In nature, organisms rarely live in isolation, but instead coexist in complex ecologies with various symbiotic relationships 
. As defined in macroecology, observed relationships between organisms span a wide range including win-win (mutualism), win-zero (commensalism), win-lose (parasitism, predation), zero-lose (amensalism), and lose-lose (competition) situations 
. These interactions are also widespread in microbial communities, where microbes can exchange or compete for nutrients, signaling molecules, or immune evasion mechanisms 
. While such ecological interactions have been recently studied in environmental microbial communities 
, it is not yet clear what the range of normal interactions among human-associated microbes might be, nor how their occurrence throughout a microbial population may influence host health or disease 
Several previous studies have identified individual microbial interactions that are essential for community stability in the healthy commensal microbiota 
, and many are further implicated in dysbioses and overgrowth of pathogens linked to disease 
. Each human body site represents a unique microbial landscape or niche 
, and relationships analogous to macroecological “checkerboard patterns” 
of organismal co-occurrence have been observed due to competition and cooperation 
. For example, dental biofilm development is known to involve complex bacterial interactions with specific colonization patterns 
. Likewise, disruption of relationships among the normal intestinal microbiota by overgrowth of competitive pathogenic species can lead to diseases, e.g. colonization of Clostridium difficile
in the gut 
. However, no complete catalog of normally occurring interactions in the human microbiome exists, and characterizing these co-occurrence and co-exclusion patterns across body sites would elucidate both their contributions to health and the basic biology of their ecological relationships. Thus, characterizing key microbial interactions of any ecological type within the human body would serve as an important first step for studying and understanding transitions among various healthy microbial states or into disease-linked imbalances.
As has been also been pointed out in macroecology, however, the analytical methodology needed to comprehensively detect such co-occurrence relationships is surprisingly complex 
. Most existing studies employ simple measures such as Pearson's or Spearman's correlation to identify significant abundance relationships 
. These methods are suboptimal when applied without modification to organismal relative abundances 
. Since absolute microbial counts are not known and measurements depend on sampling and sequencing depth, an increase in one relative abundance must be accompanied by a compositional decrease in another, leading to spurious correlations among non-independent measurements 
. In addition, sparse sequence counts can cause artefactual associations for low-abundance organisms with very few non-zero observations 
. Conversely, association methods such as log-ratio based distances 
that have been developed specifically for such compositional data are difficult to assign statistical significance, a vital consideration in high-dimensional microbial communities containing hundreds or thousands of taxa.
Here, we have addressed these issues to catalog a baseline of normal microbial interactions in the healthy human microbiome. The Human Microbiome Project (HMP) 
sampled a disease-free adult population of 239 individuals, including 18 body habitats in five areas (oral, nasal, skin, gut, and urogenital), providing 5,026 microbial community compositions assessed using 16S rRNA gene taxonomic marker sequencing 
. We have developed a suite of methods to characterize microbial co-occurrence and co-exclusion patterns throughout the healthy human microbiome while suppressing spurious correlations. Specifically, these were 1) an ensemble approach including multiple similarity and dissimilarity measures, and 2) a compendium of generalized boosted linear models (GBLMs) describing predictive relationships, both assessed nonparametrically for statistical significance while mitigating the effects of compositionality. Together, these methods provide a microbiome-wide network of associations both among individual microbes and between entire microbial clades.
Among the 726 taxa and 884 clades in the HMP data, we examined both intra-body site and inter-body site relationships as a single integrated microbial co-occurrence network. Each relationship represents co-occurrence/co-exclusion pattern between a pair of microbes within or between body sites among all subjects in the HMP (in contrast to studies within single subjects of microbial co-occurrences across biogeography, e.g. 
). This ecological network proved to contain few highly connected (hub) organisms and was, like most biological networks, scale-free. Co-occurrence patterns of the human microbiome were for the most part highly localized, with most relationships occurring within a body site or area, and there were proportionally few strong correspondences spanning even closely related body sites. Each pair of organisms was assessed for positive (e.g. cooperative) or negative (e.g. competitive) associations, and in many cases these patterns could be explained by comparing the organisms' phylogenetic versus functional similarities. In particular, taxa with close evolutionary relationships tended to positively associate at a few proximal body sites, while distantly related taxa with functional similarities tended to compete. The resulting network of microbial associations thus provides a starting point for further investigations of the ecological mechanisms underlying the establishment and maintenance of human microbiome structure.