SeqEvi

SeqEvi: Sequence Evidence is a content-addressed cache for reusable protein sequence annotation evidence.

SeqEvi identifies proteins by canonical sequence content, determines which sequences already have evidence under an exact annotation contract, runs an external annotation tool only for cache misses, and exports an adapter-specific single-file DuckDB result for the current FASTA.

Status

The target architecture and v1.1 result contracts are approved. The SeqEvi 0.2.0 source tree uses strict protein sequence identity, the single-host SQLite/POSIX Store, the external tool runner, exact cache-miss orchestration, and atomic DuckDB result materialization.

SeqEvi 0.2.0 provides managed setup for dbCAN only. The eggnog and interpro-pfam adapters remain supported through explicit runtimes and named host profiles; managed setup for them is later feature work. The Slice D gate record preserves the incomplete original public-user run and its subsequent acceptance decision. In particular, a repeat pull from that run's site is a deferred transport check rather than a 0.2.0 release blocker.

The interpro-pfam and eggNOG-mapper 2.x eggnog adapters are implemented with native-output validation and fixture parity coverage. The eggNOG adapter passes direct parity against eggNOG-mapper 2.1.13 and eggNOG DB 5.0.2. The InterPro adapter passes direct parity against InterProScan 5.77-108.0 with InterPro data 108.0 and Pfam 38.1. The Phase 5 shared Store service, HTTP client, streamed POSIX artifacts, and PostgreSQL persistence are implemented and covered by local/shared plus provisioned PostgreSQL integration tests. Phase 6 resource locks avoid repeated hashing of large immutable database files and provide an explicit full-content verification command. Annotation now uses atomic FASTA staging, file-backed artifacts, bounded Store batches, adapter-native Parquet artifacts, and an operational thread setting.

Why SeqEvi

Two FASTA files do not need to be identical to reuse annotation. If a new FASTA contains sequences seen in earlier projects, SeqEvi reuses the immutable evidence for those sequences and annotates only novel content.

FASTA A: 2000 new sequences       -> annotate 2000
FASTA B: 1000 sequences from A    -> annotate 0
FASTA C: 1000 from A + 500 novel  -> annotate 500

Reuse is exact. Tool runtime, annotation resource, semantic parameters, or adapter contract changes produce a different evidence key and never silently fall back to an older result.

Intended CLI

For repeated use, keep one machine-local TOML per adapter runtime under ${XDG_CONFIG_HOME:-~/.config}/seqevi/profiles/:

seqevi profile init eggnog-5.0.2 --adapter eggnog
seqevi profile init interpro-pfam-38.1 --adapter interpro-pfam
seqevi profile init dbcan-5.2.9 --adapter dbcan-cazyme

Each command creates a complete adapter-specific TOML file and refuses to replace an existing profile. After editing the machine-local paths, inspect and validate profiles without launching either annotation runtime:

seqevi profile list
seqevi profile show eggnog-5.0.2
seqevi profile validate \
  --config "${XDG_CONFIG_HOME:-$HOME/.config}/seqevi/profiles/eggnog-5.0.2.toml"

profile show resolves paths and operational defaults but reports only environment variable names, never their values. The original complete templates remain available through profile example --adapter ADAPTER.

These profile commands configure SeqEvi; they do not install annotation software or databases. Managed setup is available only for dbCAN and uses a runtime image published by SeqEvi. It supports a read-only preview and an explicit apply:

seqevi setup dbcan-cazyme \
  --resource /data/dbcan/db_v5-2-9_5-5-2026/raw \
  --dry-run

seqevi setup dbcan-cazyme \
  --resource /data/dbcan/db_v5-2-9_5-5-2026/raw \
  --dry-run --json

seqevi setup dbcan-cazyme \
  --resource /data/dbcan/db_v5-2-9_5-5-2026/raw \
  --yes

--dry-run never mutates state. --yes pulls the immutable image only when needed, verifies the caller-owned four-file resource, creates seqevi.lock when the resource permits it, runs an ephemeral read-only smoke, and publishes the v2 profile atomically. It never downloads or copies the database. Slice C now dispatches a managed dbCAN annotation through an ephemeral Docker container with the same caller UID/GID, read-only FASTA/resource mounts and a local-Store --network none boundary:

seqevi annotate \
  --profile dbcan-cazyme \
  --store /data/seqevi-store \
  --fasta proteins.fasta \
  --output results/dbcan.duckdb

The dispatcher and cleanup boundary are covered by fixture tests. Real direct-candidate versus managed-v2 scientific equality and later-process replay passed the release gate. A validation harness used an immutable local image ID built from the exact published inputs when site GHCR transport is unavailable; the public setup/profile surface remains pinned to the bundled GHCR digest and exposes no image override.

The real local/shared Store acceptance for eggNOG and InterPro/Pfam is recorded in the result-consumption runtime report. The managed dbCAN distribution gate is tracked in the runtime image release review.

Run repeated annotations by name:

seqevi annotate \
  --profile eggnog-5.0.2 \
  --fasta proteins.fasta \
  --output results/eggnog.duckdb

seqevi annotate \
  --profile interpro-pfam-38.1 \
  --fasta proteins.fasta \
  --store https://seqevi.example.org \
  --output results/pfam.duckdb

seqevi annotate \
  --profile dbcan-5.2.9 \
  --fasta proteins.fasta \
  --output results/dbcan.duckdb

An exact profile file can be selected with --config PATH. Complete explicit mode remains available:

seqevi annotate \
  --adapter eggnog \
  --fasta proteins.fasta \
  --store /data/seqevi-store \
  --output results/eggnog.duckdb \
  --executable /opt/eggnog-mapper/emapper.py \
  --resource /data/eggnog-5.0.2 \
  --threads 8

seqevi annotate \
  --adapter interpro-pfam \
  --fasta proteins.fasta \
  --store https://seqevi.example.org \
  --output results/pfam.duckdb \
  --executable /opt/interproscan/interproscan.sh \
  --resource /data/interproscan-5.77-108.0/data

Shared deployments expose the same Store contract:

seqevi serve \
  --database-url postgresql+psycopg://seqevi@postgres/seqevi \
  --artifacts-dir /data/seqevi-artifacts

The supported user-systemd deployment through the host rootful Docker daemon is documented in the service runbook. The service image contains SeqEvi and its server dependencies only; annotation executables and databases remain external.

Initialize or audit a database resource lock independently of annotation:

seqevi resource verify \
  --adapter eggnog \
  --executable /opt/eggnog-mapper/emapper.py \
  --resource /data/eggnog-5.0.2

V1 Scope

• Protein FASTA input with strict, deterministic canonicalization.
• GA4GH SQ. sequence identifiers plus MD5 compatibility aliases.
• Exact, immutable evidence keys.
• Explicit eggnog, interpro-pfam, and official-runtime-validated dbcan-cazyme adapters. dbCAN direct/local/shared scientific acceptance is complete; publishing the managed runtime image remains separate work, and annotation databases remain caller supplied.
• Local SQLite/POSIX Store and shared PostgreSQL/POSIX Store service.
• One self-describing DuckDB result per invocation; adapter-native normalized evidence remains Parquet inside the incremental Store.

SeqEvi does not infer species, manage projects, schedule workflows, install third-party tools, distribute annotation databases, or merge unrelated adapter schemas.

Documentation

Start with the documentation index.

• Architecture overview
• Sequence and evidence contract
• Adapter contract
• Result consumption contract
• Execution profile contract
• Storage and deployment architecture
• MVP implementation plan
• Execution profile implementation plan
• Validation strategy
• Annotate runtime and bounded-memory plan
• Bounded-memory and operational performance
• InterProScan Pfam runtime validation
• dbCAN CAZyme adapter implementation plan
• DuckDB result-consumption runtime acceptance
• dbCAN runtime image release review

Accepted managed-boundary documents; Slice B setup and smoke plus Slice C OCI execution and real candidate acceptance are implemented, while v1 profiles remain compatible:

• Managed adapter onboarding roadmap v1.1
• Managed-distribution architecture v1.2
• Execution profile v2.2 contract

Python And Result Discovery

The public Python API returns DuckDB's native relation, so the same object can be queried from a notebook or passed to Arrow/Polars without a SeqEvi wrapper:

import seqevi

annotations = seqevi.annotate(
    "proteins.faa",
    profile="interpro-pfam-38.1",
    output="results/pfam.duckdb",
)
print(annotations.columns)
pfam = annotations.select("InputID", "SignatureAccession")

An existing result can be opened read-only with seqevi.scan_annotations(). If the adapter columns are not known in advance, inspect the native relation or the stable catalog first:

annotations = seqevi.scan_annotations("results/pfam.duckdb")
print(annotations.columns)
print(annotations.pl(lazy=True).collect_schema())

The normal protein-level join key is InputID. SequenceID is the content identity used for exact Store reuse. InterPro/Pfam keeps one-to-many domain rows, so aggregate it before joining to a one-row-per-protein table when that is the desired grain. SQL, R, and workflow tasks can open the same file and query main.annotations; _seqevi.column_info, _seqevi.table_info, and _seqevi.metadata provide column descriptions, row grain, and provenance.

SeqEvi 0.2.0 is a deliberate output cutover from the 0.1.0 directory Data Package. Existing 0.1.0 packages remain readable by their own Data Package tools, but new SeqEvi invocations publish DuckDB only; rerun an annotation to produce the new result file.

External Tools

Annotation runtimes and databases are supplied by the user. The current CLI has no seqevi setup command. SeqEvi v1 targets eggNOG-mapper and InterProScan with the Pfam application, and dbCAN for protein-level CAZyme annotation. A future managed path may publish a SeqEvi-maintained runtime image built from locked upstream inputs after runtime compliance review; it would not be an upstream-official image. Annotation databases remain separate and are never bundled in the wheel or runtime image. The proposed managed path uses a public, digest-pinned ghcr.io/fuqingzh/seqevi-dbcan runtime package; callers continue to provide the database path, and internal registry mirrors remain deployment policy.

License

SeqEvi is distributed under the MIT License.