Skip to content
Tech Interview Prep
Technical interview guide

Biological Databases & Ontologies

The major public repositories a computational biologist works against daily (NCBI/GenBank, Ensembl, UniProt, and Gene Ontology) what each holds and how they cross-reference each other.

Read
49 min
Practice MCQs
25
Interview Q&A
25
Edition
v3
Editorial status
Reviewed

Scope: Current NCBI E-utilities and data policies; Ensembl release 116; UniProt, RCSB PDB, Gene Ontology, OBO Foundry, Sequence Ontology, HGNC, Identifiers.org and OLS guidance reviewed 2026-09-04.

Curated: · Written: · Reviewed:

25 questions

MCQ-2A service ingests identifiers from three source databases and joins on the bare string. The string 'ABC123' from source A denotes a gene; the same string from source B denotes a protein. The join returns the wrong record type. Which property of identifiers does this failure expose?
MCQ-6A pipeline stores variant calls as `chr7:140453136:A>T` with no other fields. Six months later a collaborator re-runs the same call set against GRCh38 and reports that a subset of the positions land on different genes than your GRCh37-based report. What single addition to each stored record would have prevented the disagreement?
MCQ-3A variant pipeline caches calls keyed on the base accession only. The provider re-releases NC_000001.10 as NC_000001.11 (hypothetical coordinates): 4 bases are deleted at position 100,000 and 2 bases are inserted between positions 120,000 and 120,001. The pipeline serves a cached call for a variant at position 150,000 against the new release. Which base does it read?
MCQ-5A clinical variant pipeline ingests gene annotations from three sources. Source A reports approved HGNC symbols, source B reports aliases, and source C reports symbols from a 2019 HGNC release. The database uses the symbol as the primary key. After a quarterly HGNC update, 1,842 variant records fail to join to gene metadata: 1,793 symbols changed, 37 aliases now match a different gene's approved symbol, and 12 symbols were withdrawn. The team needs joins to stay stable across future updates. What should they do?
MCQ-1A lab writes every submission to `raw_submissions` (append-only: accession_no, submitter, received_at, raw_value, raw_unit). A nightly job normalizes units and upserts one row per (site, analyte, sample_date) into `curated_results`, overwriting whatever was there. A report cites 12.6 mg/L nitrate for site 7 on 2024-03-11. The submitter's retained copy of accession 88431 says 1.26 mg/L, and `raw_submissions` agrees; `curated_results` holds only 12.6. Which change makes the discrepancy resolvable?
MCQ-9A variant curation pipeline re-runs nightly. On day 1, sample S-1042 gets annotation 'pathogenic' from ClinVar release 2024-03 with evidence PS3. On day 30, a re-run overwrites the row with 'likely pathogenic' from release 2024-06 and no evidence code. An auditor asks why the day-1 report said pathogenic. The pipeline currently stores only the latest label and source release. Which change makes that audit answerable without rerunning old pipelines?
MCQ-8A curator audits an annotation in a GAF file: gene RAD52, GO term 'DNA repair' (GO:0006281), evidence code IDA, reference a paper whose only relevant experiment is a plate assay showing that the rad52 deletion strain is sensitive to a DNA-damaging agent while the wild-type strain is not. What should the curator do with the evidence code?
MCQ-7A protein knowledgebase record carries a curator-reviewed badge. The record has 12 annotations: 5 from direct experiments, 4 inferred by the curator from sequence similarity, and 3 imported automatically from another database. A downstream classifier treats every annotation on a reviewed record as experimentally validated and trains on all 12 as positives. What is the correct diagnosis and fix?
MCQ-17A curator reviews an assay in which a protein carrying a kinase domain shows no detectable phosphotransferase activity against any of 40 tested substrates, while a positive-control kinase phosphorylates them normally. The curator wants to record this result in GO. Which annotation is correct?
MCQ-18A GO annotation service holds 40,000 direct annotations, each carrying a PMID, and a nightly materialized ancestor closure of 1.2M rows. Some entity-term pairs carry two direct annotations with different PMIDs. A count query for GO:0006955 returns 3,100 genes against 210 direct annotations, and a curator auditing 'gene A annotated to GO:0006955' cannot see which paper supports it. The count query must stay under 200 ms; expanding ancestors with a recursive CTE over the direct table takes 4.5 s. Which storage design keeps the 3,100-hit retrieval, restores the evidence trail, and avoids double counting?
MCQ-11A gene-annotation service stores two edge types in the same graph: `is_a` (a child term is a subtype of a parent term, e.g. `kinase activity is_a catalytic activity`) and `part_of` (a child term is a component of a parent term, e.g. `mitochondrial membrane part_of mitochondrion`). A query walks `is_a` edges transitively to collect all ancestors of a term, then walks `part_of` edges transitively from those ancestors to collect their components. A curator reports that querying `mitochondrial membrane` returns `catalytic activity` as an ancestor, which is wrong. Inspection shows the term has exactly one outgoing edge: `mitochondrial membrane part_of mitochondrion`. What is the most likely cause?
MCQ-13A loader walks every edge in the GO graph and emits an inferred annotation for each term it can reach from a direct annotation. Gene product G has exactly one direct annotation: 'mitochondrial ATP synthase complex'. The graph contains these edges: 'mitochondrial ATP synthase complex' is_a 'ATP synthase complex'; 'mitochondrial ATP synthase complex' part_of 'mitochondrion'; 'mitochondrion' has_part 'mitochondrial ribosome'. The loader currently emits all three reachable terms. Which inferred set is correct?
MCQ-12A pipeline propagates Gene Ontology annotations up the hierarchy to find enriched terms. It walks parent links from each annotated term, and the same ancestor keeps being reached by different routes: for example, the term 'mitochondrial ribosome' is reached both through 'mitochondrial large ribosomal subunit' and through 'organelle ribosome'. The engineer's first version counts each arrival separately, so a term with three parent paths gets its evidence counted three times, and the enrichment p-values come out too small. Which change to the traversal fixes the over-counting while still letting every ancestor be reached?
MCQ-19A service scores GO term pairs with Resnik similarity: sim(t1, t2) = max IC(a) over the common ancestors a of t1 and t2, where IC(a) = -log2 p(a) and p(a) is the fraction of the corpus's gene products annotated to a or to any of its descendants. The corpus holds 10,000 gene products. For the pair (t1, t2) the only common ancestors are A and B: | term | gene products annotated to it (directly or via descendants) | | --- | --- | | A | 5,000 | | B | 250 | | t1 | 40 | | t2 | 100 | What similarity does the service report for (t1, t2)?
MCQ-10A variant-annotation service stores each annotation as (sample_id, term_label, term_namespace) and resolves the term by joining term_label to a lookup table rebuilt from each Gene Ontology release. Between the 2023-05 and 2024-01 releases, 412 terms were obsoleted and their labels reassigned to other terms, and 1,180 labels were edited in place. After the upgrade, 6,300 of 40,000 annotations join to the wrong term and 900 fail to join at all. Which change removes both failure modes?
MCQ-14A clinical data platform stores records spanning 2013–2024, recorded in SNOMED CT, for longitudinal studies that count distinct patients per concept per year. In the 2024-01 international release, concept 123456789 (used in 38,412 rows) is inactive. Its historical association file lists two targets, 111111 and 222222, both typed POSSIBLY EQUIVALENT TO, and no REPLACED BY. The ETL currently joins on the concept's preferred label. What should the pipeline do with the 38,412 rows?
MCQ-16A knowledge graph ingests two sources. Source A contributes 40,000 HGNC-to-Ensembl mappings, all curated as exact. Source B contributes 900,000 pathway-to-gene xrefs; 12% are labeled broad or related, the rest are unspecified. The loader writes all 940,000 as owl:sameAs. A target-expansion query for imatinib returns 47 genes; the curated target list has 3. Which change fixes the query without discarding Source B?
MCQ-15A hypothetical product search team maps catalog terms to an ontology that records synonym scope. For the class `laptop`, `notebook computer` is an exact synonym and `computer` is a broad synonym. The index has 120,000 products. A query for `laptop` returns 8,400 products, and 18 of the top 20 are laptops. If the team expands the query with both synonyms as equal replacements, the result set grows to 31,000 products and only 6 of the top 20 are laptops. The team wants to keep the extra recall from the broad synonym without losing top-rank precision. What should it do?
MCQ-4A pipeline joins a per-gene expression table (one row per Ensembl gene ID) to a per-protein abundance table (one row per UniProt accession) on a shared gene symbol column. After the join, the row count is 3.4x the gene table's row count, and summing the joined abundance column gives a total 3.4x larger than summing the protein table directly. The team's dashboard reports per-gene totals. What is the correct fix?
MCQ-20A nightly sync calls `GET /orders?limit=100`, which returns `{"items": [...], "next_cursor": "..."}`, and writes `len(items)` to a row-count metric. The metric reports exactly 100 every night; the source table holds 47,000 rows. The job logs no errors and exits 0. What is the most likely cause?
MCQ-21A nightly sync job calls a vendor's REST API and reads the response headers on every call. It sees `X-RateLimit-Limit: 600`, `X-RateLimit-Remaining: 0`, and `X-RateLimit-Reset: 30` (seconds until the window resets). The job currently fires its next batch immediately, gets HTTP 429 with `Retry-After: 30`, and retries in a tight loop until the window resets. The vendor's docs say the limit is enforced per API key across all of that key's clients. What change makes the job stop generating 429s without slowing the sync more than necessary?
MCQ-22A team's nightly job pages a live REST API: 1,000 records per page, ~4.2M records, about 6 hours per run, and writes results to a warehouse. Two runs last month disagreed by 3.1% on the same metric, and one run double-counted ~1,800 records after a page boundary shifted mid-pagination. Raw API responses are cached for 30 days. A regulator asks the team to reproduce a figure published three months ago. The provider also publishes a monthly bulk export: one gzipped CSV, ~2.1 GB, with a SHA-256 checksum and a license file. What should the team do to reproduce the figure?
MCQ-24A production knowledge snapshot is served from a content-addressed store. A source release adds 1,200 entities, renames 3 relation types, and drops 2 sentinel records that downstream regression queries assert. The staged snapshot passes schema validation and its entity count matches the source. What should the team do before activating it?
MCQ-23A data platform ingests 4.1M images from 3,000 public datasets. The ingestion job records source_url, retrieval_date, and the SPDX license id for each dataset, then writes a combined training manifest. One dataset is CC BY-NC 4.0 and contributes 120,000 images; another is CC BY-SA 4.0 and contributes 700,000 images. The team will train a model, sell it as an API, and publish a derived text corpus built from the CC BY-SA dataset's captions. What should the ingestion job do to the manifest?
MCQ-25A team builds a nightly target–disease graph by joining pinned releases of three governed sources. The build is green: no parse errors, no empty labels, loader exit 0. Curation then finds two defect classes in the 4.6M-edge graph: 9,730 edges whose disease endpoint the pinned release marks obsolete (each with a replacement ID), and 41,000 edges whose compound mapping carries a license that is not on the redistribution allowlist. Of the 9,730 obsolete endpoints, 1,240 point to terms the newest release has reinstated as active. Which gate would have failed the build on both defect classes?