This study was approved by the Research Ethics Board of the Pitié-Salpétrière Hospital, Paris—France and is registered under the Clinical trial number NTC02627404.

The sample comprised 64 individuals who met the study eligibility criteria. The main inclusion criteria were a diagnosis of BD type I according to DSM-IV criteria (4th edition of the Diagnostic and Statistical Manual of Mental Disorders), currently in remission (i.e. no major episode or hospitalization in the last six months) and euthymic (i.e., a score of < 8 on both the Montgomery-Åsberg Depression Rating Scale (Montgomery and Åsberg 1979) and the Mania Rating Scale (Bech et al. 1978) and being willing and able to give written informed consent. Individuals with a current diagnosis of comorbid alcohol or cannabis abuse or dependency were excluded.

Eligible participants were interviewed with the French version of the Diagnostic Interview for Genetic Studies (DIGS) to confirm the diagnosis of BD and to assess lifetime comorbidities (such as alcohol and cannabis use) and with the Family Interview for Genetic Studies (FIGS) to assess family history (Nurnberger et al. 1994). All participants were currently on a stable prophylaxis medication (i.e., no change in the last six months) and are completely independent from those previously published (Marie-Claire et al. 2022, 2020).

Response to Li was rated using the “Retrospective Criteria of Long-Term Treatment Response in Research Subjects with Bipolar Disorder”, also referred to as the “Alda scale” (Grof et al. 2002). This scale was specifically developed to allow a retrospective assessment of prophylactic response to treatment in naturalistic conditions (and considers both overall response to the introduction of Li on a 0–10 scale, whilst accounting for confounders of response such as poor adherence, etc.). In accordance with the available literature (Manchia et al. 2013), patients with total score >  = 7 were characterized as good responders (GR) and patients with total score ≤ 3 were characterized as non-responders (NR). Remaining individuals were classified as partial responders (PaR).

DNA was isolated from total peripheral blood collected at inclusion. Genomic DNA was extracted using the automated Maxwell 16 DNA Purification Instrument and the dedicated RSC Blood DNA Kit (Promega). All DNA samples were stored at − 80 °C at the “Plateforme de Ressources Biologiques des Hôpitaux Universitaires Henri Mondor in Creteil France”. As previously described, genomic DNA input was 200 ng to be modified with sodium bisulfite using the EZ DNA™ methylation kit (Zymo Research, Irvine, CA, USA). Human methylated and unmethylated DNA standards were diluted before bisulfite conversion to prepare 0, 10, 20, 30, 40, 50, 60, 70, 80, 90 and 100% methylated to unmethylated template ratios. Bisulfite modified DNA was eluted in 10 µL of nuclease-free water according to the manufacturer’s instructions. The modified DNA was quantified with a NanoDrop One Spectrophotometer (Ozyme, Saint-Cyr-l’École, France).

Methylation-Sensitive High-Resolution Melting (MS-HRM) is based on the comparison of the melting profiles of PCR products from unknown samples with profiles specific for PCR products derived from methylated and unmethylated control DNAs (Šestáková et al. 2019). As previously described, MS-HRM tests have been designed to amplify three previously identified DMRs in GR as compared to NR individuals with BD-I. Primers design with the Bisearch online tool (http://bisearch.enzim.hu/) and amplification details for DMR17107 (amplicon size 169 bp), DMR106540 (amplicon size 116 bp) and DMR24332 (amplicon size 147 bp) have been previously described (Marie-Claire et al. 2022). PCR reactions were performed on a CFX384 Touch Real-Time PCR Detection System (Biorad Laboratories, Des Plaines, IL, USA) in a final volume of 10 µL, containing 200 nM of each primer, 5 µL of Precision Melt Supermix (Biorad Laboratories) and 10 ng of bisulfite treated DNA. The initial denaturation (95°, 3 min) was followed by 45 cycles of 10 s at 95 °C, 30 s at 50 °C, 30 s at 72 °C. The HRM step consisted of a denaturation of all products at 95 °C for 30 s followed by an annealing at 60 °C for 1 min. Samples were then slowly warmed to 95 °C at 0.2 °C per second, holding for 10 s after each stepwise increment and fluorescence data were collected. Each sample was analyzed in triplicate. For the DNA methylation assessment, bisulfite converted dilutions of methylated and unmethylated DNA standards were analyzed together with the samples. Peak-heights were calculated automatically with the CFX Maestro Software (Version 2.2; Bio-Rad Laboratories, Inc., Hercules, CA, USA). Linear curves of the peak-heights of the Tm first derivative of HRM curves against the methylation percentage of the standard were plotted (Tse et al. 2011). The lab technician was blinded to response status of the participants.

All statistical analyses were performed with IBM SPSS version 28 software. A p-value < 0.05 was considered statistically significant.

We performed backward stepwise logistic regression analyses (BSLR), followed by receiver operating characteristic (ROC) curve analysis with findings reported as the estimated area under the curve (AUC) with 95% confidence intervals (95% CI). The following available variables were included in model 1 of the BSLR (socio-demographic and clinical variables alone): age, sex, cigarette smoking status, lifetime number of hospitalizations, age at onset of BD, polarity at onset, psychotic symptoms at onset, family history of BD, lifetime alcohol misuse, lifetime cannabis misuse, panic disorders, Li prescribed as the first mood stabilizer (vs 2nd or 3rd choice, etc.). The BSLR analysis was then repeated, but the model included sociodemographic and clinical variables alongside the three DMRs (model 2). The following variables were not used into the models since redundant with some B items of the Alda scale: number of episodes (B1 item) and frequency of episodes, including rapid cycling (B2 item). Since there are different ways to use the Alda scale as a complement of the classical three groups (i.e. using A-scale, or A-scale with B score < 4 for instance), we also provide an analysis based on model 2, but using the A-scale only in a linear regression. Multicollinearity was examined using the Variance Inflation Factors (VIF) (a VIF above 4 indicates that multicollinearity might exist).