Large-scale inference of conjunctive Bayesian networks

Abstract

Supplementary data are available at Bioinformatics online.

Every which way? On predicting tumor evolution using cancer progression models

“…Two of the models used, CBN and CAPRI, have been used in two variants (H-CBN and MCCBN for CBN, CAPRI AIC and CAPRI BIC for CAPRI), yielding a total of six different procedures for obtaining CPMs. Only a brief overview of these CPMs is provided here; detailed descriptions can be found in the original references for each model: H-CBN (Gerstung et al, 2009(Gerstung et al, , 2011 , MCCBN (Montazeri et al, 2016), OT (Desper et al, 1999;Szabo and Boucher, 2008), CAPRI (Caravagna et al, 2016;Ramazzotti et al, 2015), and CAPRESE (Olde Loohuis et al, 2014). (Other CPMs exist but they have not been considered here because they are too slow for routine work, have no software available, or have dependencies on non-open source external libraries -see S4 Text).…”

“…A key difference between OT and CAPRESE is that CAPRESE reconstructs these models using a probability raising notion of causation in the framework of Suppes' probabilistic causation, whereas in OT weights along edges can be directly interpreted as probabilities of transition along the edges by the time of observation (Szabo and Boucher, 2008, p. 4). In contrast to OT and CAPRESE, both CAPRI (Caravagna et al, 2016;Ramazzotti et al, 2015) and CBN (Gerstung et al, 2009(Gerstung et al, , 2011Montazeri et al, 2016) allow modeling the dependence of an event on more than one previous event: the output of the models are directed acyclic graphs (DAGs) where some nodes have multiple parents, instead of a single parent (as in trees). CAPRI tries to identify events (alterations) that constitute "selective advantage relationships", again using probability raising in the framework of Suppes' probabilistic causation.…”

“…We have used two versions of CAPRI, that we will call CAPRI AIC and CAPRI BIC, that differ in the penalization used in the maximum likelihood fit, Akaike Information Criterion (AIC), or Bayesian Information Criterion (BIC), respectively. For CBN we have also used two variants, H-CBN, described in Gerstung et al (2009Gerstung et al ( , 2011) that uses simulated annealing with a nested expectation-maximization (EM) algorithm for estimation, and MCCBN, described in Montazeri et al (2016), that uses a Monte-Carlo EM algorithm. Thus, these six procedures can be divided into three groups: models that return trees (OT and CAPRESE) and two families of models that return DAGs, CBN (H-CBN and MCCBN) and CAPRI (CAPRI AIC and CAPRI BIC).…”

“…This we did for all models. From H-CBN and MCCBN we can also obtain the estimated probability of each path of tumor progression to the fitness maximum, since both H-CBN and MCCBN return the parameters of the waiting time to occurrence of each mutation (given its restrictions are satisfied; e.g., p. i729 in Montazeri et al, 2016, section 2.2 in Gerstung et al, 2009, or Hosseini, 2018 details and example in section 3 of S4 Text). It is also possible to perform a similar operation with the output of OT, and use the edge weights from the fits of OT to obtain the probabilities of transition to each descendant genotype and, from them, the probabilities of the different paths to the global maximum.…”

“…Cancer progression models (CPMs), such as conjunctive Bayesian networks (CBN) (Gerstung et al, 2009(Gerstung et al, , 2011Montazeri et al, 2016), oncogenetic trees (OT) (Desper et al, 1999;Szabo and Boucher, 2008), CAncer PRogression Inference (CAPRI) (Caravagna et al, 2016;Ramazzotti et al, 2015), or CAncer PRogression Extraction with Single Edges (CAPRESE) (Olde Loohuis et al, 2014), can be used to predict paths of tumor progression. CPMs were originally developed to identify restrictions in the order of accumulation of mutations during tumor progression from cross-sectional data (Beerenwinkel et al, 2015.…”