Pre-releaseDemonstrations on request
SI-assisted biosequence search for patent examination
Biosequence prior-art search you can defend.
Seekquence reaches the distant protein relatives that alignment alone misses, then holds every AI-found candidate to alignment before it counts as evidence. Every hit says how it was found. Every search can be replayed. The examiner decides.
Synthetic case: ranked references
References, highest priority first
| identical. similarblank: different
Identity first
Exact and reverse-complement matches are found before any AI runs.
Evidence on every hit
Each hit carries a class, from E0 exact identity to E7 unverified.
Pinned and replayable
Every search is tied to an immutable corpus snapshot and a manifest.
In your AWS account
Built from AWS managed services. We host none of it.
The problem
Similarity finds. Alignment proves. Examination needs both.
Every biotechnology application raises the same two questions about its sequences: what in the prior art is related, and can you show it?
Alignment answers the second question, but on its own it misses distant protein relatives whose sequences have drifted apart while their function has not. Protein language models reach those relatives, but a similarity score is a ranking signal, not evidence.
Seekquence runs each method for what it does well, in a fixed order, and records which one supports every hit.
| Method | Identical sequences | Close relatives | Distant relatives |
|---|---|---|---|
| Identity lookup | Proves | Not its job | Not its job |
| Alignment search alone | Proves | Proves | Often misses |
| Language-model retrieval alone | Finds, unproven | Finds, unproven | Finds, unproven |
| Seekquence, in order | E0 Identity | E1 Aligned | E3 once alignedE7 if not |
How it works
AI narrows. Alignment proves. The examiner decides.
Separate pipelines for nucleotides and proteins, each in a fixed order. Every stage is versioned, and every search records which versions ran.
- 01
Identity first
Query
5′ ATGGCGTTCAAG 3′
┆┆┆┆┆┆┆┆┆┆┆┆
3′ TACCGCAAGTTC 5′
Other strand
Stored record
5′ CTTGAACGCCAT 3′
E0 Reverse-complement identityAn exact and reverse-complement identity lane runs before any AI. A stored sequence identical to the query is listed as E0, and is never hidden or cut from the results.
- 02
AI narrows
Query Candidates Stored proteinsProtein queries go straight to retrieval: ESM-2 protein language-model embeddings, searched in an HNSW vector index, reach relatives that share little sequence. Nucleotide queries take a deterministic route on both strands, and coding sequences also reach the protein index through six-frame translation.
- 03
Alignment proves
Diagonal: the alignment
Step: a gap
E1 AlignedE3 Remote homologyCandidates are aligned with MMseqs2 on GPUs, protein hits are re-ranked by Smith-Waterman, and profile hidden Markov models add domain evidence. A candidate found by AI counts as evidence only after alignment confirms it.
- 04
The examiner decides
E0 hitE1 hitE1 hitE3 hitPatent A
Reference
Publication B
Reference
Hits roll up to the patents, publications and records their sequences occur in, ordered by a documented rule. Every acceptance, rejection and note is the examiner's, and each one is recorded.
Two depths
Interactive, for triage. Interactive searches target under two minutes.
Deep, for the record. Deep searches target a 24-hour service level.
Evidence classes
Every hit says how it was found.
A similarity score ranks candidates; it does not prove anything. So each hit carries an evidence class that tells the examiner how it was found and what supports it, from exact identity to AI similarity that alignment has not confirmed.
E4 is reserved for structure evidence, a research channel that is not part of the shipped search path.
- E0 Exact identityE0 Reverse-complement identity
The stored sequence is identical to the query, on either strand.
Found by: Exact identity lane, before any AI
- E1 Aligned
An alignment supports the hit, with its identity, coverage and E-value.
Found by: Alignment search
- E2 Profile or domain
A profile hidden Markov model matches the hit: a protein domain or an examiner-built profile.
Found by: Profile search
- E3 Remote homology
Found by AI retrieval and verified by alignment.
Found by: AI similarity retrieval, then alignment
AI similarity only, not aligned. A candidate that retrieval found and alignment has not confirmed is shown apart from the evidence, with no alignment numbers, and is never counted as a match.
Found by: AI similarity retrieval
The workbench
From sequence listing to search record, in one case.
Seekquence is built around the examiner's case: claims, specification and WIPO ST.26 sequence listing together. SI proposes. Versioned code checks every proposal. The examiner accepts, rejects or revises, and every step is kept.
Synthetic case: claim features
1. A polypeptide comprising SEQ ID NO: 4.
2. A fragment of claim 1 comprising residues 20 to 80 of SEQ ID NO: 4.
3. A variant of claim 1 with conservative substitutions.
- Accepted
Recited sequence
SEQ ID NO: 4, claim 1
- Proposed by SI
Fragment
Residues 20 to 80 of SEQ ID NO: 4, claim 2
Quote found in claimAcceptRejectRevise - Proposed by SI
Variant
Conservative substitutions, claim 3
Quote found in claimAcceptRejectRevise
01
Case intake
Claims, specification and the ST.26 listing become one search-ready case. Text is extracted once and kept with a SHA-256 fingerprint.
02
Claim features
SI proposes recited sequences, identity floors, fragments and variants. Code checks each one against the claim text before it is shown.
03
Search plan
A rule engine, with no language model, recommends the searches for each accepted feature and shows the rule that fired.
04
Query grouping
Searches are grouped by claim, and deep searches beyond your limit move to later batches.
05
Prioritized results
Identity, alignment and AI-retrieval lanes stay apart. References are ordered by a documented rule.
06
Reference roll-up
Every listed hit rolls up to the patents, publications, applications and database records it occurs in.
07
Cross-file synthesis
One roll-up covers every search of the case, at both depths, and imported GenCore 6 or BLAST result files.
08
Search summary
An editable summary is composed from the stored record. Every paragraph outside your own notes cites what it rests on.
SI, held to account
Assistance that shows its work.
A language model is used in three places: to propose claim features, to phrase a short explanation for each reference, and to write each search report's narrative. Code checks the first two before they are shown. The third is labeled as AI-generated advisory text. All three are stored, and replay reuses them.
- model
- allowlisted, via Amazon Bedrock
- prompt
- sha256 3b9c41…e1a0
- response
- sha256 a41f0c…77d2
- residues
- none sent
checks
- Quote found at the cited claim offsets
- Cites its own claim or a parent claim
- SEQ ID NO exists in the listing
- States no anticipation, obviousness or novelty
Proposals, never findings
Code refuses any proposed feature or explanation that states anticipation, obviousness or novelty.
No sequences in the prompt
The model reads claim and specification text. It never receives a residue.
Every call on the record
Each response is stored with its model id, a hash of the prompt and a SHA-256 of the response.
Determinations by code
A report's determination is computed by code. The narrative beside it is labeled as AI-generated, with its model id.
One gateway, inside your network
Language-model calls go only through Amazon Bedrock, over a VPC endpoint, to models on the deployment's allowlist.
A planner with no model
The search planner is a rule engine: the same input gives the same plan, byte for byte.
- snapshot
- corpus-2026-11-07, pinned
- pipeline
- protein, interactive
- stages
- identity, retrieval, alignment, evidence
- embedding
- esm2-650m, pinned revision
- rules
- planner_v1, ref_priority_v1
- model calls
- stored with hashes
Reproducible by design
Run it again. Get the same record.
Retrieval, validation, roll-up, coverage, priority, clustering and export are deterministic: the same inputs give the same outputs.
Immutable snapshots
Every search is pinned to a corpus snapshot that never changes underneath it.
Search manifests
Each search records what was searched, against which snapshot, with which settings and versions.
Replay
Replay runs the search again against the same snapshot, reusing the stored model responses.
Complete result sets
Complete result sets are kept and every hit is paged, so nothing is dropped silently.
Measured
Results we quote with the corpus named.
Each figure comes from our own benchmark harness, on the corpus stated. These are retrieval-stage measurements; end-to-end recall at production scale is not yet measured.
Protein family remote homologs
Pfam clan remote homologs, 500 sequences, August 2026.
Recall in the top 100, out of 1
Language-model retrieval: 0.580
Alignment search alone: 0.030
Pfam clans group protein families that share an evolutionary origin even where their sequences have diverged. Retrieval finds those candidates; alignment confirms them before they count as evidence.
Reverse-complement matches
GenBank patent nucleotide corpus, 6,899 records, 100 reverse-complement queries, September 2026.
Matches found, out of 100
Both strands, as shipped: 100
Forward strand only: 4
Why Seekquence searches both strands of every nucleotide query, with no length cap.
Deployment
Runs in your AWS account. We host none of it.
Seekquence is assembled from AWS managed services and deployed by code into the customer's own AWS account. AWS maintains the services; every Seekquence release is versioned code.
Versioned releases
Engine versions are pinned in code, and a release is applied with
pulumi up.Your identity provider
Sign-in federates to your identity provider by OIDC or SAML. Roles come from your groups, and federation never grants admin.
Closed to the internet
Model endpoints, rule engines and the pipeline have no internet ingress. People reach Seekquence through the authenticated interface and API.
Your data stays yours
You own your data, your examiners' decisions and feedback, and any model weights fine-tuned on them.
Your AWS account
private VPC, no internet ingress to the pipeline
Search
- AWS Step Functions
- AWS Batch on GPU
- Amazon OpenSearch Service, k-NN
Embeddings
- Amazon SageMaker, self-hosted ESM-2
Language model
- Amazon Bedrock, allowlisted models, VPC endpoint
Records
- Amazon Aurora PostgreSQL
- Amazon S3, pinned snapshots and complete result sets
Sign-in
- Amazon Cognito, federated to your identity provider
Operations
- Amazon CloudWatch
- AWS KMS
- AWS Secrets Manager
Defined in Pulumi code. Released with pulumi up. Nothing changed by hand.
Boundary and accessibility
- Production deployment: designed for AWS GovCloud (US) inside a customer's FedRAMP High boundary; not yet deployed there.
- Seekquence holds no FedRAMP authorization of its own.
- The examiner interface is designed and tested to WCAG 2.2 AA.
Questions
Straight answers.
Can we use Seekquence today?
Seekquence is pre-release: built and tested, and not yet in production. Demonstrations on a synthetic case are available on request, and a trial sandbox on public and synthetic data is planned. The roadmap lists what comes next.
Does the AI decide anything?
No. SI proposes, versioned code validates, and the examiner decides. Code refuses any proposal that states anticipation, obviousness or novelty, and the determination in each search report is computed by code.
Which models does it use?
ESM-2 650M, an MIT-licensed protein language model, self-hosted on Amazon SageMaker in your account, for protein embeddings. Language-model calls go through Amazon Bedrock to the models on the deployment's allowlist. A nucleotide foundation model and structure-based retrieval are research channels, not part of the shipped search path.
How are nucleotide sequences searched?
Nucleotide prior art is an identity-class problem, so nucleotides take their own deterministic pipeline: an identity lane on both strands, then MMseqs2 candidate search on both strands with no length cap. Coding sequences also reach the protein index through six-frame translation.
Can it read our existing search results?
Yes. GenCore 6 result files and BLAST tabular output can be imported into a case. Imported rows are listed as reported, rolled up with the case's own searches, and not aligned again.
Where are the limits?
Each limit is stated where it applies, in the product and in the documentation. For example, a very short nucleotide query is not found inside a longer record, though an identical record is still listed, and the search form says so before the search runs.
How is Seekquence priced?
A tiered subscription, plus metered compute for deep searches. Prices are not yet published; a price schedule is planned. The pricing page describes the model.
Who owns the data?
You do. The core product is licensed commercially; a government customer receives Government-Purpose Rights in the tailoring made for it. The customer owns its data, its examiners' decisions and feedback, and any model weights fine-tuned on them.
See it work
See a case worked from listing to record.
A live walkthrough on a synthetic case: claim features, the search plan, evidence-classed results, references and the search summary.