More than half a century has passed since the discovery that reactive electrophilic metabolites derived from xenobiotic agents covalently modify endogenous cellular proteins [1
]. Since then such covalent binding by reactive metabolites has been strongly correlated with, and is widely believed to be responsible for, the acute organ-damaging effects of a wide range of xenobiotic agents including drugs and natural products [3
]. Tissue injury is a complex phenomenon. Most tissues are comprised of more than one cell type, and macroscopic tissue injury often involves various secreted chemical mediators such as cytokines, tumor necrosis factor, nitric oxide and pro-inflammatory cells such as leukocytes and macrophages or Kupffer cells. Thus it is important to note that the same compounds that can cause organ damage in vivo can also cause acute cytotoxicity, correlated to protein covalent binding, in isolated metabolically-competent cells in vitro.
It has generally been presumed that protein adduction impairs protein function, leading to disruption of metabolic or signaling pathways, organelle failure, loss of cellular homeostasis, aberrant cell-cell interactions, macroscopic tissue damage through necrosis and/or apoptosis, and in the extreme, organ failure and death. We know that some compounds that give rise to covalently bound residues do not cause toxicity, i.e., none of their adducts trigger events leading to cytotoxicity. Likewise some, perhaps many, of the adducts of toxic compounds are ineffective at causing toxicity, but some of the adducts of toxic compounds do trigger events leading to cytotoxicity (and eventually tissue damage and organ injury in vivo). The challenge is to identify which adduct structure(s) on which amino acid residue(s) of which protein(s) are important for toxicity, and the mechanism(s) by which their appearance triggers toxicity while others do not.
Despite extensive investigation, the mechanisms by which covalent binding events trigger cytotoxic outcomes remain largely unclear [6
]. A major reason for this gap is that only recently has it become technically feasible to identify numbers of individual proteins targeted by xenobiotic reactive metabolites. Early target protein identifications were based on isolating individual adduct-bearing proteins, one at a time, using traditional protein separation methods. By 1997, only 28 proteins targeted by xenobiotic reactive metabolites had been isolated and identified, largely by N-terminal sequencing [6
]. In 1998, however, the coupling of 2D gel electrophoresis with mass spectrometric methods of protein identification literally revolutionized the field [8
]. Since then both the number of known target proteins and the number of small-molecule adduct-forming xenobiotics studied have increased nearly ten-fold. To help keep track of this information and to facilitate its analysis, we recently built the Reactive Metabolite Target Protein Database (TPDB) [6
]. All proteins listed in this database came from studies using live animals or intact living cells.
As the number of known target proteins grew, so did the hope of elucidating mechanistic pathways connecting covalent adduction events to the observed cytotoxic outcomes. For example, we and others [7
] have attempted to make sense of lists of target proteins by arbitrarily grouping them into categories according to function, but this approach has done little to reveal to a unifying mechanism of toxicity caused by a variety of reactive metabolites [6
]. Another way to analyze target proteins is to sort them into Gene Ontology (GO) categories [15
] and determine whether any show an over-abundance of target proteins relative to statistical expectations. Similarly, KEGG biological pathway analysis [16
] can identify metabolic pathways in which target proteins are over-represented. Such analyses could potentially indicate a functional connection between target protein adduction and biological consequence in the context of systems biology [12
In living cells, proteins interact extensively with other proteins, forming protein-protein interaction (PPI) networks that sense and respond to the abundance or status of other network proteins. Since endogenous post-translational modifications of proteins are well-known to perturb PPI networks comprising intracellular signaling cascades [17
], it is plausible to hypothesize that protein modification through adduction by xenobiotic reactive metabolites could constitute an aberrant form of signaling leading to cytotoxic consequences. Thus, inspection of the interacting partners of reactive metabolite target proteins might shed new light on the path from protein adduction to cytotoxicity.
In addition to the question of how protein adduction leads to cytotoxicity, it is also of interest to know what features of a protein, beyond simple abundance, determine whether or not it is likely to be a target for reactive metabolites. Electrophilic xenobiotic metabolites can be classified broadly as having acylating or alkylating activity [6
]. The former tend to attack lysine side chains, while the latter tend to attack predominantly cysteine, histidine and lysine side chains. Despite the commonplace occurrence of these side chains in most proteins, it is well established that protein adduction in living cells is remarkably selective, with some abundant proteins experiencing little adduction while some low-abundance proteins experience high levels of adduction [7
]. Nevertheless, the protein features that determine susceptibility to adduction are almost completely unknown [21
]. Since many protein features can now be calculated or predicted using software programs, the analysis of target proteins using feature selection algorithms could potentially shed light on this important question [22
]. Feature selection algorithms operate differently from conventional statistical (correlative) studies of individual features considered independently from each other. Such algorithms can identify which features among many contribute the most to determining a complex behavior such as relative susceptibility to adduction by electrophiles.
In this paper we report our efforts to use bioinformatics approaches to elucidate the interactions between reactive metabolites, their cellular target proteins, and the other proteins that interact directly with the target proteins. In brief, we first analyzed 171 proteins targeted by reactive metabolites from one or more of 18 different protoxins and found that a number of GO categories and KEGG pathways were significantly enriched with some of these target proteins. We then selected 28 proteins known to be adducted by reactive metabolites of at least 3 different protoxins and found that 21 of them had a total of 165 directly interacting partners. GO and KEGG pathway analysis of the combined 186 proteins revealed several categories to be highly significantly enriched by target proteins and/or their directly interacting partner proteins. Finally, we applied machine learning methods to analyze the properties of 62 rat liver proteins targeted by reactive metabolites of thiobenzamide and 45 rat liver proteins targeted by reactive metabolites of bromobenzene, in an effort to identify properties that help to distinguish whether a protein is likely to be a target of reactive metabolites.