- False-positive and false-negative assessment is an important part of quality control and performance evaluation in bioinformatics. Computational methods are widely used to identify genetic variants, predict genes, classify biological sequences, detect microorganisms, identify differentially expressed genes, and interpret molecular data. However, these methods do not always produce correct results. They may report findings that are not genuinely present or fail to detect findings that do exist. Understanding and measuring these errors is essential for evaluating analytical reliability, selecting appropriate computational tools, and interpreting biological results accurately.
- A false positive (FP) occurs when an analytical method reports a positive result even though the condition, feature, or biological event being investigated is absent. In variant calling, a false-positive variant is reported by the computational pipeline but is not supported by the trusted reference data or independent evidence. A false negative (FN) occurs when a genuine feature or event is present but the method fails to detect it. In variant analysis, a false-negative result means that a real genetic variant is present in the sample but is missing from the variant-calling output. These errors can occur in many areas of bioinformatics, including genome analysis, sequence classification, gene prediction, transcriptomics, metagenomics, proteomics, and disease diagnostics.
- False positives and false negatives are usually assessed by comparing computational predictions with a trusted reference or independently established result. The reference is commonly called the ground truth, truth set, or reference standard. In genetic variant analysis, this may be a well-characterized set of variants generated using independent sequencing technologies, extensive experimental validation, or carefully curated reference materials. In gene prediction, the reference may consist of experimentally supported gene annotations. In sequence classification, it may be a curated set of organisms or biological sequences with established identities.
- The comparison between predicted and reference results produces four main categories: true positives, false positives, true negatives, and false negatives. A true positive (TP) is a correct positive prediction, while a true negative (TN) is a correct negative prediction. A false positive is an incorrect positive prediction, and a false negative is an incorrect negative prediction. Together, these categories form a confusion matrix, which provides the basis for calculating several performance measures.
| Assessment category | Actual positive | Actual negative |
| Predicted positive | True positive (TP) | False positive (FP) |
| Predicted negative | False negative (FN) | True negative (TN) |
- In variant calling, a true positive represents a genuine variant correctly identified by the pipeline. A false positive represents a reported variant that is not present according to the trusted reference. A false negative represents a genuine reference variant that the pipeline failed to detect. A true negative represents a correctly identified reference state at an evaluable genomic position. In practice, counting true negatives across an entire genome can be difficult because the number of possible reference positions is extremely large. Variant benchmarking therefore commonly emphasizes precision, recall, and related measures rather than relying on overall accuracy alone.
- Precision measures the proportion of reported positive predictions that are actually correct. It is also known as positive predictive value (PPV). A high precision indicates that the method produces relatively few false-positive results among its reported positive findings.
[
\text{Precision} = \frac{TP}{TP+FP}
]
Recall measures the proportion of actual positive findings that are successfully detected. It is also called sensitivity or true-positive rate. A high recall indicates that the method misses relatively few genuine positive findings.
[
\text{Recall} = \text{Sensitivity} = \frac{TP}{TP+FN}
]
The false discovery rate (FDR) represents the proportion of reported positive findings that are false. It is the complement of precision.
[
\text{FDR} = \frac{FP}{TP+FP} = 1-\text{Precision}
]
The false-negative rate (FNR), also called the miss rate, represents the proportion of actual positive findings that are not detected.
[
\text{FNR} = \frac{FN}{TP+FN} = 1-\text{Recall}
]
Specificity, or the true-negative rate, measures the proportion of actual negative findings correctly identified as negative.
[
\text{Specificity} = \frac{TN}{TN+FP}
]
Accuracy measures the proportion of all evaluated cases that are classified correctly.
[
\text{Accuracy} = \frac{TP+TN}{TP+FP+TN+FN}
]
The F1 score combines precision and recall using their harmonic mean. It provides a single measure of the balance between avoiding false positives and detecting genuine positives.
[
F1 = 2 \times \frac{\text{Precision}\times\text{Recall}}
{\text{Precision}+\text{Recall}}
]
- These measures describe different aspects of analytical performance and should be selected according to the purpose of the study. Precision is particularly important when false-positive findings are costly or potentially harmful. Recall is especially important when missing genuine findings could lead to serious consequences. Specificity is useful when evaluating the ability to recognize negative cases, while the F1 score provides a combined measure of precision and recall. Accuracy may be misleading when positive cases are rare because a method can classify most observations as negative and still achieve a high overall accuracy.
- For example, consider a variant-calling pipeline that is evaluated against a trusted reference dataset. Suppose the pipeline identifies 900 genuine variants, reports 100 variants that are not supported by the reference, and misses 100 genuine variants. The resulting values are TP = 900, FP = 100, and FN = 100. Precision is 90%, while recall is also 90%. The F1 score is 90%. This indicates that the pipeline detects most genuine variants, but it also produces some incorrect calls and misses some real variants.
- The following Python example calculates common performance measures from a confusion matrix:
# Example confusion matrix values
TP = 900
FP = 100
TN = 9000
FN = 100
# Calculate performance measures
precision = TP / (TP + FP) if TP + FP else 0
recall = TP / (TP + FN) if TP + FN else 0
specificity = TN / (TN + FP) if TN + FP else 0
accuracy = (TP + TN) / (TP + FP + TN + FN)
f1 = (
2 * precision * recall / (precision + recall)
if precision + recall else 0
)
false_discovery_rate = (
FP / (TP + FP) if TP + FP else 0
)
false_negative_rate = (
FN / (TP + FN) if TP + FN else 0
)
print(“Precision:”, round(precision, 4))
print(“Recall:”, round(recall, 4))
print(“Specificity:”, round(specificity, 4))
print(“Accuracy:”, round(accuracy, 4))
print(“F1 score:”, round(f1, 4))
print(“False discovery rate:”, round(false_discovery_rate, 4))
print(“False negative rate:”, round(false_negative_rate, 4))
- The example illustrates how the same confusion matrix can be used to calculate multiple performance indicators. In real bioinformatics studies, the values should be calculated using clearly defined evaluation units, such as genomic positions, variant alleles, genes, transcripts, or classified sequences. The definition of a positive case and the method used to match predictions with reference results must be documented.
- False-positive variants can arise from several technical and computational problems. Sequencing errors may introduce incorrect bases or artificial insertions and deletions. Low sequencing depth can produce unreliable evidence, while poor base quality may cause incorrect nucleotide observations. Mapping errors occur when reads are aligned to the wrong genomic location, particularly in repetitive regions, segmental duplications, and regions containing highly similar gene copies. PCR artifacts, contamination, sequencing-library preparation problems, and incorrect sample identification can also contribute to false-positive calls.
- Variant-calling algorithms may additionally report false positives because of limitations in their statistical models, local assembly procedures, filtering thresholds, or assumptions about ploidy and allele frequency. For example, a caller may interpret a cluster of mismatches caused by an alignment error as a genuine variant. A variant supported by duplicated reads may appear more reliable than it actually is if duplicate molecules are not properly identified. In somatic variant analysis, sequencing artifacts and low-level contamination may be mistaken for low-frequency tumour mutations.
- False-negative variants can result from insufficient sequencing depth, poor read quality, alignment failure, overly strict filtering, or limitations in the variant-calling algorithm. A genuine variant may be missed if too few reads support it or if the supporting reads are assigned low mapping quality. Variants in highly repetitive or structurally complex regions may be difficult to detect because reads cannot be uniquely aligned. Large insertions, complex indels, and structural variants may be missed by methods designed primarily for SNVs and short indels.
- False negatives can also occur when the selected reference genome does not adequately represent the sample’s genomic structure. Highly divergent regions, insertions absent from the reference, alternate haplotypes, and population-specific genomic sequences may cause reference-mapping bias. A caller may also fail to identify a variant because its filtering criteria are too stringent. In somatic analysis, low tumour purity, low variant allele frequency, subclonal variation, and normal-cell contamination can make genuine mutations difficult to distinguish from background sequencing errors.
- The balance between false positives and false negatives is strongly influenced by analytical thresholds. In variant calling, quality filters are used to remove unreliable predictions based on measures such as read depth, genotype quality, mapping quality, allele balance, strand bias, and variant quality. Increasing the stringency of these filters may reduce false positives, but it can also remove genuine variants and increase false negatives. Relaxing the filters may improve sensitivity but allow more incorrect calls. Therefore, thresholds should be selected according to the purpose of the analysis and evaluated using suitable benchmark datasets.
- Receiver operating characteristic (ROC) analysis is commonly used to examine the relationship between sensitivity and false-positive rate across different classification thresholds. The ROC curve plots the true-positive rate against the false-positive rate at multiple threshold settings. The area under the ROC curve (AUC) summarizes the ability of a classifier to distinguish positive from negative cases across these thresholds. A larger AUC generally indicates better discrimination, although ROC analysis may be less informative when positive cases are extremely rare.
- Precision–recall (PR) analysis is often more informative than ROC analysis for highly imbalanced datasets. It plots precision against recall across different thresholds. This is particularly relevant in bioinformatics applications where the number of true positive findings is small compared with the number of negative observations. Examples include rare variant detection, identification of disease-associated mutations, detection of rare microorganisms, and prediction of functional genomic elements. The area under the precision–recall curve provides a summary of performance, although the baseline precision depends on the prevalence of positive cases in the evaluated dataset.
- False-positive and false-negative assessment is important in many areas beyond variant calling. In gene prediction, a false positive may occur when a computational tool predicts a gene in a region that does not contain a genuine gene, while a false negative occurs when a real gene is missed. In RNA sequencing analysis, false-positive differential-expression findings may arise from technical variation, inappropriate statistical assumptions, or insufficient correction for multiple testing. False negatives may occur when genuine expression differences are too small to detect with the available sample size or sequencing depth.
- In metagenomics, false-positive classification can occur when sequencing reads from one organism are incorrectly assigned to another because of shared sequence regions, contamination, or incomplete reference databases. False negatives may occur when an organism is absent from the reference database, is present at very low abundance, or cannot be distinguished from closely related organisms. In protein functional annotation, a false positive may result from transferring an incorrect functional prediction based on sequence similarity, while a false negative may occur when a genuine function is not recognized because the relevant sequence or structural evidence is too divergent.
- The interpretation of false-positive and false-negative rates also depends on the prevalence of the feature being investigated. When a positive finding is rare, even a method with high specificity may produce a substantial proportion of false-positive results among all reported positives. This is important in clinical genomics and disease screening, where the prevalence of a particular variant or condition may be low. Positive predictive value depends not only on sensitivity and specificity but also on the prevalence of the positive condition in the tested population. Consequently, performance estimates from one dataset may not transfer directly to a population with a different distribution of variants or disease prevalence.
- Benchmark datasets are essential for reliable assessment. In human germline variant analysis, resources from the Genome in a Bottle (GIAB) Consortium provide well-characterized reference samples and high-confidence genomic regions. Tools such as hap.py support comparison of small-variant call sets against benchmark truth sets using haplotype-aware methods. For structural-variant assessment, Truvari can compare call sets while accounting for differences in variant representation and matching criteria. In other areas of bioinformatics, appropriate reference datasets may include curated gene annotations, experimentally validated protein functions, known microbial reference sequences, or independently measured expression data.
- The quality of the reference dataset itself must be considered. A truth set may be incomplete or uncertain in difficult genomic regions. If a genuine variant is missing from the reference set, a correct prediction may incorrectly appear to be a false positive. Similarly, if the reference contains an incorrect annotation, a correct computational prediction may be counted as a false negative. Benchmarking should therefore use trusted reference data, clearly defined evaluation regions, and appropriate matching procedures. Uncertain or unevaluable cases should not automatically be classified as false positives or false negatives.
- A further challenge is the comparison of equivalent biological variants represented differently in computational output. Insertions and deletions in repetitive regions may have multiple equivalent coordinate representations. Complex variants may also be described using different combinations of SNVs and indels. Normalization and haplotype-aware comparison can reduce incorrect mismatches, but researchers must still use compatible reference assemblies, variant representations, and evaluation settings. For structural variants, matching criteria may include breakpoint distance, event size, sequence similarity, and variant type. The selected criteria can affect the measured false-positive and false-negative rates.
- In clinical bioinformatics, false-positive and false-negative results have different potential consequences. A false-positive pathogenic variant may lead to unnecessary follow-up testing, anxiety, inappropriate clinical decisions, or incorrect family risk assessment. A false-negative result may cause a disease-associated variant to be overlooked, potentially delaying diagnosis or preventing appropriate management. Clinical laboratories therefore use validated procedures, quality-control standards, appropriate confirmation methods, and established interpretation guidelines. Technical confirmation of a variant should also be distinguished from interpretation of its clinical significance: proving that a variant is present does not automatically establish that it causes disease.
- False-positive and false-negative assessment should be incorporated into the design and reporting of bioinformatics workflows. Researchers should define the positive and negative categories, select an appropriate reference dataset, establish the evaluation region, and choose performance measures suited to the intended application. They should report the number of true positives, false positives, false negatives, and, where meaningful, true negatives, alongside precision, recall, specificity, and other relevant measures. Confidence intervals can help communicate uncertainty in performance estimates, especially when the number of evaluated positive cases is limited.
- Reproducibility requires documentation of the sequencing technology, sample characteristics, reference genome, software versions, parameters, filtering thresholds, benchmark dataset, and comparison method. Researchers should preserve the original predictions and benchmark outputs so that results can be independently reviewed. Performance should ideally be assessed on independent test data rather than only on the dataset used to develop or tune the method. This helps reduce overfitting and provides a more realistic estimate of how the method may perform on new samples.
- False-positive and false-negative assessment is fundamental to determining the reliability of computational analysis in bioinformatics. False positives indicate incorrect findings that are reported as genuine, while false negatives represent real biological findings that remain undetected. Their measurement through confusion matrices, precision, recall, specificity, F1 scores, ROC curves, and precision–recall analysis provides a systematic way to evaluate analytical performance. When combined with trusted benchmark datasets, independent validation, appropriate thresholds, and reproducible reporting, these assessments help researchers select reliable methods and make more confident biological interpretations.