Artificial Intelligence in Drug Discovery: Scientific Translation, Patentability and AI-Enabled Patent Examination

Kandula Mahesh

7 minute read

Artificial intelligence is becoming an important component of pharmaceutical research, particularly in medicinal chemistry, molecular biology, target discovery, protein modelling, biomarker identification and clinical-development strategy. Machine-learning systems can analyse biological and chemical datasets at a scale impossible for conventional manual review, while generative AI can propose new molecular structures, proteins and therapeutic hypotheses. The critical question, however, is not whether AI can generate more candidates. It is whether AI can produce scientifically credible, experimentally validated, clinically relevant and legally protectable inventions.

AI is good where data is abundant and structure-linked — physicochemical properties, ADMET, synthesizability, safety pharmacology — and that gets you through Phase 1. It is not yet good where data is scarce, conditional and expensive: whether a target actually drives human disease. That is what kills you in Phase 2.

A particularly important development is closed-loop discovery. Instead of the conventional sequence of data, model and prediction, active-learning systems can operate as model → experiment → result → model update → next experiment. Sixteen authors from academia, pharma, biotech and investment ask what the field has avoided: has AI delivered safer, more effective medicines, faster? Their answer is that evidence of clinically relevant impact remains disappointingly limited — and these are not sceptics. Scannell defined Eroom’s Law and predictive validity; Grisoni built the activity-cliff benchmarks; Ghiandoni has published on AstraZeneca’s DMTA augmentation. Their diagnosis:

1. Clinical translation is not the objective function. Models are benchmarked on AUC and RMSE, not on the decision a project team must make. A model can be excellent and useless if it changes neither which molecule is synthesised nor which programme is killed.

2. Life-science data is conditional.

3. Problems are underspecified. Without a defined context of use — which chemical space, which decision, at what cost of a false negative — two models with identical benchmark scores behave differently in production.

4. Technology push has outrun science pull, and operationalisation takes years.

5. Benchmarks should measure improvement in decision-making, not model validation.

AI has changed how discovery is done — ADMET triage, structure prediction, phenotypic profiling, protein design — in gains that are real and cumulative but invisible, because they appear as efficiency. It has not yet changed the probability that a molecule entering the clinic will help a patient. The Phase 1 versus Phase 2 divergence says why: we are much better at designing molecules and no better at choosing targets.

Neither office obstructs patenting AI-derived medicines; the inventor must be human, which is a documentation problem, not a doctrinal barrier. What both systems demand is what the Perspective demands scientifically — real data, a stated technical effect, a disclosed reproducible method, and a result a routine run of a standard tool would not have produced. A programme built to survive an inventive-step attack at the EPO is, not coincidentally, a programme built to survive Phase 2.

Rentosertib’s Phase 3 readout is the first proper test — the first time an AI-identified target carrying an AI-generated molecule is examined at the scale.

Patentability of AI-Assisted Drug Discovery in the United States

The use of AI does not automatically prevent patent protection in the United States. Current USPTO guidance maintains that only natural persons can be inventors. AI systems are treated as tools, even when they contribute significantly to molecular generation, target selection or biological analysis.

For pharmaceutical organizations, documenting human conception is therefore important. Relevant contributions may include identifying the scientific problem, defining model constraints, selecting a promising AI-generated structure, recognizing its technical significance, designing validation studies, interpreting unexpected results or identifying a therapeutic application. An AI system itself cannot be named as an inventor.

Once inventorship is established, AI-assisted inventions remain subject to conventional U.S. patent requirements. The claimed invention must satisfy patent eligibility under 35 U.S.C. §101, novelty under §102, non-obviousness under §103 and written-description and enablement requirements under §112. The fact that AI was used to discover a molecule does not itself make the molecule obvious or unpatentable.

European Patentability of AI-Assisted Drug Discovery

The European Patent Office follows the same fundamental human-inventorship principle. Its DABUS decisions established that the inventor named in a European patent application must be a natural person. Using AI as part of research does not itself prevent a drug, biologic or technical process from receiving patent protection.

For the resulting pharmaceutical product, traditional EPC requirements continue to apply, including novelty, inventive step, industrial applicability and sufficiency of disclosure.

The analysis becomes more specific when protection is sought for the AI technology itself. Under current EPO Guidelines, artificial intelligence and machine-learning models are generally treated as mathematical or computational methods when considered in isolation. To contribute to inventive step, AI-related features normally need to contribute to solving a technical problem through a technical effect.

AI also creates opportunities for drug repurposing. If computational analysis identifies a previously unknown therapeutic use for a known compound, Europe provides potential protection through the further-medical-use framework under Article 54(5) EPC, provided the use is novel, inventive and sufficiently supported. Here, the invention may lie not in new chemistry but in an unexpected therapeutic application of existing chemistry.

AI and Patent Examination

AI is changing patent examination as well as invention creation. The USPTO has incorporated AI-assisted search capabilities such as Similarity Search within examiner tools. Such systems can rank potentially relevant patent documents using semantic similarity, supplementing conventional keyword and classification searches. The USPTO has also experimented with automated pre-examination prior-art searching. Importantly, substantive patentability decisions remain with human examiners.

The EPO is likewise using AI for classification, file allocation, pre-search and identification of potentially relevant prior art while maintaining a human-centric examination system. Improved machine reading of chemical structures, formulas, tables and complex patent documents may be particularly significant for pharmaceutical prior-art searching.

This evolution has practical consequences. Patent applicants should assume that future examiners will be increasingly capable of identifying conceptually similar inventions even where terminology differs. Patent strategies based primarily on different wording rather than genuine technical distinction are likely to become less effective.

AI is also increasingly used by patent professionals themselves for prior-art searching, claim review, invention mining and drafting assistance. Both the USPTO and EPO maintain that professional responsibility remains with human practitioners. Confidentiality is especially important in pharmaceutical research: uploading unpublished chemical structures, biological sequences, target hypotheses or experimental data into uncontrolled third-party AI systems may expose valuable trade secrets or future patent rights.

Integrating Science and IP Strategy

AI makes early collaboration between scientists and patent professionals increasingly important. Each promising AI-generated discovery should be evaluated by asking: What did the model contribute? What did the human inventor contribute? What experimental evidence validates the result? What technical effect distinguishes it from prior art? What claim scope can the evidence realistically support?

The strongest AI pharmaceutical portfolios may use layered protection covering the molecule or biologic, chemical genus, formulation, therapeutic indication, biomarker strategy, manufacturing technology and, where appropriate, the underlying AI-enabled technical platform.

References

Vamathevan J, Clark D, Czodrowski P, et al. Applications of machine learning in drug discovery and development. Nature Reviews Drug Discovery. 2019;18:463–477.
Stokes JM, Yang K, Swanson K, et al. A deep learning approach to antibiotic discovery. Cell. 2020;180:688–702.e13.
Zhavoronkov A, Ivanenkov YA, Aliper A, et al. Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nature Biotechnology. 2019;37:1038–1040.
Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589.
Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630:493–500.
Gangwal A, Ansari A, Ahmad I, et al. Generative artificial intelligence in drug discovery: basic framework, recent advances, challenges and opportunities. Frontiers in Pharmacology. 2024;15:1331062.
USPTO. Revised Inventorship Guidance for AI-Assisted Inventions. 2025.
USPTO. Patent Subject Matter Eligibility Guidance Update Including on Artificial Intelligence. 2024.
USPTO. Guidance on Use of Artificial Intelligence-Based Tools in Practice Before the USPTO. 2024.
European Patent Office. Guidelines for Examination: Artificial Intelligence and Machine Learning. 2026.
European Patent Office. Guidelines for Examination: Simulation, Design or Modelling. 2026.
European Patent Office. Artificial Intelligence System as Inventor — DABUS, J 8/20.
Bender et al., Nat. Rev. Drug Discov. (2026) DOI 10.1038/s41573-026-01496-2 · Xu et al., Nat. Med. 31, 2602 (2025) ·
Ren et al., Nat. Biotechnol. 43, 63 (2025) · Jayatunga et al.,
Drug Discov. Today 29, 104009 (2024) ·
Scannell et al., Nat. Rev. Drug Discov. 21, 915 (2022) · 90 Fed. Reg. 54636 (2025), rescinding 89 Fed. Reg. 10048 (2024) ·
Thaler v. Vidal, 43 F.4th 1207 ·
Amgen v. Sanofi, 598 U.S. 594 · KSR, 550 U.S. 398 · USPTO RFC on AI/PHOSITA and AI Tools Guidance (Apr 2024) ·
EPO Guidelines (1 Apr 2026), G-II 3.3.1, F-III 3 · J 8/20, J 9/20; [2023] UKSC 49 · G 1/19, G 2/21, G 1/23, G 1/24 · T 0161/18; T 206/83.

Disclaimer

This article is an independent review based on peer-reviewed literature and publicly available materials. Patent-related discussion is intended for scientific and educational purposes and does not constitute legal advice.