which happens later (in July).

The key outcomes from the MLD are:

1) By agreement, Regeneron’s invalidity counterclaim is deferred to the MCD, to run in parallel with Sanofi’s revocation action (see MCD decision below).
2) Amgen’s infringement action against Sanofi and Regeneron remains in the MLD and is not suspended; but it is going slower than the MCD action (MCD expected to go to trial before the summer (now confirmed – see below), with MLD trial not until October 2024).
3) Overall, the MLD seems content to continue on the basis that it can suspend the infringement action if the patent is revoked centrally. But if the patent is not revoked, it remains to be seen if any of Sanofi/Regeneron’s other grounds for suspension could succeed (e.g. vs the German national proceedings, EPO proceedings and/or antitrust issues).

The MCD goes faster and is highly prescriptive on managing the parallel actions and conduct of trial

The written procedure concluded in the MCD on 6 February, with the interim conference originally scheduled the next day on 7 February. This decision refers to from a hearing via teleconference on 22 February, thereby also taking into account the above MLD order. This order therefore appears to be the interim conference decision, which would make sense given the directions for the conduct of trial, which may therefore be the next time the parties will be heard in this case.

The key points from this decision include:

1) The MCD confirms bifurcation and that the MCD is now handling dual/parallel actions (by Sanofi and Regeneron) on validity only.
2) Trial is on 4 June 2024, implicitly putting pressure on the MCD to produce a swift judgment if it is to be before the interim conference in the MLD infringement action on 1 July (at the end of the MLD order, the judge-rapporteur is ordered to communicate the interim conference date to the MCD and is permitted (after hearing the parties) to postpone the interim conference in the MLD; inferring that the MLD interim conference may be delayed until after the MCD decision on validity if it is not so swift).
3) The judge-rapporteur is highly prescriptive on the co-claimants in parallel actions (i.e. Sanofi and Regeneron) – e.g. running the same case/pleadings/submissions. Pre-trial steps include case summaries (<7500 words), demonstratives, docs list and costs schedules, starting 8 weeks out from trial.
4) There is an indication of no expert testimony or cross examination being necessary at trial –.

The bifurcation outcome

A central reason for this bifurcated outcome is avoiding the risk of inconsistent decisions on validity, which is a foundational pillar of the UPC, notwithstanding its structure and acceptance of this procedural possibility.

With Amgen agreeing to move the counterclaim centrally, this infers (if not confirms) that they did not try to stay Sanofi’s central revocation action (e.g. on the basis that the local division could cover both infringement and validity – which did seem unlikely/unfeasible due to Sanofi and Regeneron’s split approach). Similarly, there is no suggestion that the parties considered moving the whole action (infringement and revocation) from the local to central division.

The lack of stay in the MLD is perhaps the most significant issue, and though refused, the MLD infringement trial is not until 16/17 Oct 2024 (interim conference: 1 July), whereas the MCD trial is quicker, on 4 June. So ultimately, the MLD has time to take into account the MCD decision on revocation – inferring that invalidity will be a suitable reason to stay the MLD infringement action (while leaving open whether it could be stayed for other reasons if the patent is held valid).

This willingness to proceed with both parallel/bifurcated actions ‘for now’ is in line with the UPC’s display to date of quite pragmatic case management, in particular given this advances an action that might be stayed/suspended. Perhaps this has been deemed acceptable given the scale of this dispute and resources of the parties; but there is no reasoning for this and it will be interesting to see if other cases have stronger objections in this respect (noting this could be a rare scenario).

Given the MLD’s earlier insight of the MCD decision being “expected before the summer break”, it seems likely that the MCD and MLD may have discussed and/or coordinated this outcome – but again there is no mention of this or e.g. why the infringement trial is slower / not until Oct 2024. Equally, there is nothing in the MCD order to suggest the quicker trial arose due to any request for acceleration, and it did occur in almost exactly year, so bang on the UPC’s intended timeframe.

Case management insights

The MCD order is primarily interesting from a procedural perspective (both trial conduct/preparation and managing parallel actions).

We have included a table below showing the full directions, but in terms of general points:

• The court is highly prescriptive on the co-parties (Sanofi and Regeneron) doing things “identically” and each serving pleadings in each case. Interestingly, there is reference to making identical submissions and it is not clear how that will work in the oral hearing as between Sanofi and Regeneron representatives.
• Commitment to a 1 day trial (with a 2nd day strictly only if someone goes wrong technically).
• Surprisingly perhaps (or not given the 1 day limit), is the express reference to there being no need to hear the experts. There is no reasons for this, nor any reference to either party requesting cross examination. This more strongly infers that UPC Judges may be well prepared and don’t feel the need to hear experts because they understand the technology (i.e. versus the need to scrutinise written evidence). This may place considerable onus and influence on the TQJ, which is also one of the themes implied in the 10x Genomics appeal with the 2 TQJs implicitly having a strong influence on the inventive step finding. It will now be interesting if the parties decide to bring experts or now – i.e. given the possibility of questions being asked.

On the pre-hearing preparations:

• Again, the court is highly prescriptive in terms of schedule and timing, and strict prohibition against introducing new arguments or issues.
• The directions include the use of case summaries (RoP 103) and demonstratives/slides (RoP 104) and also consolidated list of exhibits (very like EPO oppositions). This displays the trending pragmatic/hybrid approach to case management by UPC, borrowing from approaches in different jurisdictions – e.g. skeleton arguments / pleading notes (such as used in Dutch hearings) and demonstratives (e.g. like Sweden/Scandinavia).
• The submissions ordered / agreed to (upon the claimant’s request) are significant, including 7500 word case summaries (approx. 15 pages) and demonstratives (no limit on slides)

Next steps

The decision reinforces how life would have been different if the MLD had been first-seized or if Sanofi had also counterclaimed in the MCD, which in either event would implicitly have both kept infringement and validity solely within the confines of the MLD. We will now see if Sanofi/Regeneron’s strategy to push central revocation pays off – i.e. if can they successfully secure central revocation and dispense with the local infringement action.

Considering strategies focussing on central revocation, it will also be interesting to see how another big pharma/MedTech develops in 2024 – i.e. Meril v Edwards – which deployed a different strategy to progress central revocation (i.e. a different group company commencing a revocation action before central division notwithstanding the commencement of a local infringement action).