Seekquence

Seekquence is a deterministic, SI-assisted biosequence search workbench for patent examination. An examiner brings an application's claims, specification and sequence listing. Seekquence finds the stored nucleotide and protein sequences that match the listed sequences, shows the evidence behind each match, and rolls the matches up to the patents and publications they come from.

Status: pre-release, built and tested; not yet in production. Seekquence is not yet sold and has no customers. This site documents version 0.1.0.

How it works

SI proposes, versioned code validates, the examiner decides.

The assistance methods are listed below. Each one proposes; code checks what it proposes before anything is stored or shown; the examiner accepts, rejects or revises it. Code refuses a proposed claim feature, a reference explanation or a summary narrative sentence that states anticipation, obviousness or novelty.

Retrieval, validation, reference roll-up, claim coverage, priority, clustering and export are deterministic: the same inputs give the same outputs.
Every search is pinned to an immutable corpus snapshot and described by a search manifest that can be replayed.
A language model is used in four places: to propose claim features, to phrase a short explanation for each reference, to draft a narrative of the search summary when the examiner asks for one, and to write the narrative of each search's examiner report. Code checks the first three before they are shown, and a summary narrative enters the record only when the examiner applies it. The report narrative is labeled as AI-generated advisory text and carries the model id; the determination beside it is computed by code.
Each model response is stored with a hash of the prompt, the model id and a SHA-256 of the response, and replay reuses the stored response.
Every hit carries an evidence class, from E0 (exact or reverse-complement identity) to E7 (AI similarity only, not aligned). A candidate found by AI retrieval counts as evidence only after alignment confirms it. The classes are described in Results and evidence classes.

What the examiner can do

The eight examination stages, what is built for each, and its limits
Stage	What is built	Limits
Case summarization	A case holds the claims, the specification and the WIPO ST.26 sequence listing as one structured, search-ready record. Text is extracted once and kept with a SHA-256 fingerprint.	No narrative synopsis of the specification is written. ST.25 listings and image-only PDFs are not read.
Feature extraction	A language model proposes claim features: recited sequences, identity floors, fragments, variants and other limitations. Code checks each one against the stored claim text, and the examiner accepts, rejects or revises it.	A proposal that fails a check is never shown as a feature.
Strategy generation	A rule-based planner recommends the searches to run for each accepted feature, with the rule that fired and a completion estimate. No language model is used.	The planner does not choose databases.
Query grouping	Searches are grouped by claim. Deep searches past the examiner's limit move to later batches, or into a text-only request draft for a specialist search team.	Sending and tracking that request is planned.
Result prioritization	Results keep the identity, alignment and AI-retrieval lanes apart. References are ordered by a documented rule and can be grouped by how alike their matched sequences are.	The order is a triage aid. No gain in examiner speed or accuracy is claimed.
Reference mapping	Every listed hit is rolled up to the patents, publications, applications and database records its sequence occurs in.	The roll-up reads the rows each result lists, and says how many of the stored rows those are.
Cross-file synthesis	One roll-up covers every search of a case, at both depths, across databases and imported legacy result files. A language model phrases one explanation per reference; code checks every sentence and writes its own sentence when a draft fails.	Imported rows are listed as reported. They are not aligned again.
Documentation drafting	Code composes an editable search summary from the stored record. Every paragraph outside the examiner's own notes cites what it rests on, and the examiner edits the summary in new versions. On request, a language model drafts a narrative of the summary; code checks each sentence, and the narrative enters the record only when the examiner applies it.	The checks on a drafted narrative do not compare a sentence's meaning with the statements it cites; each sentence is shown with those statements for the examiner to read.

The documentation describes each step and its limits. The view below is a case's References region, where every hit of the case is rolled up to the references it comes from.

The References region of a case: references ranked by rule, with claims met, best evidence class, highest Patent Significance Score and decision, and the first reference open beside the list with its kind, dates, hits, why it ranks there, its claim coverage and the hits behind it. — A case's references, ordered by a documented rule. Screenshot of the product's test fixtures: placeholder records and patent numbers, not search results.

The methods behind it

Protein language-model embeddings (ESM-2 650M, self-hosted) for AI similarity retrieval and for grouping references.
Approximate nearest-neighbor vector search (HNSW) in Amazon OpenSearch Service.
Alignment with MMseqs2, and Smith-Waterman re-ranking with parasail for proteins. Nucleotide candidates come from MMseqs2 on both strands, with no length cap.
Profile hidden Markov models (HMMER) for protein domain evidence.
Structured extraction by a language model, validated by deterministic rules.
A knowledge-based planner: a rule engine that explains every recommendation.
Clustering of references in embedding space, by a fixed algorithm and threshold.
Claim-coverage rules and a lexicographic priority order.
Explanations phrased by a language model from facts that code established, checked sentence by sentence.
Rule-based follow-up suggestions.
A fine-tuning registry with benchmark-gated promotion. It is manual, protein only, and not used on the shipped search path.

A dataset of examiner decisions on closed cases, and prior-case context for planning, are built. A learned ordering of references is built but switched off and kept outside the evidence chain; it has no measured benefit. Structure-space retrieval and a nucleotide foundation model are research channels, not part of the shipped search path.

Built on AWS managed services

Seekquence is assembled from AWS managed services and deployed by code into the customer's own AWS account. OneGlobe LLC hosts no part of it.

AWS services and what each does in Seekquence
Service	Role
Amazon Bedrock	The only language-model path: one gateway, a VPC endpoint, and an allowlist of models.
Amazon OpenSearch Service	Protein vector search with the k-NN plugin (FAISS HNSW), inside the VPC.
Amazon SageMaker	The self-hosted ESM-2 embedding endpoint, and batch embedding of the corpus.
AWS Batch	Alignment, HMMER and deep-search shards on GPU instances.
AWS Step Functions	One state-machine execution per search, recorded in its manifest.
AWS Lambda and Amazon ECS on Fargate	Pipeline stages, the API, the user interface and in-VPC operator tasks.
Amazon API Gateway and Application Load Balancers	The JWT-authorized API, through a VPC link to an internal load balancer; the user interface behind an ingress allowlist.
AWS WAF	AWS managed rule groups on the user interface's load balancer.
Amazon Aurora PostgreSQL	Cases, sequences, occurrences, decisions and the application audit log.
Amazon S3	Immutable corpus snapshots and complete result sets, versioned and encrypted.
Amazon Cognito	Sign-in, federated to the customer's identity provider.
AWS Secrets Manager, AWS Systems Manager Parameter Store and AWS KMS	Secrets, stack configuration, and encryption with AWS-managed keys.
Amazon CloudWatch	Logs and alarms for every tier.
Amazon VPC, Amazon ECR, Amazon SNS and Amazon SQS	Private networking with interface endpoints, container images, and search-completion notifications.
AWS Certificate Manager and Amazon Route 53	Certificates and DNS.

The whole stack is Pulumi code. Engine versions are pinned in code, a release is applied with pulumi up, and nothing is changed by hand in the console.

Deployment

Seekquence runs in the customer's own AWS account. 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. A demonstration stack is defined in code and deployed for each demonstration window in a commercial AWS Region (us-east-1), with public data only. No production deployment exists, and Seekquence has no customers yet. The security page describes the design in more detail.

Licensing and data rights

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. Public databases stay under their own licenses, listed on the legal page. The pricing page describes how Seekquence is priced.

Request a demonstration

Request a demonstration by email. A trial sandbox is planned; the roadmap lists it with the other planned work.