Bioinformatics basics

In short

Bioinformatics is the use of computers, databases and algorithms to store, retrieve, compare and interpret biological sequence and structure data — DNA, RNA, proteins and whole genomes. It matters because sequencing now generates data far faster than any wet lab can analyse it: a single human genome run yields roughly 100 GB of raw reads, useless without computation. The topic is examined in GATE Biotechnology (BT) and Life Sciences (XL) under "bioinformatics: major databases, sequence and structure alignment, phylogeny, data mining", and in university Applied Biotechnology papers as database types, alignment algorithms and tools. Indian employers who test it include Strand Life Sciences, MedGenome, Eurofins, Syngene (Biocon), Novartis/AstraZeneca GCCs, TCS Life Sciences, Persistent, Bugworks and CSIR labs (IGIB, CCMB, NII), plus bioinformatics roles at NIBMG, C-DAC and the Bioinformatics Resources and Applications Facility.

Explained

What bioinformatics is, and its scope

• Bioinformatics: design and application of computational methods to manage and analyse biological data (sequences, structures, expression, pathways). Computational biology overlaps but leans toward modelling and simulation of biological systems.
• Three core activities: (1) building and curating databases; (2) developing algorithms and tools; (3) using both to answer a biological question.
• Levels of data: primary sequence (1-D), structure (3-D), function/annotation, and system-level (networks, pathways).
• In silico = done by computer, as against *in vitro* (test tube) and *in vivo* (living organism). Bioinformatics narrows the candidate list before expensive wet-lab work.
• Central dogma as a data pipeline: DNA → RNA → protein maps to genome databases → transcript/EST databases → protein sequence and structure databases.

Biological databases and file formats

• Classification by content:
  • *Primary (archival):* raw experimental submissions, author-owned. GenBank (NCBI, USA), ENA (EMBL-EBI, Europe), DDBJ (Japan) — these three exchange data daily as the INSDC, so an accession in one appears in all.
  • *Secondary (derived/curated):* built by analysing primary data — UniProtKB/Swiss-Prot (manually reviewed), PROSITE, Pfam (protein families as HMM profiles), InterPro, SCOP/CATH (structure classification).
  • *Composite/specialised:* NCBI RefSeq (non-redundant curated reference sequences), OMIM (human genetic disorders), KEGG (pathways), PubChem/ChEMBL (small molecules), TAIR (*Arabidopsis*), MIPS.
• Structure databases: PDB (experimental 3-D structures from X-ray, NMR, cryo-EM; Indian mirror/node historically at BIC, now accessed via RCSB/PDBe/PDBj), CSD for small molecules, AlphaFold DB for predicted structures.
• Indian databases: IBDC (Indian Biological Data Centre at RCB, Faridabad — the national repository, established under DBT; submission of Indian sequence data to it is now expected), IndiGen (genomes of Indians, CSIR), GenomeIndia (DBT), and the Biotechnology Information System Network (BTISnet) centres.
• Sequence retrieval: Entrez (NCBI's integrated query system), SRS, BLAST links, and programmatic access via E-utilities APIs.
• File formats to recognise:
  • FASTA: a > header line then sequence lines — used by nearly every tool.
  • GenBank/EMBL flat file: LOCUS/ID, DEFINITION, ACCESSION, VERSION, SOURCE, FEATURES (CDS, gene, exon with coordinates), ORIGIN + sequence.
  • FASTQ: read sequence plus per-base quality in ASCII (Phred+33), the standard NGS format.
  • SAM/BAM (alignments to a reference), VCF (variants), GFF/GTF (feature coordinates), PDB/mmCIF (atomic coordinates).
• Accession vs version vs GI: accession is permanent (e.g. NM_000518); the suffix after the dot is the version, which changes if the sequence is corrected.

Sequence alignment — the central algorithmic idea

• Why align: similarity in sequence suggests common ancestry (homology) and often common function or structure. Similarity is measured; homology is inferred — they are not synonyms.
• Homologue types: orthologues (same gene in different species, separated by speciation), paralogues (duplicated genes within a genome), xenologues (acquired by horizontal transfer).
• Pairwise alignment types:
  • *Global* — aligns sequences end-to-end; Needleman–Wunsch (1970), dynamic programming. Use for sequences of similar length and expected full-length relatedness.
  • *Local* — finds the best-matching segments; Smith–Waterman (1981), dynamic programming with a zero floor. Use for domain-level similarity or very different lengths.
• Dynamic programming steps: (1) initialise a matrix; (2) fill each cell with max of (diagonal + match/mismatch score, up − gap, left − gap) [plus 0 for local]; (3) traceback from the highest/corner cell to recover the alignment. Time and space complexity ≈ O(mn), which is why heuristics exist.
• Gap penalties: *linear* (w = g·k) vs affine (w = open + extend·(k−1)) — affine is biologically realistic because one insertion event of 5 bases is likelier than 5 separate events.
• Scoring matrices (proteins):
  • PAM (Dayhoff): built from closely related, globally aligned sequences; PAM1 = 1 accepted point mutation per 100 residues; higher PAM number = more divergent sequences. PAM250 for distant relatives.
  • BLOSUM (Henikoff): built from local blocks of conserved regions; BLOSUM62 (default in BLAST) is derived from blocks clustered at ≥62% identity; lower BLOSUM number suits more divergent sequences. So PAM and BLOSUM run in opposite directions.
  • DNA scoring is usually simple: match +1 to +5, mismatch −1 to −4 (BLASTN default +2/−3).
• Heuristic database search:
  • BLAST (Basic Local Alignment Search Tool): breaks the query into words (k-mers; 3 for protein, 11 or 28 for nucleotide), finds seed hits in the database index, extends them into HSPs (high-scoring segment pairs), reports those above threshold. Far faster than Smith–Waterman but not guaranteed optimal.
  • Flavours: blastn (nt vs nt), blastp (prot vs prot), blastx (translated query vs protein DB), tblastn (protein query vs translated DB), tblastx, PSI-BLAST (iterative, position-specific scoring matrix — finds remote homologues), DELTA-BLAST, megablast (highly similar), BLAT and FASTA (Pearson–Lipman) as alternatives.
  • Output statistics: raw score → bit score (normalised, database-independent), E-value (expected number of chance hits of that score in a database of that size — smaller is better), percent identity, query coverage.
• Multiple sequence alignment (MSA): needed for motifs, profiles and phylogeny. Exact DP is O(Lⁿ) — infeasible — so progressive methods are used: Clustal Omega / ClustalW (pairwise distances → guide tree → align in order), MUSCLE, T-Coffee, MAFFT; HMMER builds profile HMMs. Output formats: CLUSTAL, Stockholm, PHYLIP, NEXUS. Consensus, conserved blocks and Jalview-style colouring read the result.

Phylogenetics and molecular evolution

• Tree vocabulary: nodes (taxa), internal nodes (inferred ancestors), branches (with lengths = evolutionary distance), root (common ancestor), topology, clade/monophyletic group, outgroup (used to root the tree).
• Distance-based methods: UPGMA (assumes a constant molecular clock, ultrametric) and Neighbour-Joining (NJ — does not assume equal rates; fast, most used for quick trees).
• Character-based methods: Maximum Parsimony (fewest mutations), Maximum Likelihood (ML — best tree under an explicit substitution model: Jukes–Cantor, Kimura 2-parameter, HKY, GTR), Bayesian inference (MrBayes, posterior probabilities).
• Distance correction: observed differences underestimate true substitutions because of multiple hits at the same site; Jukes–Cantor and Kimura corrections fix this.
• Bootstrap: resample alignment columns with replacement, rebuild the tree 100–1,000 times; the bootstrap value on a branch is the percentage of replicates showing it. ≥70% is usually treated as reasonable support.
• Software: MEGA (widely taught in Indian labs), PHYLIP, RAxML, IQ-TREE, FigTree.

Sequence analysis, gene finding and annotation

• Basic sequence statistics: length, GC content, melting temperature, codon usage bias, reverse complement, translation in six frames.
• ORF finding: search all six reading frames for ATG … stop (TAA/TAG/TGA); tools — NCBI ORFfinder, EMBOSS getorf. Prokaryotic gene prediction (GLIMMER, Prodigal) is easier — genes are continuous, ORFs are long. Eukaryotic prediction (GENSCAN, AUGUSTUS) must handle introns, so it uses splice-site signals, promoter/TATA elements, poly-A sites and evidence from ESTs/RNA-seq.
• Annotation levels: *structural* (where the genes and features are) and *functional* (what they do — GO terms, EC numbers, pathway mapping).
• Protein-level prediction: molecular weight and theoretical pI (ExPASy ProtParam), hydropathy (Kyte–Doolittle), transmembrane helices (TMHMM), signal peptides (SignalP), post-translational modification sites, domains (Pfam/InterProScan), secondary structure (PSIPRED, Chou–Fasman and GOR as classical methods).
• Structure prediction: homology (comparative) modelling — needs a template with usually >30% identity (MODELLER, SWISS-MODEL); threading/fold recognition for the twilight zone; ab initio for no-template cases; AlphaFold2/3 deep-learning prediction with per-residue pLDDT confidence. Validation: Ramachandran plot, RMSD against template, PROCHECK/MolProbity.
• Structure-based drug design: molecular docking (AutoDock Vina, GLIDE) scores a ligand pose in a binding site; virtual screening of libraries; pharmacophore mapping; QSAR; molecular dynamics (GROMACS, NAMD) for stability; ADMET filters such as Lipinski's Rule of Five.
• Primer design rules of thumb (Primer3/Primer-BLAST): 18–25 nt, GC 40–60%, Tm 55–65 °C with the pair within ~5 °C, avoid 3′ self-complementarity and runs of 4+ identical bases, check specificity by BLAST.

NGS data, tools and programming

• Typical variant-calling pipeline: FASTQ → quality check (FastQC, MultiQC) → trimming (Trimmomatic, fastp) → alignment to reference (BWA-MEM, Bowtie2 for DNA; STAR/HISAT2 for RNA) → SAM→BAM sort/index (SAMtools) → duplicate marking (Picard) → variant calling (GATK HaplotypeCaller, bcftools) → VCF → annotation (ANNOVAR, VEP, SnpEff) → interpretation (ClinVar, ACMG classes).
• RNA-seq pipeline: alignment or pseudo-alignment (Salmon/kallisto) → counts (featureCounts/HTSeq) → normalisation and differential expression (DESeq2, edgeR in R/Bioconductor) → enrichment (GO, KEGG).
• Assembly: *de novo* assemblers use overlap-layout-consensus (long reads) or de Bruijn graphs (short reads — SPAdes, Velvet); quality judged by N50, total length and completeness (BUSCO). Reference-guided assembly is easier where a genome exists.
• Phred quality: Q = −10 log₁₀(P_error). Q20 = 1 error in 100 bases (99% accuracy), Q30 = 1 in 1,000 (99.9%) — the usual acceptance threshold.
• Programming and environment: Python with Biopython (SeqIO, Seq, Entrez, AlignIO), R with Bioconductor, Perl/BioPerl (legacy), shell scripting and awk for text munging, SQL for database work, Linux/HPC job schedulers, Nextflow/Snakemake for reproducible workflows, Git for version control, Docker/Conda for environments.
• Advantages of bioinformatics: fast, cheap relative to wet lab, scalable, reuses public data, enables genome-scale questions impossible by experiment.
• Limitations: predictions are hypotheses needing experimental validation; annotation errors propagate between databases; databases are biased toward model organisms and European-ancestry human data; garbage-in-garbage-out on poor-quality reads; heavy compute needs; "homology by transitivity" errors.
• Indian policy and legal frame: the Digital Personal Data Protection Act, 2023 governs personal (including genomic-linked) data — note that its rules were notified in 2025 and are being phased in, so compliance detail is still settling; the Biological Diversity Act, 2002 (amended 2023) with NBA approval for access to Indian biological resources and benefit sharing; ICMR National Ethical Guidelines (2017) and the DBT/ICMR Biomedical Research and Health Data sharing policies; the Indian Biological Data Centre is the designated national store for sequence data generated in India. The DNA Technology (Use and Application) Regulation Bill has been withdrawn, so no such Act is in force.

Key points

• Bioinformatics – computational storage, retrieval and analysis of biological data; *in silico* complement to lab work.
• Primary vs secondary database – archival author-submitted records (GenBank/ENA/DDBJ) vs curated derived data (Swiss-Prot, Pfam, PROSITE).
• INSDC – GenBank, ENA and DDBJ synchronise daily, so one accession resolves in all three.
• RefSeq – curated, non-redundant reference sequences; use for a clean reference, GenBank for the raw submission.
• FASTA – >header + sequence; FASTQ adds Phred quality; VCF stores variants; BAM stores alignments.
• Similarity vs homology – similarity is a measured percentage; homology is an inferred shared ancestry, and is not expressed as a percentage.
• Needleman–Wunsch – global alignment by dynamic programming; Smith–Waterman – local alignment, negative scores clamped to zero.
• BLAST – heuristic word-seed-and-extend local search; fast but not guaranteed optimal.
• E-value – expected number of chance hits of equal score; depends on database size, so it changes as databases grow. Bit score does not.
• PAM vs BLOSUM – higher PAM = more divergent; lower BLOSUM = more divergent (BLOSUM62 is the BLAST default).
• Affine gap penalty – gap opening costs more than gap extension, matching real indel biology.
• MSA – progressive alignment (Clustal Omega, MUSCLE, MAFFT) because exact multi-dimensional DP is computationally impossible.
• Phylogenetic methods – distance (UPGMA, NJ) vs character-based (parsimony, maximum likelihood, Bayesian); bootstrap gives branch confidence.
• Gene prediction – straightforward in prokaryotes (long continuous ORFs), hard in eukaryotes (introns, splice sites, alternative splicing).
• Homology modelling – usable above roughly 30% sequence identity with a template; AlphaFold reports pLDDT confidence per residue.
• Q30 – 99.9% base-call accuracy, the standard NGS quality cut-off.
• N50 – the contig length at which half the assembly is in contigs of that size or longer; a size statistic, not a correctness statistic.
• Indian frame – IBDC (Faridabad) as national data repository; NBA approval under the Biological Diversity Act for Indian biological resources; DPDP Act, 2023 for personal data.

Formulas and examples

Example 1 — Percent identity and the denominator trap. Alignment of two proteins: aligned region 180 positions, of which 72 are identical; query length 200, subject length 240.

• Identity over the alignment = 72/180 = 40.0%
• Identity over the query = 72/200 = 36.0%
• Identity over the longer sequence = 72/240 = 30.0%

Same alignment, three published numbers — always state the basis. At 40% identity over 180 residues the pair is comfortably homologous; at 25% over 60 residues it would be in the "twilight zone".

Example 2 — E-value and the effect of database size. A blastp hit has bit score S′ = 40. Query m = 300 residues, database n = 2 × 10⁸ residues.

• m·n = 300 × 2 × 10⁸ = 6 × 10¹⁰
• 2⁻⁴⁰ = 1/1.0995 × 10¹² = 9.095 × 10⁻¹³
• E ≈ 6 × 10¹⁰ × 9.095 × 10⁻¹³ = 5.46 × 10⁻² ≈ 0.055 → borderline, worth checking.

Now the database grows tenfold to 2 × 10⁹ residues with the same hit: E ≈ 0.546 — ten times worse, purely from database size. The alignment did not change; the significance did. To regain E ≈ 0.055 the bit score must rise by log₂10 ≈ 3.3 bits.

Example 3 — Phred scores and expected errors. A 150 bp read with every base at Q30:

• P per base = 10^−30/10 = 10⁻³
• Expected miscalled bases per read = 150 × 10⁻³ = 0.15, i.e. about 1 read in 7 carries an error.
• Probability the whole read is perfect = (1 − 0.001)¹⁵⁰ = 0.999¹⁵⁰ ≈ 0.861 (86.1%).

At Q20 (P = 0.01): expected errors = 150 × 0.01 = 1.5 per read, and P(perfect) = 0.99¹⁵⁰ ≈ 0.221 — which is why Q30 is the working threshold. ASCII characters: Q30 → chr(63) = ?; Q40 → chr(73) = I.

Example 4 — Coverage for a human exome/genome run. A run yields 400 million paired-end reads of 150 bp each, mapped to a 3.1 × 10⁹ bp genome.

• Total bases = 400 × 10⁶ × 150 = 6.0 × 10¹⁰ bp = 60 Gb
• Coverage C = 6.0 × 10¹⁰ / 3.1 × 10⁹ = 19.4× — adequate for germline SNP calling (30× is the usual target), not for somatic tumour variants (often 100×+).
• To reach 30×, bases needed = 30 × 3.1 × 10⁹ = 9.3 × 10¹⁰ → reads needed = 9.3 × 10¹⁰/150 = 620 million reads, a 1.55× bigger run.

Example 5 — Jukes–Cantor distance and tree branch length. Two homologous 500 bp genes differ at 60 positions.

• p = 60/500 = 0.12
• (4/3)p = 0.16; 1 − 0.16 = 0.84; ln 0.84 = −0.17435
• d = −0.75 × (−0.17435) = 0.1308 substitutions per site.

The corrected distance (0.1308) exceeds the observed 0.12 by about 9% because of unseen multiple substitutions; at p = 0.30 the correction is much larger — d = −0.75 ln(1 − 0.40) = 0.383 against 0.30 observed, a 28% increase.

Example 6 — Needleman–Wunsch by hand (small matrix). Sequences GATTA and GTTA; match +1, mismatch −1, linear gap −2.

        -    G    T    T    A
   -    0   -2   -4   -6   -8
   G   -2    1   -1   -3   -5
   A   -4   -1    0   -2   -2
   T   -6   -3    0    1   -1
   T   -8   -5   -2    1    0
   A  -10   -7   -4   -1    2

Final cell = +2; traceback gives

G A T T A
G - T T A

Score check: G/G +1, A/gap −2, T/T +1, T/T +1, A/A +1 → 1 − 2 + 1 + 1 + 1 = +2. Consistent.

Where it is used

• Clinical genomics and diagnostics: MedGenome, Strand Life Sciences, Neuberg, Metropolis and Mapmygenome run variant-calling pipelines for carrier screening, cancer panels and rare-disease exomes; roles — NGS data analyst, variant curator, clinical bioinformatician.
• Pharma and biosimilars: Biocon/Syngene, Dr Reddy's, Sun Pharma, Serum Institute and Bharat Biotech use sequence analysis for molecule design, codon optimisation of expression constructs, epitope prediction for vaccine candidates and host-cell-protein analysis; CROs and global capability centres of Novartis, AstraZeneca, Lilly and Bristol Myers Squibb in Hyderabad and Bengaluru hire heavily.
• Agriculture and plant breeding: ICAR institutes (IASRI, NBPGR, IIRR), NIPGR and seed companies use marker-assisted selection, QTL and GWAS analysis, and genome assembly of crop and pathogen strains.
• Public health surveillance: INSACOG's genome-sequencing network for SARS-CoV-2 lineage assignment (Pangolin, Nextclade) and tuberculosis drug-resistance prediction at NIRT/ICMR labs.
• IT services and product firms: TCS Life Sciences, Infosys, Persistent Systems, Cognizant and Excelra build pipelines, LIMS, curated data products and cloud workflows for global biopharma clients.
• Research and academia: CSIR-IGIB, CCMB, NCBS, IISc, NII, NIBMG, DBT-BIRAC-funded startups; central facilities under BTISnet and compute on PARAM/C-DAC supercomputers.

Interview questions

• Q: What is the difference between similarity and homology? A: Similarity is a measurable quantity — percent identity or alignment score between two sequences. Homology is an inference of descent from a common ancestor; sequences are either homologous or not, so "70% homologous" is wrong usage.
• Q: When would you prefer Smith–Waterman over Needleman–Wunsch? A: When the sequences differ greatly in length or share only a domain or motif, because local alignment finds the best matching segment instead of forcing an end-to-end alignment.
• Q: Why is BLAST faster than Smith–Waterman, and what do you lose? A: BLAST only extends alignments around short exact word matches instead of filling the whole matrix, giving near-linear speed; it can miss optimal alignments that contain no exact seed word, so it is a heuristic.
• Q: Two hits have the same percent identity but very different E-values. Why? A: E-value depends on alignment length and score as well as identity, and on database size — a short high-identity match scores few bits and is easily matched by chance, so its E-value is poor.
• Q: Which scoring matrix would you use for distantly related proteins? A: A high PAM number (PAM250) or a low BLOSUM number (BLOSUM45); BLOSUM62 is the default for average cases and BLOSUM80 for close relatives.
• Q: Why is an affine gap penalty better than a linear one? A: Insertions and deletions usually occur as single multi-residue events, so charging a large opening cost and a small extension cost reflects biology and avoids many scattered one-base gaps.
• Q: Why is eukaryotic gene prediction harder than prokaryotic? A: Eukaryotic genes are split by introns with variable splice sites, have alternative splicing and long intergenic regions, so ORFs are not continuous and signal-based or evidence-based models are needed.
• Q: What does a bootstrap value of 95 on a branch mean? A: In 95% of resampled pseudo-replicate alignments the same branch appeared; it measures internal consistency of the data under the method, not that the branch is biologically true.
• Q: What is N50 and what does a high N50 not tell you? A: The length such that contigs of that size or longer cover half the assembly; a high N50 shows contiguity but says nothing about base accuracy, misassemblies or completeness — use BUSCO and read-remapping for those.
• Q: Walk me through a variant-calling pipeline. A: FastQC check, trim adapters and low-quality ends, align to the reference with BWA-MEM, sort and index BAM with SAMtools, mark duplicates, recalibrate and call variants with GATK, then annotate the VCF with VEP or ANNOVAR and filter against ClinVar and population frequencies.
• Q: Your homology model has 22% identity to the template. Is it usable? A: Not reliably — below roughly 30% identity the alignment itself becomes uncertain, so use threading or a deep-learning predictor such as AlphaFold and judge the per-residue confidence before any docking.
• Q: What Indian legal requirements affect working with human genomic data? A: The DPDP Act, 2023 governs personal data processing and consent; ICMR national ethical guidelines and institutional ethics committee approval are needed for human-participant research; access to Indian biological resources requires NBA clearance under the Biological Diversity Act, 2002 as amended in 2023, and Indian sequence data is expected to be deposited in the IBDC.

Common mistakes

• Writing "X% homology" instead of "X% identity or similarity" — interviewers treat this as a conceptual error, not a slip.
• Quoting an E-value as if it were fixed; it scales with database size, so a reported E-value is meaningless without the database and date. Bit score is the stable number.
• Getting the PAM and BLOSUM direction backwards (high PAM and low BLOSUM both mean *distant*).
• Choosing global alignment for a short motif against a whole chromosome, or blastn for a protein-coding comparison where blastx/tblastn would find the relationship through the translated frame.
• Calculating percent identity without stating the denominator, or comparing identities computed on different bases.
• Trusting a database annotation chain — "similar to a hypothetical protein" propagated between records — as experimental evidence of function.
• Treating a prediction (gene model, docking score, structure, pathogenicity call) as a result instead of a hypothesis needing wet-lab or clinical confirmation.
• In NGS work, skipping quality control and duplicate marking, then reporting PCR duplicates or adapter sequence as real variants.

Related topics

Genomics and proteomics · Molecular biology and recombinant DNA technology · Biostatistics and biological data analysis · Computational drug design and molecular modelling · Protein structure and function · Biosafety, bioethics and IPR