Participants

A sample of 401 medically well White European adolescents aged 12 to 16 years at psychosocial risk for unipolar major depression (DSM–IV) ²¹ were recruited from 11 secondary schools in the Cambridgeshire region of the UK. The study was approved by the local ethics committee and written informed consent for participation from the adolescent and one parent or guardian was obtained. Participants were recontacted 12 months later and 365 (91%) agreed to a mental state re-evaluation. There were no major differences (P>0.1) between the retained sample and those who were not followed up in terms of demographics (age, gender), risk status, morning cortisol levels, and mean level of depressive symptoms or distribution of the BDNF and 5-HTTLPR alleles.

Psychosocial risk profile

Psychosocial risk was defined as having a parent with a psychiatric history (216, 54%), or two or more other risks consisting of: two bereavements over the young person's lifetime (192, 48%), a negative emotional temperament (133, 33%), chronic (>6 months) marital disharmony or parental separation (186, 46%), two recent undesirable life events or difficulties with family or friendships over the past 12 months (121, 30%). We made no attempt to stratify the sample by number of risks or type, but retained the risk score (number). The sample is not representative of the Cambridgeshire school population. As a condition of ethical approval those with a recent (past 2 years) history of overt maltreatment or who were on the UK child protection register were also excluded.

Psychopathology measures

At entry the adolescents completed a semi-structured psychiatric interview, the Schedule for Affective Disorders and Schizophrenia for School-Age Children – Present and Lifetime Version (K–SADS–PL). ^{Reference Kaufman, Birmaher, Brent, Rao, Flynn and Moreci22} Those with current or recent (past 3 months) depressive disorder were excluded from the study to restrict outcomes to new depressive episodes only. Individuals with three or four depressive symptoms but who did not meet clinical criteria for unipolar depression (n = 18) were retained. Individuals were assessed for all disruptive behavioural and anxiety disorders and retained. Participants were reassessed at 12 months using the same measures and all episodes of DSM–IV-defined major depression and probable major depression over the follow-up period (four symptoms together with psychosocial impairment defined as a Children's Global Assessment Score of <60) ^{Reference Shaffer, Gould, Brasio, Ambrosini, Fisher and Bird23} were established together with new and persistent episodes of anxiety and behavioural disorders. All participants were administered a 33-item self-report measure of current depressive symptoms, the Mood and Feelings Questionnaire (MFQ), ^{Reference Kent, Vostanis and Feehan24} at entry: 385 (96%) provided useable data. This questionnaire has established criterion-related validity for use in adolescents with depression. ^{Reference Daviss, Birmaher, Melhem, Axelson, Michaels and Brent25} The internal consistency reliability estimate in this sample was high (Cronbach α = 0.96) and the marginal reliability under item-response theory (IRT) exceeded 0.9. The sum score correlated >0.95 with IRT-based estimates.

Recent life events

At follow-up, a life-events interview was completed by the adolescents. Recent events were included for analysis if they were rated by the adolescents as having a moderate to severely undesirable impact on themselves lasting at least 2 weeks and occurred in the month immediately prior to onset as these are associated most with depression onset in young people. ^{Reference Goodyer, Herbert, Tamplin and Altham7} Those deemed possibly not severe, low in impairment or likely to have been brought about by illness-related behaviours were excluded (<5%). A panel of raters (n = 4) checked all life-events ratings and where ratings were ambiguous a consensus rating was made. Panel raters were independent of the interviewer who collected the information on events and did not know whether the adolescent had received a diagnosis at the time of their rating. Test–retest (2 weeks, n = 40) reliability for all events (κ > 0.85) and agreement with parent report (κ = 0.9) were satisfactory. Research interviewers also showed good general agreement with each other (90%) and with adolescents' subjective ratings of level of undesirability (κ = 0.8). This semi-structured procedure evaluating the quality and number of such events is considered a more appropriate procedure for genetically sensitive studies than using simple, self-completion checklists. ^{Reference Monroe and Reid18}

Cortisol assay

Participants provided samples of saliva at 08.00 h and recorded their waking time on 4 consecutive schooldays within a week of the interview (99% 2 days, 98% 3 days, 95% 4 days). Mean time between waking and morning salivary cortisol collection was 50 min. Cortisol was measured by ELISA on 20 μl samples of saliva without extraction (antibody Cambio UK). Intra-assay variation was 5.7%; inter-assay variation was 5.6%. Results are reported in ng/ml.

Genotyping

Their DNA was harvested from a separate saliva sample (Qiagen, Crawley, UK) also collected at entry and genotyped for 5-HTTLPR and BDNF. The 5-HTTLPR region was genotyped as described elsewhere. ^{Reference Goodyer, Bacon, Ban, Croudace and Herbert20} The frequency of the common Val66Met (rs6265) BDNF polymorphism was determined commercially (Geneservice, Cambridge UK, www.geneservice.co.uk) using a custom-designed assay (Applied Biosystems, www.appliedbiosystems.com).

Statistical analysis

We determined whether individual differences in morning salivary cortisol levels were influenced by genetic variants in BDNF and 5-HTTLPR using mixed-effects linear regression computing the intraclass correlation using STATA version 10 for Windows. This is reported in the online Table DS1. We predicted depression onsets using a multiple logistic regression model, ^{Reference Harrell26} implemented in R for Windows. ^{Reference Adachi, Barrot, Autry, Theobald and Monteggia27} The binary outcome was the onset of a DSM–IV major depressive episode at any point over the follow-up period of 1 year. Morning salivary cortisol level, MFQ score and total number of risks at study entry were entered as continuous variables. The presence or absence of one or more undesirable life event in the month prior to new episode onset was included as a categorical variable. Our model included main effects for all variables, including BDNF and 5-HTTLPR, together with two-way interactions between BDNF and cortisol and between 5-HTTLPR and cortisol. We did not include the two-way interaction between BDNF and 5-HTTLPR as we were primarily interested in their interactions with cortisol; however, we note that such an interaction, if present, would bias our estimates of the main effects of the two polymorphisms and their interactions with cortisol. We considered a possible three-way interaction between BDNF, 5-HTTLPR and cortisol but did not test this as the overall model fit was inferior as measured by Akaike's information criterion (AIC). The online tables (DS2–4) report the robustness of the final model to further adjustment/confounding by three variables: a grouping variable for those (n = 18) with subclinical depressive symptoms at study entry, and participant gender and age at entry.

We performed a post hoc power estimation for the logistic regression analysis using a bootstrap procedure. We repeated the procedure 1000 times and estimated the proportion of significant tests of the interaction terms. This approximates the power of our study taking the effect sizes to be their sample estimates. Results from traditional logistic regression are reported as unadjusted (univariate) and adjusted (multiple logistic regression) odds ratios with 95% confidence intervals. P-values for unadjusted estimates are reported using asymptotic values. Two analysts (T.C. and F.D.) independently replicated the general linear model analyses using both STATA version 10 and R for Windows.

Gene variants and morning cortisol

The distribution of the BDNF gene variant across the whole sample was: Val66Val n = 268 (67%), Val66Met n = 109 (27%), Met66Met n = 24 (6%) and for 5-HTTLPR l/l n = 155 (39%), l/s n = 166 (41%), s/s n = 80 (20%). These two genes are somewhat associated in this sample (exact test, P = 0.052) but as they are unlinked, this association may be a result of a selection effect in our cohort of at-risk adolescents, although this requires further confirmation. Mixed-effects linear regression showed no main effect of BDNF on morning salivary cortisol but the known main effect for 5-HTTLPR ‘s’ allele on morning cortisol was retained ^{Reference Goodyer, Bacon, Ban, Croudace and Herbert20} (see online Table DS1). For morning salivary cortisol the within- and between-participant variance components were 60% and 40% respectively. The model was not improved by inclusion of two-way interactions so we did not test the interaction of BDNF and 5-HTTLPR on morning salivary cortisol.

Predicting subsequent depressive episode onset

The differences in factors at entry together with undesirable life events over the follow-up period between those who developed a new depressive episode and those who did not are documented in Table 2. At 12 months 175 (48%) of the 365 reassessed participants reported one or more undesirable life events prior to onset of episode or month of interview in controls. Of these, 70 (19%) reported a permanent loss, 94 (26%) a personal disappointment, 17 (5%) a personal danger, 41 (11%) and 46 (13%) a danger to a meaningful other (family member or personal friend). Mental state reassessment showed that 41 (11%) had experienced an episode of depression between entry and follow-up. There were no significant differences in the proportion of participants reporting undesirable life events or depression onset by BDNF or 5-HTTLPR genotype. Individuals were more likely to report a depressive episode if they were female, had self-reported higher depression scores at entry, had higher morning waking salivary cortisol at entry and reported experiencing one or more undesirable life events.

Table 2 Factors at entry, subsequent life events and depression onset

	Depressed episode during follow-up period		Statistics
	No (n = 324)	Yes (n = 41)	F	d.f.	P
Measures at entry
Depressive symptoms, mean (s.d.)	17.6(9.0)	26.5(9.9)	33.9	1,358	<0.0001
Morning salivary Cortisol, mean (s.d.)	3.0(1.3)	3.8(1.6)	11.13	1,364	0.0009
Age, years: mean (s.d.)	13.6(1.2)	13.6(1.1)	0.00	1,364	0.96
Psychosocial risks, mean (s.d.)	2.2(0.7)	2.3(0.6)			0.23a
Gender, n (%)					0.032a
Females	147(45)	26(63)
Males	177 (55)	15 (37)
BDNF, n (%)					0.92a
Val66Val	214(66)	27(66)
Val66Met	90(28)	11(27)
Met66Met	20(6)	3(7)
5-HTTLPR, n (%)					0.82a
‘l/l’	119 (37)	17 (41.5)
‘l/s’	140 (43)	17 (41.5)
‘s/s’	65 (20)	7 (17)
Measured at follow-up
Undesirable life events, n (%)					<0.0001a
None	182(56)	8(20)
One or more	142(44)	33(80)

We analysed the risk of depression using multiple logistic regression with 357 (89%) participants who had full data. For the purposes of analysis both gene variants are collapsed into two groups each as follows: BDNF as ValBDNF (Val66Val, n = 268 (67%)) and MetBDNF (Val66Met + Met66Met, n = 133 (33%)); 5-HTTLPR as l/l-5-HTTLPR (l/l variant n = 155 (39%)) and s-5-HTTLPR (l/s + s/s variants n = 246 (61%)). The findings are shown in Table 3. There are significant effects for higher antecedent depressive symptoms, higher morning salivary cortisol and subsequent undesirable life events. The effect of higher morning salivary cortisol is qualified by the two two-way interactions with Val66Val variant of BDNF and the ‘s’ variant of 5-HTTLPR genes respectively. The latter results implicate both gene variants moderating the risk of morning waking salivary cortisol on subsequent depression onset. There were no effects of age, gender or minor depression on these results (online Tables DS2–4). The power to detect significant associations at the 5% level was 61% for the interaction between BDNF and cortisol on risk of depression, and 44% for the interaction between 5-HTTLPR and cortisol. The power for main effects of BDNF and 5-HTTLPR was 64% and 57% respectively, although these were not our main interest. The probability of a new depressive episode occurring in the presence of morning waking cortisol by each gene is shown in Fig.1.

Table 3 Factors that predict onset of a new depressive episode over 1 year^a

Factor	No major depressive disorder (n = 317) v. major depressive disorder (n = 40)
	Coefficient	s.e.	Odds ratio	95% CI	Wald χ2	d.f.	P
BDNF	2.828	1.23	16.9	1.5–188.4	5.73	2	0.057
5-HTTLPR	–3.004	1.29	0.1	1.6–252.6	5.86	2	0.053
Depressive symptoms	2.061	0.48	7.9	3.1–20.1	18.17	1	<0.0001
Morning Cortisol	1.519	0.85	4.6	0.9–24.2	17.37	3	<0.001
Life events	1.265	0.47	3.5	1.4–8.9	7.29	1	<0.001
BDNF × morning cortisol	–2.337	0.98	0.1	0.14–0.67	5.72	1	0.017
5-HTTLPR × morning cortisol	2.028	1.02	7.6	1.02–56.1	3.96	1	0.046
Multiple degree of freedom test for interactions					9.37	2	0.009
Overall model					37.63	7	< 0.0001

This shows line plots of the probability of subsequent depression for morning waking salivary cortisol × 5-HTTLPR (l/l v. l/s + s/s) and BDNF (Val66Val v. any Met variant) respectively from the model in Table 3. The figure shows the moderating effects of each gene variant on the risk that elevated morning cortisol carries for the subsequent onset of a depression episode over the 12-month follow-up period. In each panel, the predicted risk from the logistic model described in the method above was calculated for each participant and the logit-transformed risks then regressed back onto cortisol alone to give a smoothed marginal curve for our fitted model.

Fig. 1 Probability of subsequent major depression by either (a) BDNF (brain-derived neurotrophic factor gene) Val66Met or (b) 5-HTTLPR (the serotonin transporter gene promoter) and morning cortisol levels.

Genetic influences on vulnerability for depression onset

Since more than 80% of depressive episodes arise in adolescents with prior psychosocial risks and elevated depressive symptoms, ^{Reference Jaffee, Moffitt, Caspi, Fombonne, Poulton and Martin28} the current findings are likely to represent antecedent patterns of vulnerability for depression onsets and pathophysiological mechanisms underpinning psychosocial-risk processes in a considerable proportion of the adolescent population in general. The findings also show that the risk effects of genetically sensitive elevations in cortisol are not confounds of the depressive symptoms at entry or proximal undesirable life events during follow-up. The absence of gene × social environment interactions may be because this is a moderate-risk sample with no low-risk participants, a function of selecting undesirable events close to onset of disorder rather than cumulative indices of social adversities over time and/or not including those difficulties arising from earlier in the life course. ^{Reference Brown and Harris29} Alternatively it may be because the mediating psychological processes such as individual differences in emotion processing and behavioural styles associated with event perception and occurrence have not been included in this study. ^{Reference Jacobs, Kenis, Peeters, Derom, Vlietinck and Van Os30,Reference Canli, Ferri and Duman31}

Possible influences of Val66Met BDNF on depression

As noted elevated glucocorticoid levels reduce BDNF mRNA expression in the hippocampus ^{Reference Kumamaru, Numakawa, Adachi, Yagasaki, Izumi and Niyaz8} and interfere with the ability of BDNF or serotonin-acting drugs to promote neurogenesis. ^{Reference Huang and Herbert9,Reference Pinnock and Herbert10} We also noted a preliminary report suggesting that adults who carry the less common Met allele may be protected at the level of the brain from the deleterious effects associated with the short ‘s’ allele of 5-HTTLPR. ^{Reference Pezawas, Meyer-Lindenberg, Goldman, Verchinski, Chen and Kolachana14} From the current results we speculate that when adolescent carriers of the common Val66Val BDNF variant are exposed to subsequent undesirable life events the risk for corticoid-mediated neurotoxicity may be increased because of the less protective nature of this BDNF gene variant in the face of stress. If true this would result in a cascade of cellular events inhibiting synaptic plasticity and neurogenesis and diminishing the liability for adaptive response behaviour. This speculative mechanism is hypothesis generating and awaits further study preferably with a sample size that can test for a three-way relationship between BDNF Val66Met, ‘s’ allele 5-HTTLPR and morning cortisol.

Implications of genetic influences on corticoid-mediated risk

Overall the findings suggest that the two common genetic variants we studied are indexing functional interrelations between neurobiologically active systems and help to further specify morning waking salivary cortisol levels, a physiological endophenotype, as a biomarker for unipolar depressions emerging during adolescence.

Limitations

The modest sample size, data loss to follow-up and the power limitations in testing for moderator interactions limit any conclusions about causality for or specificity between our findings and affective disorder. Furthermore the selection of just two gene variants prevents a comprehensive interpretation of genetic effects on the social environment or the overall risk for subsequent affective disorders. Given the caveats surrounding the interpretation of gene–social environment interactions ^{Reference Kumamaru, Numakawa, Adachi, Yagasaki, Izumi and Niyaz8,Reference Risch, Herrell, Lehner, Liang, Eaves and Hoh17,Reference Munafo, Durrant, Lewis and Flint32} and the limitations represented by selection of single candidate gene variants in complex systems, ^{Reference Petryshen, Sabeti, Aldinger, Fry, Fan and Schaffner33} these findings require replication in other populations to understand the impact of stratification due to gender, psychosocial and other risk factors.

Further research

A prospective developmental life-course cohort study with a larger sample would yield a greater proportion of conversions to depressive status of differing forms (sporadic, recurrent and chronic unipolar, bipolar I or II and rapid cycling types) over a longer time period and provide the necessary replication or refutation. Future studies also need to include measures of lifetime childhood difficulties and life-course histories of mental illness as there may be programming effects on individual differences in morning cortisol arising from the environment perhaps in genetically sensitive individuals. ^{Reference Ouellet-Morin, Boivin, Dionne, Lupien, Arseneault and Barr34} Measuring psychological processes may further characterise the mediating pathways between genes, cortisol and the social environment. For example, both 5-HTTLPR and BDNF gene variants may differentially influence reward processing and motivational incentives as well as response to negative mood ^{Reference Forbes15,Reference Cools, Roberts and Robbins35} but the relationships of these processes to morning cortisol levels and hence to subsequent depression remains unclear.