An Automated Allele-calling System for High-throughput Microsatellite Genotyping

Microsatellite markers are widely used for genetic analysis in biomedical research, agriculture, population and evolutionary biology, as well as for forensics and diagnostics. Advances in laboratory automation and data collection have increased the throughput and have reduced the cost of large-scale genotyping. One step in the measurement process that still needs improvement is “allele calling”, where raw electrophoresis signals are converted into discrete genotypes. This is still largely a laborious manual process that constitutes more than a quarter of the genotyping cost. Automating allele calling is hampered by, among others, the presence of “stutter patterns” (artefact peaks introduced during PCR amplification) and variation in electrophoresis migration behavior. Both effects are marker specific, making it difficult to devise an algorithm that works for all markers without marker-specific calibration. This thesis proposes an allele calling method that consists of two main computer programs: (1) STRAL: a trace alignment algorithm that normalizes variation in the “time domain” of the observed chromatograms, and (2) FA: a pattern recognition algorithm that performs allele calling on the aligned chromatograms. Both are adaptive and do not require marker-specific calibration. For a given observation, each possible genotype is associated with a quality score related to the probability of calling error. This quality score can be used to rank and select the most likely genotype(s). Benchmark tests were performed on ∼33,000 genotypes taken arbitrarily from the daily output of a genotyping service laboratory. The performance is characterized by a trade-off between true calls and miscalls at a given cutoff of the quality value. At a level corresponding to less than 1% error (acceptable for most purposes), 55% of the data can be called correctly (or 70% of the data that can be called by human analysts). This performance is still far from manual calling (at 80% correct call of the total with < 0.2% error). However, it is useful for a hybrid system where up to 70% of the data is scored automatically, 15% of which is automatically rejected, and the remaining 30% needs to be manually examined, but with only 5% of them requiring corrections. We conclude that this prototype is worth implementing for actual appli-