Science & Research

Bioinformatics interview questions

Bioinformatics interviews test statistical judgment far more than tool trivia. Interviewers want to hear how you choose an aligner for the data in front of you, how you decide which variants or differentially expressed genes you believe, how you handle multiple testing and batch effects, and whether someone else could rerun your analysis next year and get the same numbers.

9 questions (2 easy · 4 medium · 3 hard), each with what a strong answer covers and where people lose the point. Free to read, no account.

On this page (9 questions)

1.What is the difference between local and global sequence alignment, and when would you use each?

Warm-up

What a strong answer covers

  • A global alignment (the Needleman-Wunsch formulation) aligns two sequences end to end and forces every residue into the alignment, which is appropriate when the sequences are expected to be homologous over their whole length and of similar size.
  • A local alignment (the Smith-Waterman formulation) finds the highest scoring subsequence pair and ignores the rest, which is what you want when a short read or a conserved domain sits inside a much longer sequence.
  • BLAST is a heuristic local search: it seeds on short exact word matches and extends them, which trades guaranteed optimality for the speed needed to search a whole database.
  • A strong answer connects the choice to a real task: comparing two orthologous proteins of similar length is a global problem, finding a domain in an unannotated ORF or searching a database is a local one.
  • Mentioning that scoring depends on the substitution matrix and gap penalties, and that those choices change which alignment wins, signals real experience.

Where people lose the point

  • Describing BLAST as a global aligner, or treating BLAST and Smith-Waterman as interchangeable rather than heuristic versus exhaustive.
  • Ignoring gap penalties and substitution matrices entirely, as though alignment were a parameter-free operation.
  • Reading a low BLAST E-value as proof of function rather than as a statement about the chance of seeing that score by chance in a database of that size.
Link to this question

2.You open the QC report for a fresh sequencing run. What do you look at, and what would make you stop before aligning?

Warm-up

What a strong answer covers

  • Start with per-base quality across the read: a gentle decline toward the 3' end is normal, a collapse partway through the read is not, and it changes whether you trim or rerun.
  • Check adapter content and overrepresented sequences. Adapter read-through indicates short inserts, and heavy overrepresentation of one sequence often points at rRNA carryover, a primer dimer or a contaminant.
  • Check GC distribution against the expected profile for the organism and library type. A second peak usually means contamination or an unexpected species in the sample.
  • Check duplication level in the context of the assay: high duplication in a low-input or heavily amplified library can be expected, while the same figure from a high-input DNA library suggests you sequenced very little complexity.
  • A strong answer names the stop conditions explicitly: a species or index mismatch, a run where quality falls off before the region you need, or read counts far below what the experimental design was powered for.
  • Mentioning that QC continues after alignment (mapping rate, insert size distribution, coverage uniformity, duplicate rate) shows you do not treat pre-alignment QC as the whole story.

Where people lose the point

  • Treating a red flag in an automated QC report as a pass or fail verdict rather than as something to interpret against the library type.
  • Trimming aggressively by reflex, which shortens reads and can bias alignment more than the low quality bases would have.
  • Skipping the check for the wrong organism or a swapped index, which is the failure that wastes the most downstream time.
Link to this question

3.How do you choose an aligner, and why can't you use the same one for DNA and RNA reads?

Core

What a strong answer covers

  • The first question is whether reads can span introns. RNA reads from a eukaryote cross splice junctions, so they need a splice-aware aligner such as STAR or HISAT2; a DNA aligner will soft-clip or misplace those reads.
  • The second is read length and error profile. Short-read DNA alignment is well served by BWA-MEM or Bowtie2, while long noisy reads from long-read platforms need an aligner built for them, such as minimap2.
  • The third is what you will do downstream. Variant calling wants accurate placement and honest mapping quality; quantification can sometimes skip alignment entirely and use a lightweight mapping approach such as a selective alignment or pseudoalignment quantifier.
  • Speed and memory are real constraints, not cosmetic ones: a suffix-array aligner that needs tens of gigabytes of RAM per instance is a scheduling decision as much as a scientific one.
  • A strong answer notes that mapping quality is the aligner's own statement of confidence, and that filtering on it is how multi-mapping reads in repetitive regions stop contaminating your calls.

Where people lose the point

  • Aligning RNA-seq with a genomic aligner and then wondering why junction-spanning reads are lost or clipped.
  • Naming a favourite tool with no reference to read length, spliced structure or downstream use.
  • Ignoring multi-mapping reads and mapping quality, then reporting confident results in repetitive or paralogous regions.
Link to this question

4.You have a raw VCF with far more variants than you expected. How do you decide which calls to believe?

Core

What a strong answer covers

  • Separate the two questions: is the site real, and is the genotype right. Site-level evidence comes from depth, mapping quality, base quality, strand balance and position within the read; genotype-level evidence comes from allele balance and genotype quality.
  • Look at the shape of the excess. A pile-up of calls in repeats, homopolymers, segmental duplications or a known problem region is an artefact signature, not biology, and region blacklists exist for exactly this.
  • Expected allele fraction depends on the biology: a germline heterozygote should sit near half, so calls far from that in a germline sample need explanation, while a somatic variant in a heterogeneous tumour legitimately can be low fraction and needs a caller built for that.
  • Global sanity checks catch problems that per-variant filters miss: transition to transversion ratio, total variant count against what that assay usually yields, and concordance with an orthogonal genotyping source if one exists.
  • Filtering approaches range from hard thresholds to model-based recalibration that learns the boundary from known sites. Model-based approaches need enough variants and a matched truth set to behave, so small panels are usually filtered with explicit thresholds.
  • A strong answer ends with visual confirmation: pulling the top candidates up in a genome browser and looking at the reads is still how experienced analysts catch alignment artefacts.

Where people lose the point

  • Applying a filter recipe copied from a whole-genome germline pipeline to a small targeted panel or a somatic study without adjusting it.
  • Filtering only on depth, which keeps deeply covered artefacts and discards real low-coverage calls.
  • Reporting a candidate variant that has never been looked at in a browser, when a five-second look would have shown it sits at the edge of every read.
Link to this question

5.Walk me through an RNA-seq differential expression analysis. What are the decisions that actually change the answer?

Core

What a strong answer covers

  • Start from raw counts per gene. Established count-based methods model counts directly with a negative binomial, so feeding them values that are already normalized for length, such as TPM or FPKM, breaks the assumptions they rely on.
  • Normalization handles differences in library composition, not just depth. Methods such as median-of-ratios or trimmed mean of M-values assume most genes are not changing, which is an assumption worth stating out loud when a treatment shuts down transcription globally or when a spike-in normalization would be more honest.
  • Dispersion estimation is where small experiments live or die: with three replicates per group, per-gene variance is badly estimated, so these methods share information across genes to shrink dispersion toward a fitted trend.
  • Design matters more than depth. Beyond a modest number of reads per sample, an extra biological replicate buys far more power than more reads on the same libraries, and technical replicates are not biological replication.
  • Filter low-count genes before testing, correct for multiple testing with a false discovery rate procedure, and consider shrinking fold changes so that a lowly expressed gene with a wild ratio does not top the list.
  • A strong answer separates statistical significance from effect size, and interprets the gene list against the design rather than reading enrichment results as mechanism.

Where people lose the point

  • Running a differential test on TPM or FPKM values instead of raw counts.
  • Reporting raw p-values, or reporting fold changes computed from unnormalized counts.
  • Treating replicates from the same culture flask, aliquot or animal as independent biological replicates.
Link to this question

6.How do you detect and handle a batch effect, and when is a batch effect unfixable?

Hard

What a strong answer covers

  • Detect it by looking at global structure before testing anything: clustering or a principal component plot coloured by processing date, sequencing run, extraction kit lot and operator, not only by biological group.
  • If batch and condition are partly confounded but not identical, batch can be included as a covariate in the model, which is preferable to correcting the data in place and then analysing it as though the correction were free.
  • Correction methods that adjust the data itself are useful for visualization and for downstream methods that cannot take covariates, but they can inflate confidence if the corrected matrix is then fed into a test that assumes untouched data.
  • A batch effect is unfixable when it is perfectly confounded with the variable of interest: all treated samples in one run and all controls in another. No statistical method can separate two effects that never vary independently, and the honest answer is that the experiment cannot answer the question.
  • A strong answer moves to prevention: randomize or block condition across runs, plates and days, keep a record of processing metadata, and involve the analyst before the samples are prepared rather than after.

Where people lose the point

  • Claiming a correction method can rescue a design where batch and condition are fully confounded.
  • Correcting the matrix and then running the standard test on it without acknowledging the effect on the error estimate.
  • Only checking for batch structure after the results look strange, rather than as a standing part of exploratory analysis.
Link to this question

7.A collaborator asks you to rerun an analysis from eighteen months ago and get the same numbers. What has to have been true for that to work?

Hard

What a strong answer covers

  • Every step has to be scripted rather than clicked. A workflow manager such as Nextflow or Snakemake gives you a declared dependency graph, restartability and a record of what actually ran, which a folder of ad hoc shell commands does not.
  • The software environment has to be pinned, through containers or a resolved environment specification, because an aligner or a statistics package that silently moved a minor version can move your results.
  • Reference data has to be pinned with the same seriousness as software: genome build, annotation release, and any variant or pathway database version.
  • Anything stochastic needs a fixed seed, and any step that depends on thread count or ordering needs to be flagged, since some tools are not bitwise deterministic across different parallel settings.
  • Parameters and inputs should live in a versioned config, and outputs should be traceable to the exact commit and config that produced them.
  • A strong answer distinguishes bitwise reproducibility from scientific reproducibility and says which one was promised: a rerun that produces the same conclusions with slightly different numbers may be entirely acceptable, but only if that was agreed rather than discovered.

Where people lose the point

  • Relying on manual steps such as a spreadsheet edit or a browser upload that leave no record.
  • Using a floating environment specification or a latest tag, so the environment cannot be rebuilt.
  • Storing only the final result table, with no link back to the code version, parameters and references that produced it.
Link to this question

8.Your differential expression run returns half the transcriptome as significant. What do you do?

Hard

What a strong answer covers

  • Treat it as a QC finding rather than a discovery. A result that large is more often a design or processing artefact than a biological one, and the credible cases (a global transcriptional shutdown, a strong developmental contrast) should be argued from the biology, not assumed.
  • Check the model first: are the sample-to-group assignments right, is the design matrix specifying the contrast you think it is, and did the covariates go in correctly. A swapped label sheet is a common cause.
  • Then check global structure: does clustering separate samples by group, or by run and date. Strong batch structure aligned with condition produces exactly this pattern.
  • Check normalization assumptions. If most genes really are changing, methods that assume most genes are stable will shift the whole distribution and manufacture significance in the opposite direction.
  • Look at the effect sizes, not just the adjusted p-values. Thousands of hits with tiny fold changes in a well-powered experiment mean something different from thousands of hits with large ones.
  • A strong answer says what it would take to believe the result: an orthogonal measurement on a handful of genes, or a replication in an independent sample set, before anything goes into a paper.

Where people lose the point

  • Raising the fold-change threshold until the list looks a reasonable size and presenting that as the result.
  • Running pathway enrichment on the huge list and reporting whatever comes back, when an enrichment on a broken input is just as broken.
  • Telling the collaborator the finding is real before checking the sample sheet.
Link to this question

9.How do you present an ambiguous analysis result to a wet-lab collaborator who wants a yes or no?

Core

What a strong answer covers

  • Lead with the answer to the question they asked, then state the confidence and what limits it, rather than walking through the pipeline before saying anything they can use.
  • Give the uncertainty in terms of decisions: what the result supports doing next, what it does not yet justify, and which experiment would resolve it fastest.
  • Show the data in a form they can inspect: a browser screenshot of the reads, a counts plot for the gene they care about, a clustering plot showing where the samples sit. Biologists can often spot a problem in their own material faster than the analyst can.
  • Be explicit about the difference between not significant and not there, especially in an underpowered design, since those get conflated in the conversation that follows.
  • A strong answer mentions writing down the caveat in the shared record, because the caveat is what disappears first when a result is passed along.

Where people lose the point

  • Delivering a p-value with no effect size and no interpretation.
  • Overstating a borderline result because the collaborator clearly wants it to be true.
  • Hiding behind method detail instead of answering the question, which reads as evasion rather than rigour.
Link to this question
No account needed

Answer one real Bioinformatics question now

A question a Bioinformatics panel actually asks, answered out loud, scored on what you said and how you said it. Under two minutes, and nothing to sign up for.

What is the difference between local and global sequence alignment, and when would you use each?

We never store the audio. Your answer is deleted within 24 hours unless you save the result.

How Bioinformatics answers get judged

The weights a Bioinformatics interviewer is holding, whether or not they say so out loud. Round Zero scores your practice answers against exactly these, and quotes your own words back as the evidence for each.

Method and tool selection

30%

Chooses aligners, callers and analysis methods from the properties of the data (read length, spliced or unspliced, ploidy, expected variant type) rather than from habit, and can say what the alternative would have cost.

Statistical judgment

30%

Handles multiple testing, replication, normalization, dispersion and confounding correctly, and distinguishes a result that is statistically significant from one that is biologically meaningful.

QC and skepticism

25%

Reads QC output critically, notices when a result is too good or too large to be real, and traces a suspicious answer back to sample handling, reference mismatch or design before believing it.

Reproducibility and communication

15%

Pins versions and references, keeps analyses scripted and rerunnable, and explains findings to collaborators who are biologists rather than programmers.

Related Science & Research skills

All skills →

Now say them out loud

You have read what strong Bioinformatics answers contain. The next thing that moves the needle is producing one under time, out loud, and finding out where it falls apart.

  • These questions asked back, with follow-ups
  • Flashcards for the ones you keep missing
  • A scored mock that quotes your own answers

Browse every skill

Practising Bioinformatics: common questions

What Bioinformatics interview questions should I practice?
Start with the core areas Bioinformatics interviewers probe: What is the difference between local and global sequence alignment, and when would you use each; You open the QC report for a fresh sequencing run. What do you look at, and what would make you stop before aligning; How do you choose an aligner, and why can't you use the same one for DNA and RNA reads. This page outlines strong answers and common mistakes, and the scored path drills each one with follow-ups.
Is the Bioinformatics practice free?
Yes. The Bioinformatics path runs free inside Round Zero: lessons, practice questions and flashcards. Drills are unlimited on every plan, free included. So is the full scorecard. Free also covers 3 complete scored interviews, no card.
How is this different from a Bioinformatics question list?
A static list gives you questions with no feedback. Round Zero runs a live scored practice that probes your actual answers, rotates difficulty, and tells you exactly what to fix, grounded in a Bioinformatics rubric.
How should I prepare for a Bioinformatics interview?
Learn the concepts, drill the questions until answers come fast, then prove it in a scored mock. Round Zero sequences all three so you know you are ready, not just that you read about Bioinformatics.
How is a Bioinformatics answer scored?
Bioinformatics answers are scored on method and tool selection, statistical judgment, qc and skepticism, reproducibility and communication, with evidence quoted from what you actually said, so feedback is specific instead of generic praise.