Despite a continuously rising number of clinical trials the rate of bringing novel medication to the clinic is stalling (Pammolli et al., 2011). Here, Omics profiling and high throughput drug screening technologies at the interface of large scale clinical data have triggered novel conceptual strategies aimed at improved patient stratification for enabling precision medicine (Trusheim et al., 2011; Hollebecque et al., 2014). For implementing such approaches a number of issues need to be addressed including: (i) mirroring the clinical categorization of a phenotype on a molecular level description, (ii) spotting molecular factors mechanistically driving disease progression, (iii) drug-based intervention specifically addressing such progression mechanisms, and (iv) predictive biomarkers allowing fit-for-purpose analysis regarding a match of relevant pathophysiology and drug mechanism of action on the individual patient level (Heinzel et al., 2012).

A clinically well-established example is HER2 positive breast cancer characterized by overexpression of a member of the epidermal growth factor receptor family (ERBB2) playing a mechanistic role in progressive disease. In case the factor is proving positive for a patient the specific presentation is amenable for treatment tackling growth signaling (Hicks and Kulkarni, 2008). Still, the clinical presentation of breast cancer shows heterogeneous pathophysiologies apart HER2 positive subtypes. In consequence, when aiming at a comprehensive assessment of progressive breast cancer phenotypes multimarker panels are needed, e.g., implemented by a multiplexed assay holding 70 individual molecular features (Buyse et al., 2006). Such multimarker panels have generally become a promising strategy for characterizing complex clinical presentations, e.g., utilizing a serum marker panel for predicting coronary artery disease in symptomatic patients, or a urinary proteomics profile for early diagnosis of diabetic kidney disease (LaFramboise et al., 2012; Zürbig et al., 2012).

Failure for identifying a single causative factor as proxy for determining progression of a complex clinical phenotype becomes apparent when comparing the performance of marker panels with single markers, with the latter e.g., reviewed by Hellemons et al. for onset and progression of diabetic kidney disease (Hellemons et al., 2012). In clinical practice a different type of biomarker may be utilized, providing a phenotypic readout primarily reflecting the functional status of an organ in contrast to the pathophysiological characteristics. In kidney disease such functional markers are used in patient management as well as clinical trial design, including the estimated glomerular filtration rate (eGFR) and proteinuria (reflecting glomerular filtration and permeation of macromolecules across the glomerular capillary wall, respectively).

Association of these parameters with worsening of diabetic kidney disease, together with increasing incidence of endpoints as cardiovascular events is undisputed (Adler et al., 2003). However, these markers do not provide information on the specific molecular characteristics of the disease. Functional markers render stratification for tailored therapy in the concept of precision medicine essentially impossible.

The molecular pathway of primary interest in the present clinical setting of diabetic kidney disease is the renin-angiotensin system (RAS), in its activity at foremost controlling blood pressure and fluid balance. Blockade of the RAS has been able to reduce the incidence of renal events in patients with and without diabetes mellitus (Ruggenenti et al., 1998; Brenner et al., 2001). In a study by Lewis et al. angiotensin receptor blockade by Irbesartan reduced the risk of a primary composite endpoint (doubling of baseline serum creatinine concentration, development of end-stage renal disease or death from any cause) during a follow up period of 2.6 years by 20% when compared to the placebo (Lewis et al., 2001). Nevertheless, 50% of patients in the Irbesartan group reached the primary endpoint after 54 months. In an effort to increase the efficacy of RAS antagonistic therapy an angiotensin receptor blocker was combined with placebo or the angiotensin converting enzyme (ACE) inhibitor Lisinopril (Fried et al., 2013). The combination therapy did not reduce the incidence of a combined renal endpoint. On the contrary an increased risk of hyperkalemia and acute kidney injury was observed confirming other reports questioning the safety of this approach (Mann et al., 2008; Parving et al., 2008).

Next to addressing RAS, organ-specific molecular processes involving inflammation and oxidative stress have been implicated in progressive tubulointerstitial fibrosis, the best histological, hence molecular mechanistic predictor of an adverse renal disease prognosis (Rodríguez-Iturbe and García García, 2010). Bardoxolone, a nuclear factor-erythroid-2-related factor 2 activator with anti-oxidative capacity increased eGFR in patients with advanced diabetic renal disease (Pergola et al., 2011). However, a large prospective controlled randomized trial with hard endpoints had to be stopped because of severe side effects (De Zeeuw et al., 2013).

As given with these examples for chronic kidney disease (but in its conceptual fundament holding true for a multitude of highly prevalent chronic diseases), many of the recent interventional studies failed to achieve their goals. Here biomarkers promise to take a key role in selecting patients for studies and/or to predict the long term effects of a drug on hard endpoints. Upfront stratification in randomized controlled trials by separating patients by drug response as measured by biomarkers serving as endpoint surrogate and then randomizing the groups separately is an approach which is, at least from a statistical point of view, preferable to post-hoc analysis (De Leon, 2012). Such an enrichment strategy is currently e.g., tested in the SONAR study (clinicaltrials.gov reference NCT01858532) addressing diabetic nephropathy (DN).

However, with respect to fit of specific drugs biomarkers need to carry predictive value, i.e., a biomarker shall on a patient-specific level identify responders benefitting from drug effect. In this setting various levels need to be considered involving genetic and environmental components defining disease presentation and progression. The drug target may see genetic polymorphism impacting drug binding, but polymorphism may further involve drug transport and drug metabolism (Johnson, 2001). A significant number of genetic polymorphisms have in the meantime become drug label-relevant regarding drug efficacy, but also toxicity and side effects (U.S. Food and Drug Administration, 2014). Pharmacogenomics has clearly demonstrated that the genetic background of an individual introduces heterogeneity in drug response.

Still, this setting assumes a homogeneous patient population with respect to the molecular mechanistic factors determining disease progression, only exhibiting differences in genetic peculiarities of one and the same molecular mechanistic context. In such setting functional biomarkers appear sufficient for identifying progressive disease, and drug variance is fully explained by the genetic background in regard to the mechanism of action of a specific drug.

A complementary perspective may be that the molecular mechanistic background and progression-relevant molecular factors are per se diverse and patient-specific, naturally determining drug response (Mayer et al., 2012). In such scenario a biomarker needs to serve as proxy of key mechanistic factors characterizing and driving a disease on a patient-specific level, combined with educating on the specific interference of disease mechanism with drug mechanism of action. For capturing these constraints a detailed molecular map of a clinical phenotype and its interference with a drug mechanism of action is needed, and here integration of Omics profiling adds to identifying such mechanisms (Fechete et al., 2011; Mühlberger et al., 2012).

An a priori stratification of patients based on an appropriately chosen biomarker panel reflecting the pathophysiology of a given patient (group) allowing to determine a match with a specific drug's mechanism of action appears as promising approach. As recently discussed by Himmelfarb et al. fresh approaches are critical in finding therapies to kidney disease benefiting patients, outlining the importance of improving the translational aspect in clinical research (Himmelfarb and Tuttle, 2013). Here, omics technologies have added significantly to the data landscape characterizing chronic kidney disease, however, in a first instance mainly expanding the candidate set of apparently relevant processes and pathways, going in hand with a large number of biomarker candidates, which individually hamper clinically relevant assessment on disease progression (Fechete et al., 2011; Hellemons et al., 2012).

Integrative approaches in the realm of Systems Biology have been proposed for reaching a consensus description of chronic kidney disease pathophysiology, including molecular models of DN as well as of the reno-cardial axis (He et al., 2012; Komorowsky et al., 2012; Mayer et al., 2012; Heinzel et al., 2013). Still, a translation process needs to be followed, joining disease pathophysiology, stratification markers allowing enrichment strategies, combined with on a molecular mechanistic level matching drugs for allowing precision medicine (Mirnezami et al., 2012). In this work we exemplify such procedure on DN being the major clinical presentation leading to end stage renal disease.

For a large spectrum of clinical presentations an impressive number of drug targets have been proposed out of translational and preclinical research, with a significant number further proceeding into clinical trials. Just in the first half of 2014 close to 10,000 new clinical studies were recorded on the platform clinicaltrials.gov. Taking a specific look at diabetic nephropathy as clinical phenotype, 124 interventional trials in any status are identified at clinicaltrials.gov specifically involving the disease term, covering 45 individual drug entities addressing 86 known targets. Via mining scientific literature additional 18 drug targets are identified.

Next to a number of trials utilizing drugs and drug combinations addressing known factors impacting DN progression as the RAS, drug targets are disparately distributed across molecular pathways, hence mechanisms assigned to the disease.

From literature mining 27 different pathways according to KEGG and Panther pathway annotation are discussed as associated with DN, of which 19 hold drug targets. These include well known mechanisms of relevance in DN including hypoxia response or fibrosis, combined with a large set of signaling components. On top, 52 drug targets are embedded in molecular context outside this literature-derived DN pathway landscape.

For biomarker candidates an equivalent situation is found. 54 unique proteins extracted from scientific literature are discussed in any biomarker context, covering 14 of the 27 pathways, with 31 biomarker candidates not assigned to any of the members of the extracted DN pathway map.

Interestingly, predictive performance regarding disease progression of any of the individual biomarker candidates proved limited value. For example, in a review by Hellemons et al. 13 relevant markers were found in the context of nephropathy in diabetes, of which five were found as significantly associated with onset as well as progression of DN again covering various mechanisms including inflammation (e.g., C-reactive protein), cell surface interaction and homeostasis (e.g., E-selectin, ICAM1) and metabolism (triglyceride levels) (Hellemons et al., 2012).

Apparently, individual biomarkers reflecting the status of an individual molecular process, pathway or mechanism cannot capture disease prognosis for the comprehensive DN population. In alternative approaches multimarker panels were included in classifiers on disease diagnosis and prognosis demonstrating improved performance also in blinded validation. In Roscioni et al. a signature of 273 peptides determined in urine were included in a support vector machine-based classifier (Roscioni et al., 2013). The signature held fragments of collagen eventually mirroring alterations in the extracellular matrix turnover and fibrosis together with markers of inflammation as e.g., the pro-inflammatory protein S100-A9, as well as uromodulin shown to be associated with interstitial fibrosis and tubular atrophy (Nkuipou-Kenfack et al., 2014).

One contributing factor for needing multimarker panels may be individual variance of baseline biomarker levels, where inclusion of multiple markers specifically in non-linear classification methods adds to robustness. However, a second factor may be generic heterogeneity of the patient population. Specific disease presentation may significantly vary not only across stages of disease progression eventually seeing a transition from protective to damaging mechanisms, but even within a specific chronic kidney disease category as defined by present clinical classification provided by KDIGO guidelines (KDIGO Board Members, 2013).

Improved prognostic performance of multimarker panels on top of strict functional classification of stage transitions in DN utilizing albuminuria but also eGFR as clinically used progression parameters clearly support the case of pathophysiological heterogeneity of a, in present clinical terms homogeneous, patient population. However, specifically for albuminuria the role of functional marker vs. factor in disease is discussed (Roscioni et al., 2014).

Deriving robust diagnostic or prognostic classifiers from e.g., proteomics or metabolomics profiling may add to clinical patient management regarding onset as well as intensity of therapeutic measures (Roscioni et al., 2013; Pena et al., 2014). Also in clinical trial design such enrichment strategies may be utilized by e.g., identifying individuals prone to fast disease progression, and randomizing in this high risk cohort into medication and placebo arm (e.g., Priority trial, clinicaltrials.gov reference NCT02040441).

Prognostic biomarkers in contrast to diagnostic parameters with known assignment to molecular processes and pathways further allow an approximation of what specific mechanisms are associated with disease progression. The DN pathway landscape discussed in this work is solely a cross-sectional representation of the disease, in a first place not allowing deciphering which of the 27 individual pathways drive disease progression, and which other pathways are just bystanders or downstream consequences of mechanistic factors of disease. Hence, evaluating biomarker candidates for their association with progressive disease in turn allows determining mechanisms associated with progressive disease. Such knowledge is vital e.g., for determining novel drug targets, demanding to be embedded in disease mechanisms being factors for progressive disease. Remaining question however is if such mechanisms are relevant to the same extent or at all for a specific patient assigned to a clinical phenotype.

A prognostic biomarker set covering all potentially relevant processes enables specific molecular phenotyping of individual patients, being however not sufficient in terms of predicting drug response as a drug mechanism of action is not factored in. Here Systems Pharmacology aims at identifying drug response also on the level of molecular processes and pathways. Rationale is to not only focus on the specific drug target and its assignment to specific mechanisms, but to include the systemic molecular changes triggered by the drug including off-target effects as well as downstream molecular changes. Having a drug mechanism of action as well as a clinical phenotype represented on a molecular process or pathway level allows intersecting both molecular states. If from prognostic biomarker profiling of a patient specific progression-associated molecular disease mechanisms are identified, and a drug exhibits functional interference in such specific mechanisms such patient may be more prone for showing response to the drug. With such setting including knowledge on molecular phenotype composition, molecular process relevance in progressive disease and knowledge on interference of drug mechanism of action biomarker candidates initially serving a prognostic purpose can be rendered into predictive biomarkers on drug response.

Omics profiling has a major contribution to characterizing both, clinical phenotypes as well as drug mechanism of action. Integrating profiling results from clinical samples frequently sees minor overlap of individual studies, being in part driven by insufficient sample size combined with diverging inclusion criteria and sample material used (Fechete et al., 2011). In the example presented here 1010 features in total are identified as differentially regulated in transcriptomics or are being assigned to DN according to literature mining, with 880 unique features. An equivalent misbalance in feature coherence across studies is also found for the ACE inhibitor transcriptomics data. All these drugs address the same functional context, but from the in total 3152 features identified for six drugs included the total number of unique features are still 2058, with 661 being identified in at least two drug signatures.

Next divergence becoming apparent is the limited overlap of enrichment analysis based on signatures from profiling and feature-based literature mining compared to explicit literature mining for molecular pathways. Of the 27 pathways extracted from scientific references only seven are confirmed, however, seeing other pathways enriched not found via literature mining. On top, a major shortcoming is restricted representation of protein coding genes in such pathway maps, e.g., for KEGG covering 6491 and for Panther 2163 protein coding genes, respectively. This limitation not only affects pathway enrichment but also assignment of biomarker and target candidates. Of the in total 104 drug target and 54 biomarker candidates 29 are neither assigned in any KEGG or Panther pathway.

Here a different approach may be followed, namely segmentation of protein interaction networks exhibiting improved coverage of the protein coding gene set. Consolidation of INTACT, Reactome, and BioGRID allows representation of in total 13,907 protein coding genes, clearly expanding beyond public domain pathway databases. In alternative approaches hybrid interaction networks are utilized for further expanding coverage of protein coding genes, but also for improving false negative rates regarding protein-protein interactions and relations (Fechete et al., 2013).

Computing a DN-specific as well as ACE inhibitor-specific induced subgraph followed by topology-based segmentation allows an alternative representation of a molecular process landscape for the clinical presentation as well as the drug mechanism of action. Interference analysis on the level of overlapping protein coding genes resulted in four process segments holding central aspects of DN pathophysiology. Seven biomarker candidates were identified in these interfering molecular processes. CCL5 (RANTES), involved in recruiting monocytes and macrophages to the renal cortex was shown to be suppressed by ACE inhibition, indicating that RANTES expression is mediated via Angiotensin II type 2 receptor (Kashiwagi et al., 2002). Equivalently, in animal models TGFB1 expression was shown to be reduced by ACE inhibitors. Activation of NFKB1 by angiotensin II was shown in vascular smooth muscle and mesangial cells (Hernández-Presa et al., 1997). In a study by Dong et al. analyzing cost effectiveness of ACE inhibitor treatment for patients with type 1 diabetes mellitus the level of glycosylated HbA1c showed clear impact on cost effectiveness of drug use per quality-adjusted life year (QALY) (Dong et al., 2004). The authors concluded that next to patient age also other factors need to be included in therapy considerations.

Apparently, drug mechanism of action affects numerous molecular processes, as exemplified for ACE inhibitors, many of these also afflicted with DN progression. Analyzing the molecular process interface of disease progression-relevant pathophysiology and drug mechanism of action allows proposing predictive markers. Testing such predictive biomarker candidates may educate on relevance of individual processes on a patient level, directly linking to likelihood of drug response.