How Science Works | Bioinformatics — Sequences, Genomes, Omics, Databases and Turning Biological Data into Knowledge

Safety and scope: worked examples use harmless invented sequences and non-operational data. This article does not provide pathogen engineering, harmful sequence design, wet-lab protocols or optimisation of biological agents.

Reading route: represent biological data → compare sequences → build genome evidence → analyse omics → connect sequence to structure → make the analysis reproducible.

1. The scientific job is to preserve biological meaning while changing representation

NCBI describes bioinformatics as an interdisciplinary field applying computational approaches to collect, store, manipulate and analyse biological data.

The key phrase is not merely “use computers.” The computational representation must preserve the distinctions that matter biologically.

2. Biological data are heterogeneous

Sequences are strings. Structures are three-dimensional coordinates. Gene expression is a matrix. Taxonomies are trees. Pathways are graphs. Images are arrays. Scientific papers are text.

Bioinformatics succeeds by choosing data structures that fit the biological object rather than forcing everything into one table.

3. Sequence alphabets are simple; sequence meaning is not

DNA can be represented using A, C, G and T plus ambiguity symbols; proteins use amino-acid alphabets.

The string is compact, but biological meaning depends on orientation, genomic location, reading frame, species, sample and experimental provenance.

4. Metadata can be as important as the sequence

A sequence without sample identity, organism, method, date and coordinate system can be scientifically ambiguous.

Good biological databases attach provenance because the same nucleotide string can appear in different biological contexts.

5. Identifiers make records addressable

Accession numbers and stable database identifiers allow researchers to cite a record precisely.

Human-readable names can change or collide; persistent identifiers reduce ambiguity across versions and publications.

6. Versioning matters because biological databases change

Reference assemblies are corrected, gene models are revised and annotation databases grow.

An analysis that records “used database X” without a version can become impossible to reproduce later.

7. Controlled vocabularies reduce semantic drift

Ontologies define shared terms and relationships for genes, phenotypes, anatomy, diseases and experimental processes.

This makes “same concept, different wording” computable while preserving hierarchy and context.

8. Worked example: the same string can mean different things with different metadata

Harmless toy example. The sequence ACGTTGCA appears in two files.

One record is genomic DNA; the other is an adapter control used for software testing. Without metadata, identical characters invite a false biological conclusion.

9. Sequence alignment asks which positions correspond

Alignment inserts gaps and matches characters to represent a hypothesis of positional correspondence.

For related biological sequences, this correspondence can reflect shared ancestry, structural constraint or repeated motifs.

10. Global and local alignment answer different questions

Global alignment compares sequences across their full lengths; local alignment searches for strongly matching subsections.

A conserved domain inside otherwise different proteins is therefore a local-alignment problem, not necessarily a whole-sequence problem.

11. Alignment requires a scoring model

Matches, mismatches and gaps receive scores reflecting a model of plausible change.

Protein substitution matrices encode the fact that some amino-acid replacements are more evolutionarily plausible than others.

12. Gap penalties model insertion and deletion events

Opening one long gap can be biologically more plausible than creating many independent one-position gaps.

Affine gap penalties therefore distinguish the cost of opening from extending a gap.

13. Worked example: alignment is an optimisation under assumptions

Harmless toy sequences. Compare ACGTAC with ACGAC.

One plausible alignment inserts a gap after the third character. A different scoring system could prefer another placement if repeated characters make the correspondence ambiguous.

14. Similarity and homology are not synonyms

Similarity is a measurable degree of resemblance. Homology is a historical claim of common ancestry.

Sequences can be highly similar because they are homologous, but “70% homologous” is conceptually wrong: homology is not a percentage.

15. Orthologues and paralogues have different evolutionary relationships

Orthologues diverge after speciation; paralogues diverge after gene duplication.

Function can be conserved or diverge in either case, so relationship type informs but does not determine functional annotation.

16. Database search balances sensitivity and speed

Exact dynamic-programming alignment against every database sequence can be expensive.

Heuristic tools such as BLAST use fast candidate-generation strategies before detailed scoring, trading some exhaustive sensitivity for practical scale.

17. E-values describe expected chance matches under a model

In sequence search, an E-value estimates how many matches with similar score might occur by chance in a database of that size.

Database size matters: the same raw alignment score can have a different significance when the search space changes.

18. Multiple sequence alignment reveals conserved columns

Aligning many related sequences can reveal positions that evolve slowly and insertions that occur in subsets.

Conservation suggests functional or structural constraint but still requires experimental interpretation.

19. A reference genome is a coordinate system, not a perfect individual

A reference assembly provides ordered sequence coordinates used to align reads and describe variants.

It cannot contain every population allele, structural variant or haplotype found across a species.

20. Genome assembly reconstructs long sequence from shorter observations

Sequencing instruments observe fragments rather than reading an entire chromosome from end to end in one perfect pass.

Assembly algorithms exploit overlaps or graph relationships to infer a longer sequence, with repeats creating ambiguity.

21. De Bruijn graphs compress overlap structure

Short reads can be decomposed into k-mers whose overlaps form a graph.

Branches arise from repeats, variants and errors, so graph cleaning is an inference problem rather than simple concatenation.

22. Long reads change the ambiguity structure

Longer sequencing reads can span repeats and structural variants that short reads cannot resolve.

They may carry different error profiles, so assembly quality depends on technology and correction strategy.

23. Variant calling compares observed sequence with a reference model

Algorithms infer whether mismatches and gaps represent true biological variants or sequencing and alignment error.

Read depth, base quality, mapping quality and local sequence complexity all contribute to confidence.

24. Quality scores are probabilities in disguise

Phred-style quality scores transform estimated error probability onto a logarithmic scale.

A ten-point score increase corresponds to roughly tenfold lower estimated error probability under the convention.

25. Worked example: Q30 is not “30% accurate”

Original calculation. Q30 corresponds to an estimated error probability of 10⁻³, or about one error per thousand calls under the model.

The transformation prevents a common interpretation error: quality score is logarithmic, not a percent accuracy label.

26. Gene annotation is a layered claim

Structural annotation identifies genes, exons and regulatory regions; functional annotation proposes what those elements do.

Computational predictions can use sequence similarity, domains, expression and comparative genomics, but experimental evidence remains stronger for specific functions.

27. Variant annotation adds biological context

A genomic variant can be described by location, predicted molecular consequence, population frequency and evidence from functional or clinical databases.

Prediction is not diagnosis; clinical interpretation belongs to qualified medical genetics and healthcare owners.

28. Pangenomes reduce single-reference bias

A graph or collection of multiple representative genomes can preserve structural alternatives absent from one linear reference.

This improves representation of diverse populations and changes the computational problem from one path to many possible paths.

29. Phylogenetics infers historical relationships

Aligned characters are used to estimate trees or networks that explain observed sequence differences under evolutionary models.

Branch support measures uncertainty in inferred splits; one tree diagram should not hide competing plausible histories.

30. Gene trees and species trees can disagree

Incomplete lineage sorting, duplication, loss and horizontal transfer can make one gene’s history differ from the species history.

Bioinformatics must therefore avoid treating one locus as an unquestioned species tree.

31. Transcriptomics measures RNA abundance across many genes

RNA sequencing converts sampled transcripts into count-like measurements that reflect both biology and technical processes.

Library size, transcript length, composition and batch must be handled before expression differences are interpreted.

32. Normalisation makes samples comparable under assumptions

Raw counts cannot be compared directly when sequencing depth differs.

Normalisation methods estimate scaling factors, but strong global biological shifts can violate simple assumptions.

33. Differential expression is a statistical model

Algorithms compare groups while modelling count variability and experimental design.

A small p-value indicates incompatibility with a null model, not automatically a large or biologically important effect.

34. Multiple testing changes the false-positive problem

Testing thousands of genes creates many opportunities for chance significance.

False-discovery-rate procedures control the expected proportion of false discoveries among selected results under stated assumptions.

35. Worked example: 20,000 tests make uncorrected p-values misleading

Original conceptual example. If 20,000 truly null genes were each tested at p < 0.05 independently, roughly 1,000 false positives would be expected on average.

The calculation is simplified, but it shows why genome-scale analysis needs multiplicity control.

36. Batch effects can dominate real biology

Samples processed on different days, instruments or reagent lots can acquire systematic differences.

Randomisation and balanced experimental design prevent batch from becoming perfectly confounded with the biological condition.

37. Principal-component plots are diagnostic summaries, not proofs

Dimension-reduction can reveal dominant patterns, outliers and batch structure.

Axes are mathematical directions of variance; they do not automatically correspond to one biological mechanism.

38. Proteomics adds another measurement layer

Mass spectrometry can identify and quantify peptides that infer protein abundance and modification state.

Protein inference is complicated when peptides are shared among multiple proteins.

39. Metabolomics observes chemical state downstream of genes

Metabolite abundances reflect enzymes, nutrients, environment and time.

Annotation can be difficult because many molecules have similar masses or spectra, requiring standards and orthogonal evidence.

40. Single-cell data separate mixtures of cell states

Bulk samples average many cells together.

Single-cell sequencing reveals heterogeneity but adds sparse counts, dropout, cell-state continua and difficult annotation.

41. Metagenomics studies mixed communities without isolating every organism

Environmental DNA can be sequenced from communities containing many organisms.

Computational classification, assembly and abundance estimation infer community structure; this article stays at the analytical level and does not provide pathogen cultivation or engineering procedures.

42. Multi-omics integration connects layers but can amplify confounding

Genomes, transcripts, proteins and metabolites capture different parts of biological state.

Integration is valuable only when sample matching, timing and batch structure are controlled across all layers.

43. Structural bioinformatics asks how sequence maps into three-dimensional form

Protein sequences fold into structures whose geometry constrains binding, catalysis and mechanical behaviour.

Experimental structures and computational predictions complement one another but have different uncertainty profiles.

44. Structure prediction is a model, not direct observation

Modern machine-learning systems can predict many protein structures accurately from sequence and evolutionary information.

Confidence scores, flexible regions, ligand state and multimeric context still determine what conclusions are justified.

45. Domain annotation decomposes proteins into reusable units

Proteins often contain domains with conserved sequence and structural patterns.

Finding a domain can support a broad functional hypothesis while leaving substrate specificity or regulation unresolved.

46. Interaction networks are graphs with measurement bias

Protein interactions, regulatory relationships and metabolic reactions can be represented as nodes and edges.

Highly studied genes often appear more connected partly because they have been tested more frequently.

47. Pathway enrichment depends on the background universe

An enriched pathway is one containing more selected genes than expected relative to a defined comparison set.

Changing the background from “all genes” to “genes measurable in this experiment” can change the result substantially.

48. Pipelines turn analysis into an explicit sequence of transformations

Quality control, trimming, alignment, quantification, statistics and reporting are encoded as steps with inputs and outputs.

A pipeline is scientifically useful when each transformation is versioned and inspectable rather than hidden behind one final figure.

49. Containers preserve software environments

Analysis can change when library versions or operating-system dependencies change.

Containerised environments help preserve executable software state, though data and database versions must still be recorded separately.

50. Workflow languages expose dependencies

Workflow engines track which task consumes which file and can rerun only affected steps after a change.

This makes large analyses easier to audit and scale across computing systems.

51. Data provenance is part of the result

A reproducible figure should trace backward to source data, preprocessing, code, parameters and database versions.

Without that chain, later researchers may reproduce the picture but not the computation that produced it.

52. Privacy constrains biological data sharing

Human genomic and health data can be identifying and sensitive.

Access controls, consent, de-identification limits and governance therefore belong inside bioinformatics architecture, not as an afterthought.

53. Benchmark leakage can make algorithms look stronger than they are

Closely related sequences or duplicated samples can appear in both training and test sets.

Performance then reflects memorised similarity rather than generalisation to genuinely new biological cases.

54. External validation tests whether a computational result travels

A classifier or biomarker found in one dataset should be evaluated on independent samples collected under different conditions.

Failure outside the discovery dataset can reveal overfitting, population shift or technical confounding.

55. Common bioinformatics failure modes

• Sequence similarity equals same function: ignoring divergence and context.
• Reference genome equals species genome: ignoring population variation.
• Raw counts are comparable: ignoring depth and composition.
• Small p-value equals biological importance: ignoring effect size and multiplicity.
• Pipeline completed equals result valid: ignoring biological assumptions.
• Reproducible code equals reproducible science: ignoring changing databases and data provenance.

56. How to think like a bioinformatician

Start from the biological question. Define the data object and metadata. Choose algorithms whose assumptions fit the experiment. Preserve versions. Inspect quality before modelling. Separate exploratory from confirmatory analysis. Validate computational claims with independent data and, where appropriate, laboratory evidence owned by the relevant biological discipline.

57. A staged learning route

First encounter: DNA strings, genes, databases, sequence matching and genomes.

Secondary-to-JC bridge: alignments, quality scores, variants, reference genomes, expression matrices and phylogenetic trees.

Higher resolution: assembly graphs, probabilistic variant calling, single-cell models, pangenomes, structural prediction, workflow provenance and benchmark design.

58. Checkpoints with answers

Is a sequence match proof of identical function? No. Similarity supports a hypothesis that requires evolutionary and experimental context.

Why does a reference genome need a version? Assemblies and annotations change, altering coordinates and downstream interpretation.

Why correct for multiple testing? Genome-scale analyses perform thousands of tests, making chance discoveries common without adjustment.

Why can a reproducible pipeline still be scientifically wrong? Software can execute perfectly while the biological model, sample design or interpretation is invalid.

59. The final skill is keeping computation attached to biology

A complete bioinformatics explanation follows biological material into digital representation, through algorithms and statistical inference, and back into a biological claim whose provenance, uncertainty and validation remain visible.

Sources and connected subjects

Useful foundations include the National Center for Biotechnology Information, the NCBI Handbook and public biological sequence and annotation resources. Worked examples above use harmless invented data.

Continue to Genetics, Molecular Biology, Data Science and Astrobiology.

Return to How Science Works or the How X Works Hub.