Evaluating and improving next generation sequencing performance in the absence of gold standard - Thèses Université Claude Bernard Lyon 1 Accéder directement au contenu
Thèse Année : 2023

Evaluating and improving next generation sequencing performance in the absence of gold standard

Évaluation et amélioration des performances du séquençage de nouvelle génération en absence de gold standard

Résumé

Evaluating next-generation sequencing (NGS) performance suffers frequently from the absence of gold standard. Without gold standard, researchers often carry out replicates from the same individual and use concordance between replicates to evaluate NGS performance, whereas the appropriateness of that criterion is still debated. Furthermore, for a better performance, the replicates are often combined using various models to reconstruct a new high-performance callset. This work aimed to investigate these two aspects of NGS performance evaluation and improvement in the absence of gold standard. In the first part, we examined the contributions and limitations of the concordance-discordance criterion. We analyzed the relationship between the probability of discordance and that of error using conditional probability under conditional independence and conditional dependence between two sequencing results. We compared the probabilities of discordance and error with various combinations of sensitivity, specificity, and correlation between replicates, then on real results of sequencing genome NA12878. We examined covariate effects on discordance and error using generalized additive models with smooth functions. The results showed that, with conditional independence of two sequencing results, the concordance-discordance criterion seems acceptable; however, it becomes questionable in presence of high correlation because of high percentages of false concordant results. Covariate effects’ functional forms were close between discordance and error models, though the parts of covariate-explained deviance differed. In the second part, we investigated the statistical methods able to combine callsets from replicates to reconstruct a new callset. Three technical replicates of genome NA12878 were considered and five model types were compared (consensus, latent class, Gaussian mixture, Kamila–adapted k-means, and random forest) regarding four performance indicators: sensitivity, precision, accuracy, and F1-score. We concluded that the compared non-supervised clustering models that combine multiple callsets are able to improve sequencing performance vs. supervised models previously tested elsewhere. Among the models compared, the Gaussian mixture model and Kamila offered non-negligible precision and F1-score improvements. These models may be recommended for callset reconstruction (from either biological or technical replicates) for diagnostic or precision medicine purposes
Les travaux de cette thèse ont porté sur la place du modèle de concordance-discordance dans l'évaluation de la performance du séquençage à haut débit et sur des comparaisons de performance entre modèles de classification dans la reconstitution de résultats de séquençage haut débit à partir de réplicas techniques. Le premier travail a étudié le problème de l'évaluation du séquençage haut débit en l'absence de ‘gold' standard et, dans ce cadre, la pertinence des critères de concordance-discordance. Il a examiné les relations entre le taux de discordance et le taux d'erreur dans diverses situations théoriques. Il a ensuite analysé les effets des covariables sur ces deux taux en utilisant un modèle additif généralisé avec des données réelles issues de réplicats de séquençage du génome NA12878. Le second travail a étudié le problème de la fusion de résultats de séquençage haut débit de réplicas techniques en vue d'obtenir un nouveau jeu de données susceptible de comporter moins d'erreurs. Il a évalué et comparé les aptitudes des principaux modèles de partitionnement à améliorer la performance finale du séquençage à partir des résultats de trois séquençages du génome NA12878. L'étude fournit des arguments pour choisir le modèle le plus convenable et utiliser ces résultats en matière de diagnostic ou de médecine de précision.
Fichier principal
Vignette du fichier
TH2023ZHAIYUE.pdf (2.17 Mo) Télécharger le fichier
Origine Version validée par le jury (STAR)

Dates et versions

tel-04649060 , version 1 (16-07-2024)

Identifiants

  • HAL Id : tel-04649060 , version 1

Citer

Yue Zhai. Evaluating and improving next generation sequencing performance in the absence of gold standard. Ecosystems. Université Claude Bernard - Lyon I, 2023. English. ⟨NNT : 2023LYO10121⟩. ⟨tel-04649060⟩
0 Consultations
0 Téléchargements

Partager

Gmail Mastodon Facebook X LinkedIn More