Bioinformatics File Formats Explained: FASTA, BAM, GFF, and VCF

Quick answer: Every bioinformatics pipeline is a chain of file formats: sequencing machines emit FASTQ reads, aligners produce SAM/BAM alignments, annotations live in GFF3, and variant callers write VCF. Each format exists because it answers one specific question at one stage of the pipeline. Learn which format answers which question, and every tool manual you read afterward starts to make sense.

A companion to the two FASTA and FASTQ format guides already on this blog. This article is the wider map: what each format in a modern project stores, when it appears, and what reads it.

One job, one format

Biology produces many kinds of claim, and no single file can hold them all. Each major format is organized around exactly one claim type:

• Here is a sequence and (possibly) its quality per base — FASTA, FASTQ
• Here is where these reads align against that reference — SAM, BAM, CRAM
• Here are the ways this sample differs from the reference — VCF
• Here is where genes, exons, and other features sit on the sequence — GFF/GTF, BED

This division is the real organizing principle of the field, and it explains awkward-sounding facts such as why a BAM file keeps the reference packed inside CRAM, or why a VCF header is enormous. When a file confuses you, ask first: which single claim is this format built to store?

FASTA and FASTQ: the sequence layer

Format	Stores	Typical size	Read with
FASTA	One or more sequences plus headers	0.1 to 4 GB	Everything; genome browsers, BLAST
FASTQ	Per-base sequences plus Phred-scaled quality scores	10 to 200+ GB	FASTQ processing tools, QC packages

These two formats are covered in depth in their own guides on this blog; what matters in the wider map is that everything upstream of the aligner is FASTQ and everything else is downstream of it. A FASTQ record is four lines: header, sequence, a separator, and a quality string of equal length. The quality line encodes base-call confidence as printable characters, so quality scores, not just read lengths, are one less thing the aligner must guess.

SAM and BAM: the alignment layer

SAM (Sequence Alignment/Map) is a huge tab-separated text file: a header section, then one line per alignment record. BAM is exactly that file, binary-compressed and, crucially, sorted and indexed for random access. Practical facts that solve most beginner problems:

• A BAM is only as good as its index. After sorting, always produce the .bai (or .csi) index; most downstream tools refuse or crawl without it.
• Alignment records are not only positions. The FLAG field alone encodes forty-plus facts about the read (paired, forward or reverse, mate mapped, secondary alignment, duplicates).
• CRAM is the space-efficient successor: it achieves its compression partly by referring to the reference sequence, which trades file size for a file that cannot be opened without the right reference handy.

The core toolkit is samtools:

samtools view -b -o aligned.bam aligned.sam     # convert SAM to BAM
samtools sort -o aligned.sorted.bam aligned.bam
samtools index aligned.sorted.bam
samtools view -c aligned.sorted.bam chr11       # count reads on one chromosome
samtools flagstat aligned.sorted.bam            # the one-line alignment health check

flagstat is the quickest sanity check for any alignment: it prints how many reads mapped, how many are properly paired, and how many are duplicates or singletons.

VCF: variant calls with provenance

VCF (Variant Call Format) stores the positions where a sample differs from the reference, and it is the format through which most population genetics, clinical screening, and association studies actually flow. The fixed columns — CHROM, POS, ID, REF, ALT, QUAL, FILTER, INFO — describe the variant itself; the FORM and FORMAT columns then describe, per sample, what evidence supports it. Three habits worth forming early:

• Read the header. VCF headers are documentation, not decoration: the INFO and FORMAT keys are all defined there, and two files can look identical in columns while presenting entirely different quantities.
• Use a bgzip-compressed, tabix-indexed VCF for anything real, because then tools can stream only the regions you ask for.
• Treat quality and FILTER columns with respect; downstream association tools differ in whether they are applied on top of variant quality or instead of it, so it matters where filtering occurs.

The tool trio for the variant layer is bcftools (query, filter, merge, compare) plus bgzip and tabix for compression and indexing.

GFF and GTF: annotation coordinates

GFF3 and GTF are the annotation counterparts to the sequence files: one line per feature (gene, transcript, exon, repeat) with its coordinates and attributes. This is the layer your genome browser displays and what the genome annotation tools produce, as covered in the genome annotation guide. The bioinformatics detail that saves hours later: GFF3 and GTF use inclusive, one-based coordinates (position 1 is the first base), while almost every programmatic API and the BED format use zero-based, half-open intervals. A glide across the boundary is how real analyses end up shifted by exactly one base.

For a browser-track of regions rather than gene models, BED is the lightweight cousin: chromosome, start, end, optional columns, usually zero-based.

The format map of a real project

Stage	Format in	Format out	Key tools
Raw sequencing	images and base calls	FASTQ	bcl2fastq-type converters
QC trim and filter	FASTQ	FASTQ	fastp-style preprocessors
Alignment to reference	FASTQ	SAM/BAM/CRAM	BWA, Bowtie, minimap2
Post-processing alignments	BAM	BAM + index	samtools suite
Variant discovery	BAM + reference	VCF	bcftools, GATK-class callers
Annotation and features	genome FASTA + BAM	GFF3, bedGraph	annotation pipelines
Variants in context	VCF + annotation	annotated VCF, tables	bcftools annotate

Read this table top to bottom and you are reading the storyline of a project: sequences become alignments, alignments become variants, variants meet annotation. Every job on a bioinformatics job board assumes you can move a dataset across these format boundaries without hand-wringing.

Common failure modes

• Working on unsorted, unindexed BAM files. Most tools stream; when the file is not coordinate-sorted and indexed, some run to completion but silently behave badly, and others stall forever.
• Assuming positions agree between formats. One-based GFF lines next to zero-based BED regions are a real, documented off-by-one; always note which convention the tool you are writing against uses.
• Treating FASTQ quality as a cosmetic column. Trimming and filtering thresholds are chosen from the quality string; ignoring it means downstream errors are inherited rather than caught.
• Editing a header by hand and breaking what follows. BAM and VCF headers are structured records. When you must add or change something, use the tooling rather than a text edit, and keep a validation step.
• Copying one file out of a multi-file group. A BAM travels with its index and a VCF with its tbi. Moving one without the other can produce a dataset that quietly opens in some tools and fails in others.

Key takeaways

• Each format answers one claim type: sequences, alignments, annotations, or variants; shape your reading around that claim.
• FASTQ and FASTA are the sequence layer; SAM/BAM/CRAM is the alignment layer; VCF is the variant layer; GFF/GTF/BED is the annotation layer.
• Sorted and indexed is the state that real tools expect: .bai for BAM, tbi for VCF.
• GFF/GTF coordinates are one-based and inclusive; BED and most APIs are zero-based half-open. Off-by-one is a real bug, not a pedantic worry.
• The pipeline story — sequences to alignments to variants to annotation — is the same in every lab; only the tools change names.

Challenge

Take any public dataset associated with a published study, walk one sample through the table above, and name the file format present at every stage. If your dataset has no explicit VCF or GFF, note which stage your lab substitutes instead. Tight scope on one dataset makes the format landscape less abstract and gives you something concrete to discuss in an interview.

Related: The BAM to VCF half of the format chain has a pipeline of its own — see our variant calling guide.

Need help reading real datasets in a course or research project? Ampersand Academy offers one-to-one mentoring in bioinformatics and data analysis — part of Mahadhi, a development and digital marketing company.

FAQ