Adverse drug reaction are divided into types A and B ADR (Figure 1). Type A ADR, the so-called “pharmacological ADR,” are caused (i) by a change of dosage and/or pharmacokinetics and consequently of its pharmaco- or toxicodynamic action or (ii) solely by a change in the target structure leading to different affinity of the drug to the target and/or a different agonist-directed trafficking at the (target-)receptor. In contrast, type B ADR, drug hypersensitivity ADR, are caused by allergic or non-allergic mechanisms involving the immune system and/or mediators such as histamine (Figure 2). Type A were estimated to account for approximately 80% of ADR occurring in clinical practice (Borda et al., 1968). However, this figure has undoubtedly changed over the last 50 years due to differences in drug prescriptions, pharmacovigilance activities and a better understanding and thus demarcation of type B ADR. 34 years later, maybe owing to these advances in medicine, type A were reported to account for 91% of all ADR (Mjorndal et al., 2002).

In the past, it was postulated that type A ADR are usually a feature of the drug property and thus predictable, while type B ADR are strongly dependent on the genetic features of the host. Pharmacogenetic polymorphisms are now known to aggravate certain type A ADR (cf. descriptions of AERD and red-man-syndrome below). Type B ADR appeared to be non-predictable and dose-independent. However, dose-dependency has been shown for some hypersensitivity reactions (Rive et al., 2013). Rising knowledge of genetic polymorphisms of the immune system have helped to predict at least some type B ADR by applying genotyping (Rive et al., 2013).

Depending on the mechanism of activation of the immune systems, most type B ADR (~75%) can be classified as either non-allergic hypersensitivity ADR (formerly called “pseudoallergic”), i.e., direct effect on mast cells causing histamine release, or as type I reaction according to Gell and Coombs, i.e., IgE-mediated histamine release. Type IV reactions, i.e., T-cell-mediated delayed hypersensitivity reactions, are less common (~25%). Types II and III reactions are uncommon among drug hypersensitivity reactions.

Besides these immune reactions types I to IV, a direct pharmacologic action on immune receptors (“p-i concept”) of T-cells without prior presentation of the drug via MHCs (coded by HLA) has been proposed (Posadas and Pichler, 2007). Finally, some drugs are thought to alter the functioning of the immune system, e.g., alpha-methyldopa can induce the production of drug-independent autoimmune antibodies (Pierce and Nester, 2011), and statins potentiate the shifting of T-helper 1 to T-helper 2 immune responses (Suchak et al., 2007).

While type A ADR are usually a feature of the drug, drug hypersensitivity reactions are strongly dependent on the genetic features of the host. However, pharmacogenetic markers have been described for both types of ADR.

Several databases exist nowadays to summarize the various findings (Table 1).

Drugs can elicit hypersensitivity ADR by several non-allergic and allergic mechanisms at once, e.g., penicillines can induce immediate and delayed ADR (Blanca et al., 2009), and vancomycin can lead to both IgE-mediated and non-allergic hypersensitivity ADR (Polk et al., 1993).

To date, the different type IV reactions are not easily separated, i.e., type IVc reactions appear to be a dominant mechanism that is also occurring during types IVa, IVb, or IVd reactions (Posadas and Pichler, 2007).

Depending on the HLA-genotype, the risk agranulocytosis mediated by the antipsychotic clozapine could be significantly increased (Goldstein et al., 2014). However, a second mechanism was proposed before hypothesizing that clozapin would be metabolized to highly reactive nitrenium ions that deplete the ATP and glutathione-content of neutrophils and would ultimately lead to neutropenia (Tesfa et al., 2009).

The general state of the immune system may influence the occurrence or non-occurrence of ADR. E.g., vancomycin-related red-man-syndrome is very frequently seen (up to 90%) in non-infected, healthy patients, e.g., those receiving prophylactic treatment, but much less (3.7–47%) in those suffering from infections (Sahai et al., 1990; Sivagnanam and Deleu, 2003). Vice versa, reactions to ampicilline (notably, not amoxicilline) are much more common in EBV-infected patients (Haverkos et al., 1991; Chew and Goenka, 2016) and reactions to sulfonamides occur predominately in HIV-infected patients (Jaffe et al., 1983). The mechanism for sulfonamide-induced hypersensitivity in HIV-patients was attributed to a decrease of cellular glutathione (Rieder et al., 1995).

It can be speculated whether the clinical manifestation of a drug hypersensitivity reactions is connected to specific genetic alterations. E.g., is there a common pool of polymorphisms leading to either DILI, DIA or SCAR? Existing data do unfortunately deny the existence of such a simple genetic constellation.

Flucloxacilline and amoxicilline/clavulanic acid are powerful first-line antibiotics which may cause DILI (Daly et al., 2009). Due to their wide-spread usage and inherent ability to cause DILI, it is vital to further understand the mechanism. Unfortunately for testing purposes, the numbers needed to screen are high and routine testing is (currently) not feasible. This in part due to the rareness of DILI which is estimated to occur in about 14 of 100,000 inhabitants per year (Bell and Chalasani, 2009). This comes unexpected when taking into account that the liver is a metabolically highly active organ, able to produce a variety of bioactivated and thus putatively immunogenic compounds from the parent drug. However, the liver is considered a tolerogenic environment, i.e., there are mostly no or locally restricted immune reactions in the liver (Crispe, 2003).

Amoxicilline/clavulanate has been linked to DILI in patients with HLA-DRB1*15:01 and recently with HLA-A*30:02, HLA-B*18:01 and the complex genotype HLA-DRB1*15:01-DQB1*06:02 (Stephens et al., 2013). Interestingly, the detected populations at risk also differed in age and speed of reaction onset.

Several drugs can elicit a life-threatening destruction of blood cells. E.g., the antipsychotic clozapine is commonly (0.8% of drug users) causing neutropenia (less than 1500 neutrophil granulocytes/μl) or agranulocytosis (less than 500 granulocytes/μl). Other triggering drugs for DIA comprise thyreostatics like methimazole and the analgesic metamizole (USAN: dipyrone). Pharmacogenetic associations are known for methimazole and HLA-DRB1*08:03:2 (Tamai et al., 1996), for metamizole and various HLA genotypes (Vlahov et al., 1996) and for clozapine and a complex HLA genotype in Jewish subjects (Lieberman et al., 1990) for more than two decades.

Since clozapine appears to be the most effective antipsychotic drug available, there is a need to increase our understanding of its safe or putatively unsafe usage. The initial findings for clozapine were subsequently refined and applied to the general population (Dettling et al., 2001). Various polymorphisms leading to DIA after clozapine exposure have been identified, e.g., a complex genotype consisting of HLA-DRB5-DRB4, HLA-C-B-DRB5 (Dettling et al., 2007), HLA-DQB1 (Athanasiou et al., 2011), HLA-DQB1 (126Q) and HLA-B (158T) (Goldstein et al., 2014). The latter two studies deserve additional attention since the association was identified to be due to single amino acid changes at a defined position in the HLA gene product rather than to alleles described before. This is similar to the incidental finding that a single amino acid change in HLA-A, HLA-B, and HLA-C at position 152 might explain altered susceptibility of T-cells to drugs and lead to DILI (Stephens et al., 2013). Amino acids exchanges at this position alter the antigen binding pocket E of the MHC I receptor and possibly the interaction with T-cells. These findings suggest that a single amino acid change and not a serology-derived typing of HLA may provide a better prediction of the observed hypersensitivity reaction because of the mechanistical explanation inferred by the change in the binding pocket.

Next to the liver, the skin, including the mucosa, is involved in bioactivation of drugs. Additionally, since the skin is the barrier that protects our body from the environment, it is rich of immune cells. Due to this combination of a huge amount of putatively immunogenic compounds, an abundancy of immune cells and the constant pre-sensitization of the dermal immune cells due to their contact with pathogens, the skin is a prime location for the manifestation of immune reactions. It has been proposed that the increased reactivity of EBV- or HIV-infected patients to aminopenicillines might be due to a lower activation threshold of T-cells (Posadas and Pichler, 2007). Differential pathways of activation have been show for flucloxacillin (Yaseen et al., 2015).

Severe cutaneous ADR encompass DRESS and SJS/TEN. Commonly used substances which can cause SCAR are abacavir, lamotrigine, and carbamazepine (Pirmohamed et al., 2015). The latter has an aromatic moiety and is grouped together with phenobarbital and phenytoin to the aromatic anticonvulsants. These aromatic anticonvulsants were associated with various HLA-A and HLAB variants. Interestingly, a hypersensitivity reaction to the non-aromatic drug lamotrigine was recently also associated to HLA-B*15:02 and SJS in a Han Chinese population (Man et al., 2007; Zeng et al., 2015).

The finding that abacavir is very likely to induce DRESS in HLA-B*57:01 positive patients led to a guideline of the CPIC which recommends testing because of the severity of the reaction and the high risk in absence of genetic prescreening (Martin et al., 2012). It is estimated that 6% in the general population are carriers of HLA-B*57:01 and that around 50% of carrier will develop DRESS. The recommendation is thus classified as “strong.” However, it was argued that 50% of positively tested individuals will be denied an effective treatment option (Martin et al., 2012). Therefore, CPIC reviews guidelines on a regular basis. Recently, the unchanged recommendation was confirmed (Martin et al., 2014).

Table 2 summarizes all affected drugs and clinical manifestations of late-onset hypersensitivity reactions.

As nicely illustrated by the abacavir example, any improvement of signal detection and risk calculations requires correct assessment of the event (Phillips et al., 2011). Physicians in clinics tend sometimes to misdiagnose the reaction, focus on only one aspect, fail to conduct further investigations that would strengthen or invalidate the finding, or fail to document the event appropriately (Thien, 2006). The solution that worked for the abacavir-findings was to conduct two tests, both a clinical assessment using a structured query form and a skin patch test to confirm involvement of the immune system (Phillips et al., 2011).

Starting with the thalidomide tragedy, the powerful pharmacovigilance system was installed in most countries. The strength of this system is the huge data pool on adverse events like ADR encompassing millions of cases world-wide. A major drawback is the paucity of information routinely entered into this systems (Böhm et al., 2016). Genetic information is usually missing and the patients mentioned in the individual safety reports cannot be recontacted for further investigations. It is thus only possible to generate hypotheses which need further evaluation.

To increase the collection of cases and their quality, interested parties have formed their own collaborative research teams, e.g., RegiSCAR for severe cutaneous reactions (Phillips et al., 2011), the Berlin Case–Control Surveillance Study group (e.g., Huber et al., 2015) and the International CIA Consortium for DIA or DILIGEN, iDILIC, DILIN, and others focusing on DILI (Nicoletti et al., 2016). These networks allow standardized collection of patient and event data as well as genetic material, and possibly re-identification of patients for further data or material sampling.

Of note, such a network requires tremendous resources if the drug-event-combination is rare. E.g., finding the association of flupirtine and HLA-DRB1*16:01-DQB1*05:02 causing DILI took more than 10 years of careful preparation and data collection (Nicoletti et al., 2016). The Berlin Case-Control Surveillance Study group collected data for 10 years (Huber et al., 2015). The increasing availability of electronic health records, the ongoing deployment of biobanks and the advances regarding the accompanying ethical, legal, technical, and social challenges (Strech et al., 2016) might deliver a global new data source for research on drug hypersensitivity reactions.

All data essentially boils down to a 2x2 contingency table of one biomarker and the occurrence of the ADR. Several statistical measurements can be derived from this: p-values (e.g., derived from Chi squared with Yates’ correction), odds ratios, sensitivity/specificity, PPV, NPV as well as NNS.

It is sometimes desirable to combine several polymorphisms to a complex genotype or, analogously, several drugs to one drug class or several reactions to a syndrome. Grouping of individual findings will enhance the usefulness of the data for constructing novel detection techniques and to reach statistical and clinical significance (Böhm et al., 2016). E.g., genotyping of individual polymorphisms in cases of clozapine-induced agranulocytosis revealed no findings, whereas grouping polymorphisms to complex genotypes revealed three associations (Dettling et al., 2007). Another example is the 75% cross-reactivity of carbamazepine, phenytoine and phenobarbital and possibly lamotrigine on SJS/TEN. Arene oxide is a common moiety of metabolites of the first three anticonvulsant drugs and is considered to be the immunologic active substance. Grouping these drugs might help to strengthen signals for certain polymorphisms.

Traditional statistical approaches are not suited for analyses of small numbers. The fewer cases are reported, the wider the 95%-confidence interval becomes. The work of iDILIC shows that it is possible to find a convincing signal with just six cases (Nicoletti et al., 2016). However, in the future Bayesian approaches are expected to replace traditional frequentists’ methods for such small numbers of cases (Yum et al., 2014). Bayesian statistics is increasingly used in early clinical trials in order to efficiently screen for signals where numbers are low.

Each of the statistical measurements given above should be used for different purposes:

Table 3 shows several examples where statistical, technical, clinical and economic significance differ.

The example of the pooled data (dataset #3) as extracted from the HLADR database shows that combining data does not substantially improve PPV, NPV, or NNS if the individual datasets are of sufficient quality (Mallal et al., 2002; Martin et al., 2004).

A common problem is the acquisition of data due to the low numbers of cases and matching controls: often, not all cells of the 2x2 contingency table contain a count. Zero counts are problematic for the calculation of measurements of disproportionality. There are several ways to cope with this situation: A mathematical tool is the usage of Haldane’s modification for Odds Ratios. By adding 0.5 to every cell, zero counts are eliminated. Another option is carefully matching other (known) cohorts to the existing data (Nicoletti et al., 2016). The theoretical gold standard is to keep gathering reports until all cells can be filled in with sufficiently large counts which is often not feasible.

The full dataset (#4) using complex genotypes illustrates that the combination of genotypes enhances the performance of the test, in this case the PPV (Mallal et al., 2002).

The dataset #5 illustrates that even test with an extremely low sensitivity can lead to acceptable high PPV and NPV and thus NNS (Nicoletti et al., 2016).

The findings in dataset #6 shows a scenario in which testing appears very reasonable if calculating PPV and NPV from the original data (PPV = 91.49%, NPV = 88.23%, NNS = 1.3). However, when applying the real population-wide prevalence instead of relying solely on the counts reported in this publication, a much lower PPD of 0.12% results and thus a very high NNS of approximately 13,000 (Daly et al., 2009). Still, testing is desirable from the patient’s point of view (NPV 100%). However, it is economically not feasible.