Article

Artificial intelligence and biotechnology: challenging an in silico invention
Patents

Artificial intelligence and biotechnology: challenging an in silico invention

What in silico modelling can offer opponents before the EPO

The first two articles in this series took the applicant’s perspective: how the EPO assesses in silico data, particularly with regard to sufficiency of description and inventive step, and then what practical approaches to adopt, right from the filing strategy stage, to construct applications that are as robust as possible.

But in silico modelling is not solely the patent holder’s ally. In the hands of the opponent, its function changes. It is no longer a matter of establishing the credibility of an invention, but of testing that of the patent, assessing its scope, or providing evidence for which an experimental equivalent would be cumbersome, costly, or even out of reach. In silico modelling thus becomes less a tool for discovery than a tool for providing evidence.

A natural first step, of course, is to examine how the patent holder themselves has used the modelling. If the patent relies solely on the use of a model, the opponent must be able to assess the robustness of the model and determine whether it is based on sufficiently relevant real-world data; and whether the assumptions made are verifiable. Failing that, an objection on the grounds of insufficiency or a lack of credibility of the technical effect may be considered.

Similarly, if the proprietor submits, after filing, a model intended to achieve a therapeutic effect that was initially absent, the opponent may point out that, in accordance with the line of reasoning set out in T 609/02, subsequent evidence can only confirm a teaching already rendered credible by the application as filed; it can never create it. With regard to inventive step, the opponent may also argue, in the context of G 2/21, that the claimed effect was not encompassed by the original technical teaching, or was not embodied by the invention as it was initially disclosed. These principles are not specific to in silico evidence: they apply to any evidence produced after filing, whatever its nature.

This assessment remains important. However, it extends the approaches already examined from the applicant’s perspective. The most significant benefit for the opponent lies elsewhere: in the possibility of submitting their own modelling as evidence, potentially supported by targeted experimental data, even against a patent that was never based on an in silico approach.

Revealing an inherent property and undermining novelty

Decision T 1121/23 provides a particularly clear illustration of this.

The patent claimed an oligonucleotide capable of forming an intramolecular stem-loop structure. A prior art document described a structurally identical oligomer, but without mentioning this folding property. The opponents then produced predictions of the RNA secondary structure, obtained using several different software, to establish that the prior art oligomer already formed the claimed structure.

The Board dismissed the proprietor’s objections. The fact that the software used were subsequent to the priority date of the prior art was not considered decisive: a secondary structure is an inherent property of a given primary structure. It either exists or does not exist by virtue of that structure, irrespective of when it is revealed. Nor did the Board require the structure to be predominant or permanent. A transient occurrence, present in a significant proportion of the possible configurations, was sufficient.

The practical significance of this decision is considerable. Where a prior art document describes an entity that is structurally identical to that claimed, but remains silent on one of its properties, the opponent may use in silico modelling to establish that this property was already necessarily present. In silico modelling does not, therefore, serve to supplement the technical teaching of the prior art: it serves to reveal what was already inherent in the disclosed structure.

For an opponent, this line of reasoning can be applied to other biomedical situations: nucleic acid folding, ligand-target interaction, protein conformation, epitope exposure, relative affinity or physicochemical properties arising directly from a known structure. Caution remains necessary: the more the property depends on complex or variable experimental conditions, the more robust the computational evidence must be. However, where the alleged property derives directly from the previously disclosed subject-matter, modelling can become a particularly effective tool for undermining novelty.

Challenging the claimed scope

Before the EPO, the technical advantage invoked by the proprietor must be capable of being reasonably generalised to the claimed scope. This does not necessarily mean that every covered embodiment must produce the effect with the same intensity, nor that an isolated counterexample is always sufficient to invalidate the argument. On the other hand, if the effect disappears across a significant part of the scope, becomes marginal for certain covered families, or can no longer reasonably be expected in the absence of sufficient technical guidance, the claimed advantage loses its force.

The opponent may then use in silico modelling to explore the claimed scope and identify areas of weakness: variants that no longer bind to the stated target, mutations that eliminate an alleged activity, molecules covered by the general formula but lacking the claimed affinity, combinations predicted to be non-synergistic , or subpopulations in which the alleged therapeutic benefit does not manifest itself.

The aim is not necessarily to experimentally replicate the entire scope, nor to demonstrate that every untested variant fails. It may suffice to show that the technical advantage cannot be credibly generalised to the entire claim. If this demonstration is convincing, the objective technical problem may need to be reformulated in a less ambitious manner, for example as the mere provision of an alternative. The inventive step then becomes more difficult to defend.

This strategy is particularly relevant, for example, against broad claims defined by a function, a class of compounds, a family of sequences or a therapeutic combination, of which only a few aspects have actually been tested. When adapted to the property in question, the model can enable the scope to be mapped out quickly and the most useful counterexamples to be identified: those which reveal the limits of the generalisation on which the claimed technical advantage is based. 

The same reasoning can also form the basis of an objection on the grounds of insufficiency.

A functional or therapeutic claim is only sufficiently described if a person skilled in the art can carry it out across its entire scope without undue effort. If, in reality, achieving the claimed effect requires selecting, on a case-by-case basis, the rare effective variants from within a very large set, the opponent may argue that the patent imposes a process of trial and error rather than routine work.

In silico modelling can then be used to objectively demonstrate this difficulty. It can demonstrate that the structure-activity relationship proposed by the patent is insufficient, that certain essential parameters have not been taken into account, or that large portions of the claimed scope are not compatible with the alleged technical mechanism. It can also highlight that the examples in the patent do not provide generalisable guidance, but merely isolated results.

This approach does not always replace experimental evidence. In cases where the claimed effect is therapeutic, or depends on a complex biological environment, corroboration through in vitro data or targeted experimental data will often considerably strengthen the case. However, in silico analysis can already play a decisive role: selecting relevant counter-examples, prioritising lines of attack, and demonstrating that the burden placed on the person skilled in the art exceeds what the patent actually teaches.

Challenging the reasonable prospects of success

Modelling can also be used in the opposite direction: not to demonstrate that the invention does not work, but to establish that it was foreseeable.

If, on the priority date, the person skilled in the art had access to commonly used computational tools, accessible databases, and a limited number of candidates to test, the opponent may argue that there was a reasonable chance of success. In this scenario, in silico modelling no longer serves to reveal a flaw in the patent, but to reconstruct what a person skilled in the art might routinely have done on the basis of the prior art.

The argument must, however, be handled with precision. It is not enough to assert that a model existed or that screening was possible. It must also be shown that the method was accessible to a person skilled in the art, that it formed part of their standard toolkit or that they had a reason to use it, and that it would have led, in a sufficiently predictable manner, to the claimed subject-matter. Otherwise, the reasoning risks being based on a retrospective reconstruction.

This is where the distinction highlighted by T 265/23 remains useful. A model may be sufficiently robust to render an effect credible in terms of sufficiency of disclosure, without necessarily forming part of the general knowledge that would render the invention obvious. For the opponent, the question is therefore not merely whether modelling was possible, but whether it was actually available, relevant and sufficiently motivating at the priority date.

Evidence subject to the same requirements

In silico is not exempt from the standard requirements for evidence.

An opponent who submits their own modelling must accept the same level of rigor that they would require of the patent holder: an identifiable method, documented parameters, verifiable input data, explicit assumptions, and, ideally, validation through several convergent approaches or targeted experimental data.

The probative value of a model depends closely on what it purports to establish. To demonstrate a property inherent in a structure that has already been disclosed, robust and convergent predictions may suffice, as illustrated in T 1121/23. To challenge a therapeutic effect or a complex biological activity, the burden of proof should be greater: the EPO will naturally pay close attention to the link between the model, the available biological data and the claimed effect. In such cases, in vitro data – even if limited but well-chosen – can transform a plausible model into evidence that is difficult to dismiss.

In addition to this technical requirement, there is a procedural requirement. Computational evidence remains evidence like any other. If produced late – particularly only at the appeal stage or in a delayed response to the proprietor’s arguments – it is subject to the ordinary rules of admissibility. In opposition proceedings, it is therefore in the opponent’s interest to submit their simulations as early as possible, together with the elements required to reproduce them and a precise explanation of the technical fact they seek to demonstrate.

In summary

In silico modelling is not the sole preserve of the applicant preparing their application. It offers the opponent a particularly powerful evidential tool: to reveal an inherent property, challenge the scope of a claim, undermine a technical advantage, support an objection regarding the sufficiency of the description or demonstrate that a line of research was foreseeable on the priority date.

Case T 1121/23 shows that such evidence can be decisive, even where the contested patent itself is not based on in silico modelling. However, it also implicitly reminds us that the strength of a model depends on its rigor: a simulation is only convincing if it is well-documented, verifiable and technically linked to the fact it purports to establish.

Assessing what a model can demonstrate – and what it cannot – therefore requires combining experience in the pharmaceutical and biotech sectors with a detailed understanding of the computational tools used. It is precisely this dual expertise that Plasseraud IP brings together: a healthcare team specialising in the pharmaceutical and biotech sectors, combined with a team dedicated to artificial intelligence. The experience of the patent engineers in these two teams in opposition and appeal proceedings before the EPO enables them to provide comprehensive support to clients, whether this involves building a case based on in silico data or defending a portfolio against this type of evidence.

Share on: