Reading Your Own Genome: What Reads Are, How Mapping Works, and Which Tools to Run

There is no single thing called a genome reader. The phrase collapses three separate pieces of machinery. The sequencer produces short strings of bases called reads. The alignment and variant-calling stack places those reads on a reference and works out where you differ from it. The browser or annotation layer turns a list of differences into something you can look at. If you have a whole-genome sequencing result and want to read it yourself, you are working with the second and third of those. This page covers the file formats, the tools, the flags, and the places the pipeline quietly gets things wrong.

(A fourth meaning exists in the literature: “reader” proteins, the domains that bind chemical marks on histones. More on that at the end, because searches for the term land on both.)

What a read is

A read is a short stretch of sequence the instrument believes it observed, plus a per-base confidence score. On Illumina, a read is typically 150 bases, produced in pairs from opposite ends of a ~350 bp fragment. On Oxford Nanopore or PacBio, a read is 10–100 kb of contiguous sequence.

Reads arrive as FASTQ: four lines per read, with the fourth line encoding Phred quality per base. Phred 30 means a 1-in-1,000 chance that base is wrong. A 30× human genome at 2×150 bp is roughly 300 million read pairs and around 100 GB of gzipped FASTQ. “30×” means each position in your genome is covered by about 30 reads on average, which is the coverage at which germline SNV calling becomes reliable. Coverage is not uniform: GC-rich promoters and long homopolymers dip well below the mean.

Reads are the atomic unit. Everything downstream is inference on top of them, and when a result looks strange, you go back and look at the reads.

From reads to a genome

Three steps: align, sort and dedup, call variants.

For short reads we would use bwa-mem2 for alignment. It is the same algorithm as bwa mem with better memory layout, roughly 2–3× faster, and produces near-identical output:

bwa-mem2 mem -t 32 -R '@RG\tID:s1\tSM:sample\tPL:ILLUMINA' \
  GRCh38_no_alt.fa R1.fq.gz R2.fq.gz \
  | samtools sort -@ 8 -o sample.bam -
samtools index sample.bam

For long reads, minimap2 -ax map-ont or -ax map-hifi. Use samtools markdup (or gatk MarkDuplicates) before calling, because PCR duplicates inflate apparent support for an allele.

On reference build: use the GRCh38 no-alt analysis set. Not the plain GRCh38 with ALT contigs, unless your aligner is ALT-aware, because otherwise reads scatter across alternate haplotypes and mapping quality collapses to zero in exactly the regions you care about. T2T-CHM13 is a better reference biologically, with the centromeres and acrocentric arms finished, but most annotation databases are still keyed to GRCh38 coordinates. Our practice: call on GRCh38 for interpretation, and keep a second T2T alignment if you want to look at previously unresolvable regions.

Then variant calling. DeepVariant converts the read pileup at each candidate site into an image-like tensor and classifies the genotype with a convolutional network. It replaced hand-tuned statistical filters with a learned model and generally beats GATK HaplotypeCaller on SNV precision and recall in benchmark regions, especially on non-Illumina chemistries where the error profile is unusual. This is the most concrete place deep learning sits in a consumer genome pipeline, and it is part of a broader shift toward learned models across sequencing analysis¹.

run_deepvariant --model_type=WGS --ref=GRCh38_no_alt.fa \
  --reads=sample.bam --output_vcf=sample.vcf.gz \
  --output_gvcf=sample.g.vcf.gz --num_shards=32

Output is VCF: one row per variant site, with genotype, depth, and allele balance. A 30× human genome yields roughly 4–5 million variants against GRCh38, the large majority of them common and already catalogued.

Store the alignment as CRAM. Do not store it as BAM. CRAM references the FASTA instead of storing matched bases, cutting a ~70 GB BAM to roughly 20 GB. The tradeoff is that you must keep the exact reference FASTA forever, or the file is unreadable.

Reading the variants

Four to five million rows is not something you read. Annotate first.

Ensembl VEP, run offline against a local cache, is what we would use:

vep -i sample.vcf.gz --cache --offline --assembly GRCh38 \
  --everything --fork 8 --vcf -o annotated.vcf

--everything pulls in consequence terms, SIFT/PolyPhen, gnomAD frequencies, and regulatory overlap. Add ClinVar as a custom annotation. Then filtering is ordinary data work: drop anything with gnomAD allele frequency above ~1% if you are looking for rare coding variation, keep ClinVar pathogenic and likely-pathogenic calls, keep predicted loss-of-function in genes with known LoF mechanisms.

Two things to hold onto. First, most of your variants are non-coding, and the functional consequence of non-coding variation remains poorly characterised. The ENCODE pilot annotated functional elements across 1% of the genome and already showed pervasive transcription and regulatory activity outside genes². Sequence availability was never the same as understanding³. Second, a variant flagged pathogenic in your VCF is a hypothesis. It is not a finding. Research-grade sequencing is not a diagnostic test, classification depends on family history and phenotype, and anything that looks medically consequential needs confirmation in a CLIA/CAP laboratory and interpretation with a clinical geneticist or genetic counselor. Do not act on a VCF row.

Where the pipeline lies to you

• Segmental duplications and pseudogenes. PMS2, CYP21A2, SMN1, and STRC all have near-identical paralogs. Short reads map ambiguously, mapping quality drops, and callers either drop the site or produce false heterozygous calls. Check MQ in the VCF before believing anything in these genes.
• CYP2D6. Structural variation, hybrid alleles, and copy number changes mean a generic caller will misassign the star allele. Use a purpose-built tool (Cyrius, StellarPGx) or do not report it.
• Homopolymers and short tandem repeats. Indel calls in runs of 8+ identical bases are unreliable on short reads. Long reads fix most of this.
• Repeat expansions. ExpansionHunter handles known loci with a targeted model. A general-purpose caller will not see them at all.
• Structural variants. A standard SNV pipeline finds none of them. Manta or Delly for short reads, Sniffles2 for long reads, and expect a high false-positive rate without manual review.

For manual review, IGV loaded with the CRAM and the VCF. Look at allele balance, strand bias, and whether the supporting reads all start at the same position (a duplicate artifact). The UCSC Genome Browser and NCBI’s Genome Data Viewer are better for the reference context: conservation tracks, regulatory annotation, gene models.

Where machine learning genuinely helps

Gene structure prediction has been a modelling problem since GENSCAN used a probabilistic model of exon-intron structure to predict complete gene structures from raw sequence⁴. What changed is the model class and the amount of labelled data. Deep networks are now used across variant calling, splice-site prediction, regulatory element classification, and functional effect scoring⁵. Transformer-based DNA language models trained on raw sequence are the current active area, aiming to produce general-purpose representations that transfer across tasks⁶. Machine learning methods for sequence analysis are being applied across classification and assembly problems as well⁷. Clinically, AI-assisted whole-genome interpretation is being evaluated for decision support, and reviews are consistent that the evidence base is still thin outside a few well-defined settings⁸.

The other genome reader: reader proteins

In the molecular biology literature, a reader is a protein domain that binds a specific chemical modification on a histone or on DNA and translates it into a downstream effect. Quantitative interaction proteomics has mapped these readers systematically by using modified histone peptides as bait and identifying what binds⁹. Bromodomains, the readers of acetylated lysine, are a druggable example: domain-selective BET ligands distinguish the first and second bromodomains of BET proteins and produce different cellular outcomes depending on which they engage¹⁰. If you arrived here from a chromatin paper, that is the sense being used.

Questions people also ask

What are genome reads? Short sequences the instrument reports, each with per-base quality scores, stored in FASTQ. Illumina reads are typically 150 bases in pairs, long-read platforms produce 10–100 kb. A genome is reconstructed by mapping many overlapping reads onto a reference, which is why coverage depth matters more than total data volume.

Is it true that researchers are using AI to decode the human genome? Yes, in specific places. DeepVariant uses a CNN to call genotypes from read pileups, splice and regulatory effect predictors are neural networks, and deep learning is applied broadly across next-generation sequencing analysis¹. It is applied statistics on sequencing data. It is not a system that reads out meaning.

Can AI decode the human genome? It can predict some functional consequences from sequence, with accuracy that varies by task. Coding-variant effect prediction is decent, non-coding regulatory prediction is improving, and cell-type-specific and long-range effects remain hard⁵. Nothing available today converts raw sequence into a reliable account of what it does in you.

Has the human genome been decoded? The sequence is essentially complete: the T2T-CHM13 assembly published in 2022 closed the remaining gaps in one haplotype. Function is a different question. Annotation of functional elements has been an ongoing project since the ENCODE pilot covered 1% of the genome², and the practical gap between having sequence and understanding it was clear from the start³.

Footnotes

1. Wardah S. Alharbi, Mamoon Rashid. A review of deep learning applications in human genomics using next-generation sequencing data. Human Genomics, 2022. https://doi.org/10.1186/s40246-022-00396-x ↩ ↩²

2. The ENCODE Project Consortium. Identification and analysis of functional elements in 1% of the human genome by the ENCODE pilot project. Nature, 2007. https://doi.org/10.1038/nature05874 ↩ ↩²

3. Eric S. Lander. Initial impact of the sequencing of the human genome. Nature, 2011. https://doi.org/10.1038/nature09792 ↩ ↩²

4. Chris Burge, Samuel Karlin. Prediction of complete gene structures in human genomic DNA. Journal of Molecular Biology, 1997. https://doi.org/10.1006/jmbi.1997.0951 ↩

5. K.O. Mohammed Aarif, V. Mohammed Yousuf Hasan, Afroj Alam, et al. Decoding DNA: Deep learning’s impact on genomic exploration. Deep Learning in Genetics and Genomics, 2025. https://doi.org/10.1016/b978-0-443-27574-6.00005-9 ↩ ↩²

6. Muhammad Nabeel Asim, Muhammad Ali Ibrahim, Arooj Zaib, et al. DNA sequence analysis landscape: a comprehensive review of DNA sequence analysis task types, databases, datasets, word embedding methods, and language models. Frontiers in Medicine, 2025. https://doi.org/10.3389/fmed.2025.1503229 ↩

7. Sravani C., Pavani P., Vybhavi G.Y., et al. Decoding the Human Genome: Machine Learning Techniques for DNA Sequencing Analysis. E3S Web of Conferences, 2023. https://doi.org/10.1051/e3sconf/202343001067 ↩

8. Kokiladevi Alagarswamy, Wenjie Shi, Aishwarya Boini, et al. Should AI-Powered Whole-Genome Sequencing Be Used Routinely for Personalized Decision Support in Surgical Oncology—A Scoping Review. BioMedInformatics, 2024. https://doi.org/10.3390/biomedinformatics4030096 ↩

9. Michiel Vermeulen, H. Christian Eberl, Filomena Matarese, et al. Quantitative Interaction Proteomics and Genome-wide Profiling of Epigenetic Histone Marks and Their Readers. Cell, 2010. https://doi.org/10.1016/j.cell.2010.08.020 ↩

10. Ashraf Mohammed, M. Brett Waddell, Ieva Sutkeviciute, et al. Domain-Selective BET Ligands Yield Next-Generation Synthetic Genome Readers/Regulators with Nonidentical Cellular Functions. Journal of the American Chemical Society, 2023. https://doi.org/10.1021/jacs.3c06297 ↩