AI for patent sequences: Avoiding costly sequence errors

At Evolve we build our own AI agents for pharmaceutical IP. In biotech patents, few things matter more than getting the sequences of your molecules right, and sequence errors in patent applications are far more common than most people expect. Often the error sits in a sequence that is not part of a key embodiment and does no harm. Sometimes, though, it sits in the lead embodiment that is supposed to protect your blockbuster product, and that can be a very expensive mistake. We have therefore built a set of AI agents for working with patent sequences. Athena, our strategy agent, keeps the sequence work aligned with the IP strategy for the case or family. Darwin, our drafting agent, turns the sequences in an application into a WIPO Standard ST.26 sequence listing. Gilbert, our quality checking agent, checks that the claims, the listing and the specification really cover the lead candidate, at the drafting stage and throughout prosecution. Albert extracts and checks the sequences in granted patents.

Why getting patent sequences right matters

For a biologic, the sequence is often the invention. Antibodies, CARs, T cell engagers, mRNA constructs and gene therapy vectors are defined in the claims by the sequence of their various domains. If the sequence listing is wrong by a single letter, the claims can therefore define something other than what the applicant meant to protect. In the worst case, this can mean that the patent does not cover the commercial product.

The case law shows not only how expensive a sequence error can be but also how difficult it is to put right once the application has been filed.

In the Soliris saga, Alexion’s claim to eculizumab defined the light chain as “consisting of” SEQ ID NO: 4. That sequence included a 22-amino-acid leader sequence, which is cleaved off and is not part of the marketed antibody. Alexion asked the EPO to correct the claim to SEQ ID NO: 4 without the leader, and the Board of Appeal refused (T 1515/20). The Board accepted that antibody chains usually carry a signal peptide, but said the error was not obvious from the application itself. An amendment excluding the leader sequence was also refused for added matter, so the patent was granted with the uncorrected sequence. In a sister application the same correction was refused and the application failed altogether (T 1435/20).

The uncorrected sequence then decided the litigation. The High Court held that the claim did not cover eculizumab, so the biosimilars did not infringe (Samsung Bioepis v Alexion [2025] EWHC 1240 (Pat)). The UPC Court of Appeal also declined to correct the error by interpretation, because neither the error nor the correct version was certain enough to the skilled person, and it thought the claims were more likely than not insufficiently disclosed (UPC_CoA_402/2024 and UPC_CoA_405/2024, 20 December 2024). Alexion is far from the only patentee to get a sequence wrong (T 2126/08, T 1111/02, T 1805/06), and even a successful correction does not always save the patent (T 0580/05).

Sequence errors can also cost you your priority date. In T 1213/05, the priority document contained inadvertent sequence errors. The Board held, following T 70/05, that a priority claim to an incorrect sequence cannot be maintained, whatever the reason for the mistake, and the claims to the corrected sequences fell to intervening prior art (more on T 1213/05).

These errors are easy to make and hard to spot. A listing can contain hundreds of sequences, many of them hundreds or thousands of residues long, and a single wrong amino acid halfway along a 400-residue chain is very easy for a proof-reader to miss.

Getting sequences right is about much more than turning the sequences in the specification into a sequence listing. A listing can only be as good as the specification it comes from, and the harder work starts earlier and goes on for longer. We need to know that the sequence in the application is the lead in the form that will actually be made and sold, without a leader sequence, tag or linker that the product does not have. The CDRs, chains and constructs have to fit together the way the claims say they do, and any nucleotide sequence has to encode the protein it is said to encode. The sequence in the patent should match the one in the priority filing, the INN submission and the regulatory dossier, and the claims need to keep covering that molecule after each amendment during prosecution. For that reason our Sequence Agents are used throughout the life of the patent, and not only at filing.

The limits of WIPO Sequence

Since July 2022, sequence listings have been filed in the ST.26 XML format, usually prepared in WIPO’s own desktop tool, WIPO Sequence. As we have discussed before, it is an improvement on what came before, but it still has real limitations, and in a number of situations sequences still have to be added by hand. The XML files that the patent offices require are also not designed to be read by people, which makes proof-reading a listing against the specification slow and indirect.

Why we use both code and LLMs

A large language model is the wrong tool for copying a sequence. LLMs are probabilistic, so they can drop, add or invent residues in a long string of letters, and they tend to simplify and tidy up text, which is the last thing you want in a sequence listing. As we have argued before, sequence listings should be prepared with code, not with an LLM. Public sequence databases worked out how to handle large volumes of annotated sequence data years ago, and patent practice has been slow to catch up.

Code cannot do the whole job either. Finding every sequence in a patent application means reading it, because sequences turn up in figure legends, tables, claims, examples and the text itself. Whether a sequence needs its own SEQ ID NO, whether a construct counts as a synthetic construct, and whether a claim preamble requires a transmembrane domain that the lead does not have are all questions of understanding the document. That is the kind of reading the latest models now do well.

• Code handles the residues. It builds the ST.26 XML, validates it against the Standard, compares every residue with the specification, checks the sequences against each other and translates coding sequences. The results are exact and can be repeated.
• The LLM handles the reading. It finds every sequence in the application, works out what each one is, reads the claims against the lead candidate and explains each issue in terms the attorney can act on.

Agents that understand biological molecules

Our agents work with the biology behind a sequence and know how the parts of a biologic fit together. An antibody heavy chain is a variable domain followed by constant domains. A CAR runs from a signal peptide through a binding domain, hinge and transmembrane domain to its signalling domains. An mRNA has a cap, untranslated regions, a coding sequence and a poly(A) tail. Knowing this is what allows them to spot a leader sequence left in a mature chain, or a CDR in the wrong place.

Between them, the language model and the code can pick out:

• Molecule type and format. Whether a sequence is DNA, RNA or protein, and whether the molecule is an antibody, scFv, bispecific, T cell engager, CAR, fusion protein, ADC, mRNA, plasmid or viral vector.
• Domains and features. Signal peptides and leader sequences, variable and constant domains, CDRs and framework regions, linkers, hinges, transmembrane and intracellular signalling domains, purification tags and common engineered changes such as Fc mutations.
• Coding relationships. Translation in every reading frame, open reading frames, stop codons in the wrong place, reverse complements, and whether a nucleotide sequence really encodes the protein it is said to encode.
• How sequences relate to each other. Alignments and percentage identity, where a CDR or domain sits within a longer chain, which sequences are exact concatenations of others, and which differ by a single residue.
• Sequence motifs and annotation. Glycosylation motifs, unpaired cysteines, ambiguity codes, modified residues and non-standard nucleotides, and the ST.26 features needed to record them.
• Comparison against a reference. The lead candidate against the claims, the specification against the listing, and the priority application against the later filing.

The language model reasons about the biology, and the residue-level work is done in code that builds on decades of development of bioinformatic sequence analysis tools, so every alignment, translation and comparison is exact and repeatable. The attorney checks every call that matters.

The process also works in reverse. Given an existing listing, whether yours or the other side’s, Darwin validates it, checks it against the application and its bibliographic data, and produces a readable version and a report. A filed listing is part of the application as filed, so it is never edited. Instead, Darwin separates defects in the listing itself from defects in the application that the listing simply reproduces, and says which of them can still be corrected.

Do our Sequence Agents replace attorney review?

Getting sequences right is much easier when your patent attorney is part of the team. As your fractional IP counsel, we work directly with your scientists to check that the sequences in the patent match the clinical lead and the sequences in your regulatory filings, such as the INN submission and the BLA, and that the claims keep covering the right molecule throughout prosecution. Alexion got the eculizumab sequence wrong in both its INN submission and its patent, and a mismatch like that is far cheaper to catch early than to argue about in court.

How our Sequence Agents handle your files and client confidentiality

Our AI runs inside the files and software we and our clients already use, so documents do not need to be copied into and out of a separate piece of software.

Our Sequence Agents are built directly on the latest foundational models rather than on a third-party product, which means we can update them as the models improve. Any Evolve attorney can call on them by name when working on a client project.

As with the rest of our AI, there is no storing of users’ data, no training on inputs and enterprise-grade security.

Evolve’s AI agents for pharmaceutical IP

Our Sequence Agents are part of a wider set of AI agents we have built for pharmaceutical IP, each covering a different part of the work:

Agent	Main use	What it does
🪶 Darwin	Drafting	Pharma and biotech claims, specifications and ST.26 sequence listings, built from the source data and the examples.
🌱 Mendel	Amendments	Jurisdiction-ready claim sets for the EP, the US and elsewhere, with the basis for each amendment set out.
🌍 Popper	Prosecution	Global prosecution and office action responses, argued on the data and on local practice.
🧐 Gilbert	Quality assurance	Independent checking of claims, specifications, sequences and prosecution against the IP strategy.
🧠 Albert	Validity	Basis, novelty, inventive step, support, priority, claim construction and the ranking of experimental data.
🚦 Cassandra	Due diligence	Contract and licence review for pharma and biotech transactions.
💡 Franklin	Disclosures	Publication review, invention spotting and competitive intelligence landscaping.
🦉 Athena	Strategy	The IP strategy on the record, kept current for a case or a whole family.
🌈 Ammonite	Portfolio visualisation	Our own charts for competitive landscaping and for showing an exclusivity position.

Working with Evolve

If you would like to hear more about how we use our Sequence Agents and our other proprietary AI tools, or how Evolve can help you with your IP, please contact one of our attorneys. You can also read the original article on getting patent sequences right.