The CLR Algorithm: Methods and Comparison to Other Approaches

(A) A schema of the CLR algorithm. The z-score of each regulatory interaction depends on the distribution of MI scores for all possible regulators of the target gene (z_i) and on the distribution of MI scores for all possible targets of the regulator gene (z_j).

(B) Precision and recall for several different network inference methods applied to all genes in the E. coli microarray compendium were calculated using RegulonDB. The number of correctly inferred interactions (within RegulonDB) for each recall value is labeled on the top of the chart. All algorithms performed far better than the random method. Both CLR and relevance networks reach high precisions, but CLR attains almost twice the recall of relevance networks at some levels of precision.

(C) Using 60 well-chosen arrays, we can infer a network, nearly equivalent in recall and precision to the network inferred using all 445 microarrays in the compendium (dotted horizontal line), reflecting the redundancy of the compendium and the potential for improvement in choosing subsequent perturbations to profile.