Can AI Save Biotech IP?

The long-standing problem of biotech patents—lack of sufficiency of disclosure (Art. 83 EPC) or an inventive step that cannot be substantiated (Art. 56 EPC)—seems, with the rise of artificial intelligence (AI), finally solvable. Or is it?

Biotech companies know the dilemma: patent claims often have to be significantly narrowed in order to overcome objections from the European Patent Office (EPO) and attacks by competitors. More data, more experiments—but where to get them when time and resources are scarce?

This is where hopes are pinned on AI: systems that learn from real experiments could simulate countless variants in silico and deliver convincing results. In theory, they could thus substantiate the technical effect of a broad claim. Sounds tempting—but practice is sobering. Even AI-based systems must clear the same patentability hurdles as biotech inventions themselves: sufficiency of disclosure and plausibility. According to decision T 0161/18, an AI is only “sufficiently disclosed” if training data, characteristics of the neural network, and learning methods are specifically disclosed. The trained model must be calibrated and validated with reference experiments so that its output qualifies as a biologically valid statement. Merely stating that a “neural network” has been trained is not enough.

Applied to complex biological systems, this means that a “biotech AI black box” is neither patentable as an invention nor suitable for providing plausible in-silico data to support technical effects.

If AI itself is claimed as an invention, strict requirements apply regarding reproducibility and technical disclosure. If, on the other hand, it is used as a research tool—for example, to simulate biological reactions—its output comes into focus. But even then, reproducibility, validity, and plausibility are decisive. An unevaluated or insufficiently described AI cannot substantiate reliable technical effects.

For “virtual biotechnology,” the following also applies: the AI system must correctly model the (bio)medical problem to be solved. General references to training or validation data are insufficient (T 1191/19). If, for example, the binding of a CAR-T cell, an antibody–antigen interaction, or an immune response is to be simulated, biochemical, molecular-biological, and other parameters familiar to the skilled person must be described in detail. If an inadequately documented analysis method is merely replicated digitally, the objection of lack of sufficiency remains (T 1462/22).

Whether these standards will change in the future remains open. However, with the growing role of AI in research and development, EPO practice is also likely to evolve—for example, through clearer guidelines on disclosure requirements for training data and AI validation. A coordinated framework between AI regulation (EU AI Act) and patent law could also emerge to enable the legally secure assessment of AI-generated data. In the long term, a “verifiable virtual evidence base” could thus be conceivable, in which validated simulations are recognized as supplementary evidence.

So far, there is no EPO decision that specifically addresses AI-simulated data used to support an invention. The decisive rulings remain T 0161/18, T 1191/19, and T 1462/22 in conjunction with G 1/19 and G 2/21. Whether only highly advanced virtual twins of a living cell or of the biological systems under investigation—such as in the EU project CERTAINTY (“A Cellular Immunotherapy Virtual Twin for Personalised Cancer Treatment”)—will enable a robust demonstration of technical effects remains to be seen. Under the principle of free evaluation of evidence, AI-simulated data may in principle be taken into account. However, they will be persuasive only if they credibly demonstrate the claimed technical effects in the individual case (G 2/21). At the same time, the standard applied to a virtual evidence base must not exceed that of the existing EPO case law in the field of “classical” biotechnology.