Reading Between The Base Pairs
Mutations are the raw material of evolution, but they're also a headache if you spend any time actually looking at genomic data. I'm going to walk through what genetic mutations are, how to classify them properly, and where most people get it wrong. The short version is that most introductory courses oversimplify the whole process, and that leads to sloppy interpretation later on. A mutation is simply any change in the DNA sequence. That's the textbook definition and it's technically correct, but it doesn't tell you much about what happens when you're staring at a VCF file from a sequencing run. The real work starts when you categorize the change: point mutations, insertions, deletions, copy number variations, inversions, translocations, and so on. Each class behaves differently in downstream analysis. I spent about three years working in clinical genomics before moving into research, and the thing that always tripped people up was the distinction between somatic and germline mutations. Germline mutations are inherited and present in every cell. Somatic mutations accumulate over a lifetime and are restricted to certain tissues. You need to know which one you're dealing with before you even think about interpreting pathogenicity, because the framework for assessment changes completely.
Here's a practical example that came up recently in my lab. We were looking at a patient with a suspected hereditary cancer syndrome and found a missense variant in BRCA1. On paper, it looked like a straightforward pathogenic finding. But when I checked the allele frequency in gnomAD, it was sitting at about 0.03%, which is higher than expected for a highly penetrant disease gene. We ended up discovering it was a somatic event in a contaminating cell line rather than a germline variant. The sample prep had been done on a mixed culture, and the bioinformatics pipeline flagged it as heterozygous because the variant allele fraction was around 40%. A quick check of the tumor purity and a re-extraction from the original biopsy resolved the issue.
How Mutations Actually Happen
DNA replication errors account for a large portion of spontaneous mutations. Every time a cell divides, polymerases copy roughly three billion base pairs, and despite proofreading activity, there's an error rate of about one in every ten billion nucleotides. That seems low until you consider how many cells divide in a human body over a lifetime. Most of these errors get repaired by mismatch repair pathways, but some slip through. Environmental factors are the other major driver. UV radiation causes pyrimidine dimers, which lead to C to T transitions if not repaired properly. Ionizing radiation creates double-strand breaks that get misrepaired, often resulting in deletions or translocations. Chemical mutagens like benzo[a]pyrene from cigarette smoke form DNA adducts that distort the helix and cause mispairing during replication. The thing about mutagenesis that people don't always appreciate is that mutation rates vary across the genome. CpG islands are hotspots for C to T transitions because cytosine can be methylated, and methylated cytosine spontaneously deaminates to thymine. This is one of the most common single nucleotide changes in human DNA, and it's completely predictable if you know where to look.
Get the Full Details
Interpreting Mutation Data
When you're analyzing sequencing data, the first step is variant calling. Tools like GATK's HaplotypeCaller or DeepVariant will scan your aligned reads and identify positions that differ from the reference genome. The output is usually a VCF file containing information about each variant: chromosome, position, reference allele, alternate allele, quality scores, and various metadata fields. The quality metrics matter more than most beginners realize. A variant with a low Phred-scaled quality score might look interesting but could be a sequencing artifact. I've seen people waste weeks chasing variants with quality scores below 20, which means there's only about a 1 in 100 chance that the call is correct. Filtering for quality above 30 or 40 is standard practice, and it eliminates a huge amount of noise without sacrificing real signal. Annotation is the next critical step. You need to know what effect each variant has on the resulting protein or regulatory element. Tools like ANNOVAR, VEP, or SnpEff will add information about whether a variant is synonymous, missense, nonsense, frameshift, splice site, or intergenic. This is where you separate the noise from the biologically relevant changes. A synonymous variant in an exon might look like it does nothing, but it could affect splicing or mRNA stability, and we have better tools now for detecting those effects than we did five years ago.
One nuance that caught me off guard early in my career: not all loss-of-function variants are created equal. A frameshift mutation near the beginning of a gene is likely to produce a truncated, nonfunctional protein. But a frameshift near the 3' end might leave most of the protein intact and functional. Similarly, in-frame deletions that remove just a few amino acids can have dramatically different effects depending on where they occur. A deletion in a catalytic domain will wreck enzyme activity, while the same size deletion in a flexible linker region might have no effect at all.
The Complications Nobody Warns You About
Structural variants are where things get messy. Standard short-read sequencing struggles with large insertions, deletions, inversions, and translocations. You can detect some of these from paired-end read mapping and split-read analysis, but the sensitivity drops significantly for variants larger than 50 base pairs or in repetitive regions. If you need to resolve complex structural rearrangements, long-read sequencing from Pacific Biosciences or Oxford Nanopore is far more reliable, though the cost per sample is still higher. Ploidy issues are another common problem. Most variant callers assume diploidy, but cancer samples are frequently aneuploid. When you're calling variants in a tumor with abnormal chromosomal content, the allele frequencies don't match the expectations for a diploid genome. You need specialized callers like FacETS or ABSOLUTE that account for copy number changes when estimating variant allele fractions. I once analyzed a melanoma sample where a known driver mutation appeared at about 15% allele frequency instead of the expected 50% for a heterozygous call, and it took two days to figure out that the locus had been deleted on one allele and amplified on the other. Contamination is the silent killer of variant calling. Cross-sample contamination can introduce false positives, especially at low allele frequencies. The typical threshold for acceptable contamination is below 2%, but even 1% contamination can create artifacts that look like real low-frequency somatic mutations. Tools like VerifyBamID can estimate contamination levels from your own data by checking x chromosome heterozygosity in male samples or using known SNP profiles.
What To Do When Your Variant Doesn't Make Sense
Sometimes you find a mutation that appears pathogenic based on all the standard criteria, but the phenotype doesn't match. This happens more often than you'd think. Penetrance is rarely 100% for most variants, and modifier genes, environmental factors, and epigenetic changes can all influence whether a mutation actually manifests as disease. I had a case where a family carried a known pathogenic variant in MLH1 associated with Lynch syndrome, but only one out of six carriers developed cancer. The non-penetrant carriers happened to have higher expression of alternative mismatch repair genes, which partially compensated for the defective MLH1. When interpretation hits a wall, check the ClinVar database and the LOVD databases for disease-specific genes. These aggregate submissions from multiple laboratories and can reveal whether other groups have seen the same variant with different interpretations. Discrepancies between labs are common, especially for variants of uncertain significance, and having a larger sample size of observations helps resolve ambiguity. Functional assays remain the gold standard for establishing causality. If computational predictions and segregation analysis are inconclusive, you might need to express the variant protein in a cell line and measure its activity directly. This is expensive and time-consuming, usually taking anywhere from three to six months for a well-designed experiment, but it's often the only way to settle a disputed variant classification.
Where The Field Is Heading
Single-cell sequencing is changing how we think about mutations. Bulk sequencing averages signals across thousands or millions of cells, which masks heterogeneity within a tumor or tissue. Single-cell DNA sequencing can reveal subclonal populations and trace how different mutations co-occur within individual cells. The technology is still developing, and error rates are higher than bulk sequencing, but the resolution it provides for understanding mutation patterns is unmatched. Machine learning approaches to variant interpretation are improving rapidly. Tools like REVEL and MVP integrate multiple prediction algorithms into single scores that correlate better with pathogenicity than any individual method. These aren't perfect, but they reduce the number of variants that end up classified as uncertain significance, which is still a significant portion of clinically actionable findings. The biggest limitation in the field right now is the reference bias problem. Most variant calling pipelines align reads to a single reference genome, which is derived from a small number of individuals of predominantly European ancestry. Variants that are common in other populations may map poorly to this reference, leading to missed calls or false positives. This is a known issue with ongoing efforts to build more diverse reference panels, but it hasn't been fully resolved yet, and it affects the accuracy of mutation detection in underrepresented populations.
If you're getting into this work, start by practicing on publicly available datasets from the 1000 Genomes Project or the Genome in a Bottle consortium. These have well-characterized truth sets that let you validate your pipelines against known ground truth. The skills you build there translate directly to real clinical and research data, and having a validated workflow before you touch patient samples saves a lot of headaches down the line.
