Four cases with AMX-induced SCAR were recruited from Huashan Hospital of Fudan University in Shanghai, China from January 2006 to December 2022. Phenotypes of SCAR were classified according to the consensus definitions of RegiSCAR study criteria (Auquier-Dunant et al., 2002; Cacoub et al., 2011; Wang et al., 2018). The drug causality for each enrolled case was determined by the dermatologist using the algorithm of drug causality for epidermal necrolysis (ALDEN) score published by the RegiSCAR study group or the Naranjo algorithm (Naranjo et al., 1981; Sassolas et al., 2010; Kardaun et al., 2013). Data on patients’ demography, clinical patterns, causative drugs, treatments, and clinical outcomes were collected from patients’ case notes. Patients’ files with incomplete data were excluded. For general population and drug-tolerant controls, we collected clinical data and DNA samples of 1,000 individuals in general population without history of adverse drug reactions as the WES control group as well as 100 drug-tolerant patients who had received amoxicillin for several times and the cumulative duration was for more than 3 months without evidence of adverse reactions were enrolled as tolerant controls (Supplementary Table S1). The latency period between the intake of the drug and the onset of the skin eruption was recorded. Written informed consent has been collected from all subjects, and the study protocol was approved by the institutional review board of Department of Dermatology, Huashan Hospital.

The whole-exome sequencing (WES) method include steps such as DNA sample extraction, library building, and sequencing, following the methods reported previously by (Jiang et al., 2022). Typing of HLA alleles were performed using HLA-HD version 1.3.0 (Kawaguchi et al., 2017) with sequence mapped to the reference sequence of six-digit HLA alleles from the IMGT/HLA database in default settings (Barker et al., 2023). High-resolution four-digit HLA-B allele genotyping was performed by Weihe Biotechnology (Jiangsu, China) using the polymerase chain reaction-sequence based typing (PCR-SBT) method.

We conducted molecular docking studies via the Maestro (version 2019) software to investigate the interaction patterns between AMX and HLA protein. The protein and nucleotide sequences of HLA-B*15:01, HLA-B*46:01, HLA-B*15:02, and HLA-B*35:05 were obtained from the IMGT/HLA database. The structure of these four HLA protein molecules were obtained from SWISS-MODEL server, and HLA-B*15:01, HLA-B*15:02, HLA-B*35:05 and HLA-B*46:01 were modeled based on the following crystal structures, respectively: 5TXS (100% identity), 6UZM (100% identity), 7SIF (100% identity) and 4LCY (100% identity). AMX molecule was used as ligands and HLA molecule was used as receptors for docking modeling. The receptor molecule obtained the dominant conformation after repairing the missing residues and minimizing the energy using Maestro software. Ligprep module was used to generate 32 3D conformations of AMX molecules, and the optimal conformation was selected for modeling. The above receptor and ligand molecules were protonated at human physiological PH value. All parameters were set to the software default parameters. The docking simulation of the optimized ligand receptor was performed, and the docking models of the top 20 outputs were displayed. According to the docking scores and results, the optimal binding mode of AMX docking with each HLA molecule was selected.

The statistical analyses were carried out using R version 4.2.3 (The R Foundation for Statistical Computing, Vienna, Austria). Individuals who tested positive for at least one copy of a specific allele were considered as carriers. Fisher exact test was applied to compare the carrier frequency between specific HLA allele for AMX-SCAR patients and population controls as well as AMX-tolerant controls. Overall odd ratios (ORs) with corresponding 95% confidence intervals (CIs) were calculated to quantify the association. When zero cell counts were included, ORs were calculated using Haldane’s modification, which adds 0.5 to all cells to accommodate possible zero counts to reduce bias in estimating ORs. In the HLA protein sequence analysis, the χ² test and Fisher’s exact test were used to analyse the polymorphic position and amino acid residue between the AMX-SCAR group and two control groups, respectively. A two-tailed p-value of <0.05 was considered as statistically significant.

Herein, we described four patients with AMX-SCAR recruited from the dermatology ward from Huashan Hospital of Fudan University between 2010 and 2021, who have been diagnosed with SJS, TEN, DRESS and Acute Erythroderma, respectively (Table 1). One out of four patients underwent oral AMX monotherapy, and the remaining three were on oral AMX combined with other drugs, for which we have ruled out other possible combined drugs as the culprit based on the medical history and ALDEN score or Naranjo algorithm. Latency period ranged from 1 to 9 days, with an average of 6.5 ± 2.6 days. Among them, three out of four patients had fever (≥37.5°C), 2 cases presented increased white blood cell count (≥0.5×10⁹/L), 1 case exhibited drug induced liver injury, and 2 cases had a history of penicillin allergy implying cross-reactivity with other penicillins.

The HLA analysis detected three carriers of HLA-B*15:01 in four patients with AMX-SCAR, and the carrier frequency of HLA-B*15:01 is 10.7% in 1,000 population controls and 11.0% in 100 AMX-tolerant controls, implying a significant association of HLA-B with AMX-SCAR (Supplementary Table S2, Table 2). Previous studies suggested that amino acid motifs rather than classical alleles were the major drives conferring disease risk (Illing et al., 2012; Raychaudhuri et al., 2012). In the present study, three out of four patients harbored HLA-B*15:01, suggesting it may play an important role in the pathogenesis of AMX-SCAR (Table 1). Then, we further analyzed the frequencies of amino acids in HLA-B proteins in cases and 100 AMX-tolerant controls, detecting 28 amino acids significantly enriched in patients (p < 0.05). Among them, S140 (p = 5.18 × 10⁻⁴) demonstrated the strongest association, followed by A48 (p = 4.45 × 10⁻³), L187 (p = 4.45 × 10⁻³), R121 (p = 6.84 × 10⁻³), W14 (p = 6.84 × 10⁻³), V17 (p = 6.84 × 10⁻³), and W180 (p = 7.34 × 10⁻³) (Supplementary Table S3).

Given the strong associations of AMX-SCAR with specific amino acids of HLA-B, molecular modeling was conducted to further confirm the interaction between AMX and HLA-B proteins to eliminate false-positive results. The docking results revealed that amino acid at position 140 and 180 in HLA-B play a crucial role in binding to AMX. In HLA-B*15:01, S140, K170, W171, and E176 can form hydrogen bonds, R121 and W180 formed cation-π hydrophobic interactions with AMX. These combined effects gave rise to relatively low steric hindrance, which was conducive for AMX binding in the pocket with a low-energy conformation (Figure 1A). In HLA-B*15:02 (belongs to the same serotype and differs from HLA-B*15:01 by only 5 amino acids), W180 became L180, which greatly increased steric hindrance (Figure 1B). In HLA-B*35:05 (presents in patient 4), although the conversion of W180 to L180 slightly weakens the binding to the AMX molecule, S140 remains to be a key amino acid in forming hydrogen bonds. In addition, W171 and N104 also form hydrogen bonds, and this binding mode allows AMX to be surrounded by polar amino acids, promoting the formation of stable complexes in the dynamic binding process (Figure 1C). In HLA-B*46:01 (shows a high frequency in tolerant controls), AMX’s phenolic head is surrounded by polar amino acids such as E87, and the increased steric hindrance prevents AMX from remaining in a low-energy conformation in the pocket. Consequently, AMX loses the hydrophobic effect with W180 and the polar effect with S140 (Figure 1D). Overall, AMX tends to bind with S140 in the antigen-binding cleft of HLA-B, differing from unrelated HLA molecules HLA-B*46:01, which indicates that predisposition to AMX-SCAR is related to one essential amino acid residue in the peptide binding pocket of HLA-B.

For HLA-B proteins, amino acid residues S140, L187, A48, R121, W14, V17, I306, T329, and W180 are encoded by coding sequence variants C419, CT559-560, G142, G363, G41, G49, A916, A985, and T538, respectively. The distribution of these genetic variants in the AMX-SCAR and AMX-tolerant groups is consistent with amino acids. From the perspective of genetic variants as predictors of AMX-SCAR, where C419, CT559-560, G142, G363, G41, G49, A916, and A985 showed the highest sensitivity (up to 100%), followed by T538 with 75%. In terms of specificity, variant T538 exhibited the highest value (up to 89%), followed by C419 (87%), CT559-560 and G142 (76%), G363, G41, and G49 (73%), and A916 and A985 (72%). According to the current results, S140 of HLA-B or its genetic allele C419 (also known as rs4997052), present in all 4 subjects with AMX-SCAR but only 13 of the 100 AMX-tolerant controls, can serve as a clinical predictor of AMX-induced SCAR with a sensitivity of 100% and specificity of 87% (Table 3).