How AI-predicted structures, public databases and decades of folding research reshape novelty and obviousness

Introduction

For fifty years the “protein folding problem”, predicting a protein’s three-dimensional shape from its amino acid sequence, was among the great open questions in biology. The Nobel Prize in Chemistry 1972 recognised Christian Anfinsen’s work showing that sequence determines structure. Progress was slow and incremental through homology modelling, threading, fragment assembly and physics-based simulation, and experimental methods such as X-ray crystallography, NMR spectroscopy and cryo-electron microscopy produced structures that were deposited in the Protein Data Bank (PDB), founded in 1971.

That changed with the performance of DeepMind’s AlphaFold 2 at the 14th Critical Assessment of Structure Prediction (CASP14) in late 2020, followed by the publication of its method in Nature in 2021 and the public release of the AlphaFold Protein Structure Database with the European Bioinformatics Institute. The database was later expanded to cover over 200 million predicted structures. In parallel, the Baker laboratory published RoseTTAFold in Science in 2021, AlphaFold 3 appeared in 2024, and the 2024 Nobel Prize in Chemistry recognised David Baker for computational protein design and Demis Hassabis and John Jumper for protein structure prediction.

For patent law this creates a distinctive validity problem. Patents claiming protein structures, structure-based drug targets, engineered proteins defined by structural features, or computational folding methods may now face prior art from predicted structures that anyone can download and from a rapidly growing literature. This article explains how such references can be used to challenge validity under novelty, obviousness, enablement, written description and eligibility doctrines, how to establish their prior art status, and what patentees can say in response.

Note: This article is general information and is not legal advice. Doctrines on prior art, inventorship and patent eligibility for AI-related inventions are evolving and differ between jurisdictions. Confirm current law and agency guidance with a registered patent attorney before acting.

1. Technical Background

1.1 What structure prediction tools produce

Deep-learning predictors take an amino acid sequence, often together with a multiple sequence alignment of related proteins and sometimes template structures, and output atomic coordinates together with confidence measures. AlphaFold reports a per-residue confidence score (pLDDT) and a predicted aligned error between residue pairs. Predictions are generally accurate for well-folded single-chain domains, and less reliable for intrinsically disordered regions, alternative conformations, the effect of point mutations, the influence of ligands, membrane environments and large dynamic complexes. AlphaFold 3 and similar systems extend predictions to complexes involving nucleic acids, small molecules and ions, with varying accuracy.

1.2 Why these facts matter for validity

The strengths and limits of predictions determine what a reference actually discloses. A predicted model with high confidence may disclose the fold of a protein, the arrangement of domains and the positions of residues. It will generally not prove that a mutation increases stability or that a particular ligand binds with a given affinity. Validity arguments therefore turn on carefully mapping what the prior art teaches against each claim limitation.

Table 1: Selected milestones relevant to prior art dating

DateEvent
1971Protein Data Bank established as a public archive of experimentally determined structures
1972Nobel Prize in Chemistry to Anfinsen (sequence determines structure) and others for related work on ribonuclease
1990s to 2000sHomology modelling tools (such as MODELLER, SWISS-MODEL) and Rosetta fragment assembly methods become widely used; CASP competitions begin in 1994
2018AlphaFold (first version) ranks first at CASP13
Late 2020AlphaFold 2 results at CASP14 announced and widely reported as near experimental accuracy for many targets
July 2021AlphaFold 2 paper published in Nature; code released; AlphaFold Protein Structure Database launched with EMBL-EBI; RoseTTAFold published in Science
2022AlphaFold database expanded to more than 200 million predicted structures
May 2024AlphaFold 3 described in Nature, extending predictions to biomolecular complexes
October 2024Nobel Prize in Chemistry awarded for computational protein design and protein structure prediction

The dates are important, but the precise date on which any particular item became publicly accessible, such as a preprint, a database release or a GitHub commit, must be established with evidence, as discussed in Section 3.

2. Categories of Claims Vulnerable to Structural Prior Art

  1. Claims to proteins defined by structure. Examples include a polypeptide having a particular fold, a binding pocket geometry or a set of residue positions. If a public database entry already shows a protein having those features, the claim may be anticipated or obvious.
  2. Claims to variants and mutants. These claim a protein with a specified percentage identity to a reference sequence or with substitutions at structurally defined positions, often to improve stability, activity or specificity. Predicted structures can supply the motivation for selecting residues for mutation.
  3. Structure-based drug discovery claims. These claim a compound that binds a target, or a method of treating a disease with a compound identified through a structure-guided process. Public structures of the target, whether experimental or predicted, are relevant to obviousness.
  4. De novo designed proteins. Designs generated by computational pipelines may be anticipated by previously published designs or may be obvious variations of them, especially where design software was publicly available.
  5. Computational method claims. Claims to methods of predicting structure using neural networks, attention mechanisms, multiple sequence alignments or confidence scoring are exposed to published AlphaFold and RoseTTAFold papers and open source code, in addition to eligibility problems under § 101.
  6. Use claims and screening methods. Methods of virtual screening or of designing antibodies and enzymes using predicted structures can be attacked on the basis that predicted structures and docking tools were already available.

3. Establishing Prior Art Status

3.1 Statutory framework

Under 35 U.S.C. § 102(a)(1), a claim is anticipated if the invention was patented, described in a printed publication, in public use, on sale, or otherwise available to the public before the effective filing date. Subject to the exceptions in § 102(b), including a one-year grace period for the inventor’s own disclosures, the challenger must show that the reference was available before that date. In Europe, Article 54 EPC defines the state of the art as everything made available to the public by written or oral description, by use, or in any other way before the filing or priority date, with no grace period apart from narrow exceptions.

3.2 Public accessibility of databases and preprints

A reference is a printed publication if it was publicly accessible, meaning that interested persons exercising reasonable diligence could locate it. The Federal Circuit applied this test to online and unconventional materials in In re Klopfenstein (2004), looking at factors such as the length of display, the expectation of copying and the ease of locating the material. Later cases such as Blue Calypso, LLC v. Groupon, Inc. (2016) emphasised the role of indexing and searchability. For structural biology references the following sources are generally strong:

  • Journal articles with publication dates, including supplementary data.
  • Preprints on servers such as bioRxiv and arXiv, which carry posting dates and are widely indexed and routinely cited as prior art.
  • Database entries in the PDB, UniProt, the AlphaFold Protein Structure Database and the ModelArchive, each of which records deposition and release dates and version histories.
  • Conference materials, such as CASP abstracts and posters, if they were publicly available.
  • Source code repositories such as GitHub, where commit logs, release tags and archive snapshots show dates.
  • Theses and dissertations indexed by university libraries.
  • Patent publications that describe prediction methods or structures.

3.3 Practical evidence

Challengers should collect dated evidence: official release notes, database version archives, web archive captures from services such as the Internet Archive, declarations from librarians or database curators, publisher metadata and download logs. The risk with databases is that entries are updated over time, so a record visible today may not match the record available on the critical date. Retrieve archived versions or version-specific files, and state which release is relied upon. For predicted structures, identify the release version of the model and the date on which a specific entry was added, since the database has been updated with new model versions and sequences.

3.4 Is a prior art reference enabling?

To anticipate, a prior art reference must disclose each element and enable a person of ordinary skill to make the claimed invention. Printed publications are presumed enabling in ex parte examination and, in practice, are generally treated as enabling absent contrary evidence (In re Antor Media Corp., Fed. Cir. 2012). A patentee may argue that a predicted structure does not enable a protein with the claimed properties because it does not teach how to make or use it. Challengers can respond that, for claims directed to a protein sequence or structure itself, the reference disclosing the sequence and structure enables the protein, since expression and purification techniques were routine. For obviousness, the reference need not be independently enabling in the same way, since it is considered for what it teaches the skilled person together with the rest of the art.

4. Novelty (Anticipation) Arguments

Anticipation requires that a single reference disclose every limitation of the claim, arranged as in the claim. It is rare in this field for a single predicted structure to disclose a claim that recites both structure and function, but there are realistic scenarios.

  • Natural proteins claimed by structure. If the claim covers a protein having a certain fold or motif and the protein occurs in nature, then a published sequence and an experimentally determined structure may anticipate it. Eligibility problems may also arise for natural products.
  • Inherent anticipation. A reference can inherently anticipate if the claimed feature is necessarily present in what the reference discloses, even if the reference does not recognise it. A prior sequence necessarily folds into the structure determined by that sequence, so a challenger may argue that a later-recognised structural feature of a known protein was inherent. Inherency requires necessity rather than probability, and it must be supported by evidence, which can include later experimental structures or predictions, though the use of post-dated evidence to show inherency has limits.
  • Predicted structure as disclosure. If a database entry gives coordinates for a protein that match the claimed structural parameters, for example a distance between residues or a binding pocket volume, then the entry may disclose those features. The patentee may argue that a prediction is only a model and does not establish what the real protein looks like. The response is that the claim language, not the truth of the structure, controls, and that the reference does disclose coordinates with stated confidence levels. A related issue is whether the reference discloses sufficient confidence to be reliable for the purposes of the claim.
  • Prior designs. Published designed proteins with sequences that fall within a claimed identity range can anticipate claims to such sequences.

5. Obviousness Arguments

5.1 The legal framework

Under 35 U.S.C. § 103, a claim is invalid if the differences between it and the prior art would have been obvious to a person of ordinary skill in the art. KSR International Co. v. Teleflex Inc. (2007) rejected rigid application of the teaching-suggestion-motivation test and emphasised common sense, design needs and market pressures, and finite predictable solutions. Europe applies the problem-and-solution approach, asking whether the skilled person, starting from the closest prior art and faced with the objective technical problem, would arrive at the claimed solution in an obvious way.

5.2 How predicted structures influence the analysis

  • Motivation to modify. A public structure showing a flexible loop, an exposed hydrophobic patch or an unstable region can supply a reason to introduce stabilising mutations, glycosylation changes or linkers.
  • Reasonable expectation of success. The accuracy of modern predictors, and the existence of tools to model mutations, can strengthen the argument that a skilled person would expect to succeed in designing variants, although patentees will emphasise that predictions of the effect of mutations remain uncertain.
  • Obvious to try. In In re Kubin (Fed. Cir. 2009), the court held that isolating a gene was obvious where the art provided a reason and a reasonable expectation of success using routine techniques. Where predicted structures reduce the search space to a finite number of candidates, the argument for obviousness to try gains force.
  • Combination of public tools. Using a known prediction tool on a known sequence to obtain a structure and then applying standard docking software to a known compound library can be characterised as the predictable use of known techniques.
  • Level of ordinary skill. The level of skill rises as tools become widely available, and a post-2021 skilled person is routinely assumed to be capable of running and interpreting AI-based predictions. The relevant date is the effective filing date, so the time of filing matters greatly.

5.3 Patentee responses

  • Unpredictability. Evidence that predicted structures did not reliably reveal the effect of the claimed mutations, or that experimental results were surprising.
  • Unexpected results. Data showing activity, stability or specificity beyond what the art would have suggested, with a demonstrable nexus to the claimed features.
  • Teaching away. Literature that discouraged the approach, for example by reporting poor predictions for the relevant protein class.
  • Hindsight. Arguing that the challenger used the claimed invention as a template to select and combine references.
  • Secondary considerations. Commercial success, long-felt need, failure of others and industry praise, tied to the claimed features.
  • Date issues. Showing that the reference was not publicly available before the critical date, or that the claim has an earlier effective date.

6. Enablement and Written Description Challenges

Prior art is not the only basis for invalidity. Claims defined by structure or function across a wide class of proteins can be challenged under § 112(a). In Amgen Inc. v. Sanofi (2023), the Supreme Court held that claims to a genus of antibodies defined by function were invalid for lack of enablement because the specification did not enable the full scope without undue experimentation. A claim to “any protein having at least 70% identity to SEQ ID NO: 1 and having improved stability” invites similar scrutiny. Predicted structures can be used by both sides. A challenger can point out that the specification describes only a few examples and that residue-level effects are hard to predict, and the patentee may argue that modern prediction tools make it easy for the skilled person to design and test variants, which supports enablement. This is a notable irony: arguments that tools make the art predictable, helpful on enablement, can hurt on obviousness, so patentees need to align their positions across proceedings.

In Europe, the question of plausibility, that is whether the application makes it plausible that the claimed technical effect is achieved, is particularly relevant for structure-based claims where the application presents only modelling and no experimental data. An application that relies on predictions alone may face objections that the technical effect is not credible, and later data may not be able to fill the gap.

7. Eligibility and Inventorship Issues Specific to AI-Based Claims

7.1 Subject matter eligibility

Method claims for predicting structure may be directed to mathematical concepts or mental processes. Under Alice, the claim must include an inventive concept beyond the abstract idea. The Federal Circuit held in Recentive Analytics, Inc. v. Fox Corp. (2025) that claims merely applying generic machine learning techniques to a new data environment were ineligible. The USPTO has also issued guidance on the eligibility of AI-related inventions. A claim that recites a specific improvement to the technical operation of the model, or that results in a tangible physical product such as a purified engineered protein, may fare better. Challengers in district court or before the Patent Trial and Appeal Board can use § 101 as an independent ground in court, but not in inter partes review, where only §§ 102 and 103 based on patents and printed publications are available.

7.2 Inventorship and AI assistance

In Thaler v. Vidal (Fed. Cir. 2022) the court held that an inventor under U.S. law must be a natural person. Where AI tools contribute to the design of a protein, the human contribution to conception must be significant enough to support inventorship. Misjoinder or non-joinder of inventors can in principle be corrected, but errors found late may form part of an invalidity or unenforceability challenge. USPTO guidance on AI-assisted inventions has been revised in recent years, and practitioners should check the guidance currently in force. Maintaining records showing how humans selected, evaluated and validated AI outputs is good practice.

8. Procedural Routes for Challenging Validity

RouteKey featuresRelevance for structural prior art
Third-party pre-issuance submission (35 U.S.C. § 122(e))Allows submission of printed publications during a limited window before allowanceLow cost way to put database papers and preprints before the examiner
Ex parte reexaminationAny person can request; based on patents and printed publications; no estoppel on the requesterUseful for focused challenges using AlphaFold literature
Inter partes review (IPR)Challenge on §§ 102 and 103 based on patents and printed publications; petition after nine months from grant or the end of any post-grant reviewStrong venue for obviousness over published structures and methods; database entries must qualify as printed publications
Post-grant review (PGR)Within nine months of grant for first-to-file patents; any ground of invalidityAllows §§ 101 and 112 challenges as well as prior art
District court litigationAll defences, including § 101, § 112 and inventorship; clear and convincing evidence standardAllows expert testimony on predictions, inherency and public accessibility
EPO oppositionWithin nine months of grant; grounds include novelty, inventive step, sufficiency and added matterEffective for centralised attack on European patents; consider plausibility and sufficiency
Third-party observations (Art. 115 EPC)Can be filed during examination and oppositionProvide prior art without joining as a party

The choice of route depends on the available art, the grounds, cost, timing and risk of estoppel. Because IPR is limited to patents and printed publications, challengers using database entries should be prepared to show that they qualify as printed publications. The Patent Trial and Appeal Board has treated dated, publicly accessible database records and preprints as printed publications where supported by evidence, but outcomes depend on the record.

9. Building an Invalidity Case: A Step-by-Step Approach

  1. Determine the effective filing date of each claim, taking account of priority claims and whether each claim is supported by the earliest application.
  2. Construe the claims. Identify structural terms, percent identity ranges, functional language and the computational steps, and propose constructions.
  3. Search broadly. Cover journals, preprints, PDB and AlphaFold database entries, CASP materials, code repositories, theses and patents. Search by sequence using BLAST and by structure using tools such as Foldseek and DALI.
  4. Date the references. Collect release logs, web archives and declarations. Record which version of each database was relevant.
  5. Map claim limitations to references, with a chart showing where each limitation is disclosed, including specific coordinates, residues or figures.
  6. Develop motivation and expectation of success evidence, supported by expert declarations and contemporaneous literature on tool accuracy and routine practice.
  7. Anticipate rebuttals on unpredictability, unexpected results and enablement of the prior art, and gather counter-evidence.
  8. Assess parallel grounds such as § 112 and § 101, and plan consistent positions on the level of skill and predictability.
  9. Choose the forum and timing, and prepare expert support for a technically complex record.

10. Illustrative Scenario

Suppose a patent, filed in 2022, claims an enzyme having at least 85% sequence identity to a known bacterial hydrolase and containing substitutions at three residues that line the substrate channel, the enzyme showing at least two-fold higher thermostability. The patent presents experimental stability data for four variants.

  • Prior art. The AlphaFold database entry for the hydrolase was publicly available from 2021, a 2019 paper reports a crystal structure of a homologous enzyme and identifies the substrate channel, and a 2020 paper reports stabilising mutations at two of the three positions in a related enzyme.
  • Challenger’s case. The structure and the channel position are disclosed. The related enzyme paper provides a reason to try substitutions at those positions. Predicted and experimental structures narrow the choices. The skilled person would reasonably expect improved thermostability, since the technique was routine. The 85% identity range reads on known homologues.
  • Patentee’s case. The third position is not taught and the combination produced a surprising synergy shown in the data. Stability effects of substitutions are unpredictable, as the challenger’s own enablement arguments elsewhere show. The data for four variants may not be commensurate in scope with a claim covering all variants at 85% identity.
  • Likely outcome pressures. The claim may survive in narrowed form limited to the tested combination, while the broad identity genus is at risk under §§ 103 and 112.

The scenario shows that structural prior art rarely eliminates a patent outright, but it can significantly narrow claims toward what was actually demonstrated and non-obvious.

11. Drafting and Prosecution Lessons for Applicants

  • Include experimental data. Predictions alone are poor support. Provide measured activity, binding, stability or structure data.
  • Draft layered claims. Pair broad genus claims with specific fallback claims tied to tested variants and sequences.
  • Avoid claiming predicted features without support. Structural parameters from models are fragile and may later be shown to be inaccurate.
  • Disclose the human contribution. Describe the roles of the inventors in selecting targets, interpreting predictions and validating results.
  • Search for AlphaFold-era art. Include database entries and preprints in prior art searches and information disclosure statements.
  • Keep positions consistent. Statements about predictability made to support enablement can be quoted to resist non-obviousness, and vice versa.
  • Monitor effective dates. The pace of change in the field means that a few months can alter the prior art landscape.

Conclusion

AlphaFold and related tools have changed the evidence available in patent disputes over protein-related inventions. Public structure databases, prominent papers and open source code now form part of the prior art, and they raise the level of skill and the expectation of success that fact-finders assume. Challengers can use dated, publicly accessible structures, preprints and code to argue anticipation, obviousness, lack of enablement and ineligibility, while patentees can respond with evidence of unpredictability, unexpected results and strong experimental support. The most important practical points are to establish public accessibility with proper evidence, to read each reference for what it really discloses, to keep arguments consistent across doctrines, and to ground claims in measured data rather than in predictions. As the technology and the law continue to develop, current guidance and case law should be reviewed before any filing or challenge.

Leave a Reply

Your email address will not be published. Required fields are marked *