Archival serum samples were from the Child Health and Development Studies (CHDS, Public Health Institute, Oakland, California). The CHDS recruited women residing in the Oakland, California area who were members of the Kaiser Foundation Health Plan and received obstetric care for pregnancies between 1959 and 1967 [27]. Breast cancer cases were identified by linkage to the California Cancer Registry and the California Vital Status Records for cases diagnosed through 1997 as previously described [28]. Record abstraction is based primarily on pathology reports, and case identification is considered to be >99% complete after a 2-year lag for cancer diagnoses [28]. Cases were defined as women with incident invasive or noninvasive breast cancer diagnosis with available prenatal serum. The median age at diagnosis was 54 years (interquartile range, 13 years). Third trimester samples of 172 women who subsequently developed breast cancer were analyzed along with samples from 351 women from the cohort who did not develop breast cancer [29]. Of women with breast cancer, 34 had cancer diagnosis within 15 years of pregnancy. From the age distributions, approximately half of these women had breast cancer diagnosis before age 50, the usual age used as a cutoff for early-onset breast cancer. The institutional review board of the Public Health Institute approved the present study (IRB# I20–002), and we complied with all federal guidelines governing the use of human participants. Blood samples were collected with informed consent from all subjects. This was an observational study and had no experiments on humans.

Serum samples were randomized into batches of 20 samples for extraction and analysis using liquid chromatography-high resolution mass spectrometry (LC-MS) as described previously [30, 31]. HPLC grade acetonitrile and methanol, LC-MS water and 98% formic acid were obtained from Sigma-Aldrich (St. Louis, MO). A mixture of 14 stable isotopic chemicals were used as internal standards [32], including [¹³C₆]-D-glucose, [¹⁵N]-indole, [2-¹⁵N]-L-lysine dihydrochloride, [¹³C₅]-L-glutamic acid, [¹³C⁷]-benzoic acid, [3,4-¹³C₂]-cholesterol, [¹⁵N]-L-tyrosine, [trimethyl-¹³C₃]-caffeine, [¹⁵N₂]-uracil, [3,3-¹³C₂]-cystine, [1,2-¹³C₂]-palmitic acid, [¹⁵N,¹³C₅]-L-methionine, [¹⁵N]-choline chloride and 2’-deoxyguanosine-¹⁵N₂,¹³C₁₀-5’-monophosphate from Cambridge Isotope Laboratories, Inc (Andover, PA).

For processing, 50 μL serum sample was treated 2:1 (v/v) with acetonitrile, and 2.5 µL internal standard of a mixture of the 14 stable isotope standards. Following mixing, extracts were incubated at 4°C for 30 min, and protein precipitates were removed by centrifugation for 10 min at 21,000 x g at 4°C. Supernatants were placed in autosampler vials and maintained at 4°C in an autosampler prior to analysis. A pooled human reference material [Qstd, pooled plasma from 2 lots (Equitech-Bio, Inc, Kerrville, Texas)] was analyzed at the beginning and end of each batch of 20 samples. A second pooled human reference material (NIST SRM1950) was analyzed at the beginning and end of the study. Samples were analyzed in triplicate with Q-Exactive HF mass spectrometer (Thermo Fisher) using two complementary analytic platforms: reversed phase C18 chromatography and negative electrospray ionization (ESI) to measure more lipophilic compounds and hydrophilic interaction chromatography (HILIC) with positive ESI to measure more polar metabolites. Data collection occurred continuously throughout 5 min of chromatographic separation from m/z 85 to 1,275.

Peak detection, noise removal, and alignment of retention times and features were performed using the adaptive processing for LC-MS data (apLCMS) v.6.6.8 R package [29–31, 33].Additional filtering of low-quality samples and features and batch correction were conducted with xMSanalyzer v.1.3.2 [34]. Mass spectral features for each sample consisted of mass to charge (m/z), retention time (rt), and associated integrated ion intensity; interpretations are limited to features with annotation or identification as metabolites and are therefore termed metabolites. Only samples with Pearson correlation ρ >0.7 among 3 replicates were retained. Metabolic features were filtered to retain those with median coefficient of variation among technical replicates >70%. Intensities below the detection threshold were not retained by the extraction procedures, and detected intensities were median summarized across triplicates for statistical analyses. The final data included 9,042 metabolic features from C18 chromatography and 10,324 from HILIC.

Identities of amino acids and other selected metabolites were confirmed by accurate m/z match, MS/MS spectra, and co-elution with authentic standards, which is Level 1 identification by criteria of Schymanski et al] [35]. Other features were annotated with a multistage clustering algorithm, xMSannotator [36] using the Human Metabolome Database [37] at 5 ppm tolerance. Annotations by xMSannotator in combination with Mummichog pathway enrichment analyses are considered equivalent to Level 3 identification criteria of Schymanski et al. [35].

To test for amino acid pathway associations with subsequent breast cancer diagnosis, we used log2 transformed feature intensities to decrease heteroscedasticity and a two-step metabolome-wide association study (MWAS) design. This approach has been previously used to balance Type 1 (false positives) and Type 2 (false negatives) statistical error [38] and was considered important in the present study due to the expectation that the population may not be in an active disease state at time of blood collection, and effect sizes may be small. The first step in the MWAS was to use the ANOVA-like program Limma to select metabolites differing at raw p<0.05 (protecting against Type 2 error) [39]. This was performed using an in-house xmsPANDA R package with the R v.4.0.3. The Benjamini-Hochberg procedure was used to test for false discovery rate (FDR)[40]. Selected features at raw p <0.05 were used for the second step in which pathway enrichment analysis identified metabolic pathways associated with breast cancer. This analysis uses permutation testing and protects against Type 1 statistical error in identification of pathways associated with breast cancer outcome. In these analyses, metabolites for both columns were analyzed together using Mummichog version 2 (MetaboAnalyst, www.metaboanalyst.ca) [41]. Partial Least Squares-Discriminant Analysis (PLS-DA) was used to identify features that best distinguished the control and breast cancer groups. Variable Importance in Projection (VIP) values from the PLS-DA were used as selection criteria for network analysis.

To investigate relationships of amino acids with other metabolites associated with breast cancer risk, we used a data-dependent community detection tool, xMWAS [39], with all metabolites selected at p<0.05 and VIP>1.5. This tool enables visualization of networks of communities, which are groupings of network members based on a distance metric, for which we used PLS correlation with a threshold of r >0.4. xMWAS includes an Eigenvector Centrality Score (ECS) which provides a quantitative measure of centrality in network structures. For this analysis, in addition to breast cancer status (BC) and age, we included breast cancer diagnosis within 15 y (BC15) and other factors previously associated with breast cancer risk, i.e., primiparous status, >3 live births, and log-transformed estrogen concentrations for E1, E2, E3, and the ratio of E3/(E1+E2). We also plotted Pearson correlation matrices of amino acids with hierarchical clustering using the heatmap library in R [31]; to simplify the interpretation of the heat map, amino acid adducts were only included if they possessed at least one correlation with another network node where |r| > 0.4.

Characteristics of the subjects studied are summarized in Table 1 (detailed data are available in [26]). Briefly, breast cancer cases were approximately 4 years older than non-cases, and women who were diagnosed with breast cancer within 15 years of blood sampling were about 9 years older than non-cases and 5 years older than all cases. Non-cases had differences in parity compared to cases and cases diagnosed within 15 years of sampling. Most of the women were white, with 16 to 21% African American and <10% Hispanic, Asian or Mixed race.

Limma showed that 275 metabolites in C18 analyses and 261 metabolites in HILIC analyses differed at raw p<0.05 between women who subsequently were diagnosed with breast cancer and those without known breast cancer (Fig 1A; Supplement 1). For both analyses, more metabolites were negatively associated with breast cancer than were positively associated. No metabolites were significant after FDR correction, indicating that none is likely to be useful individually as a biomarker of risk.

Pathway enrichment analysis of the metabolites that differed between women with breast cancer diagnosis and women without breast cancer diagnosis showed statistically significant enrichment in nine amino acid pathways (tyrosine; histidine; arginine and proline; alanine and aspartate; aspartate and asparagine; lysine; methionine and cysteine; valine, leucine and isoleucine; and tryptophan) and the urea cycle (Fig. 1B) along with other pathways (Supplement 2). Identified amino acid and amino acid metabolites included proline, histidine, urocanate, lysine, threonine, phenylalanine, phenylpyruvate, 5-oxoproline (pyroglutamate). Additional supportive evidence for amino acid pathway identifications was obtained from correlation analyses (Supplement 3). Although no individual amino acids could be confidently associated with breast cancer risk after FDR adjustment, the results showed that multiple amino acid pathways are associated with risk.

We used PLS-DA to test whether metabolites that differed at raw p<0.05 separated individuals with subsequent breast cancer (BC) from those without known breast cancer (NonBC). Partial but incomplete separation was observed for both C18 analyses (Fig. 1C) and HILIC (not shown). Loadings from the PLS-DA were used to select metabolites contributing to separation, and 136 metabolites from C18 data had VIP>1 (Fig. 1D, Supplement 4). Of these, 33 were positively and 103 negatively associated with breast cancer. Analyses for HILIC showed that 128 metabolites contributed to separation, 33 positively and 95 negatively (Supplement 4). We used VIP, which measures the importance a given metabolite in the PLS-DA model, as a selection filter. The top 32 metabolites with VIP>1.5 included threonine, histidine and amino acid metabolites, pyroglutamate and phenylpyruvate (Supplement 5). These 32 metabolites with greatest contribution to separation had relatively low fold-changes (0.78 to 1.28) for breast cancer relative to non-breast cancer. Thus, from this approach examining thousands of mass spectral signals, the results show that amino acid pathways are associated with breast cancer risk and, although no individual amino acid metabolite appears to have useful prognostic value, amino acids are among the metabolites that contribute most to metabolic differences between women subsequently diagnosed with breast cancer and those who were not.

Visualization of amino acid-associated networks showed three communities (Fig 2), with a simplified version provided in Fig 2A to highlight the most the dominant characteristic of pregnancy serum amino acids associated with breast cancer outcome. Four amino acids (threonine, histidine, lysine, proline) are negatively associated (red edges) with pregnancy-related measures [>3 live births, primiparous status, E3/(E1+E2)] previously found to be protective against breast cancer. Notably, age did not have associations that were strong enough to be included in this network. Thus, this network analysis shows that amino acids are among the most central metabolic features in the organizational structure linked to subsequent breast cancer diagnosis, and these are negatively associated with protective phenotypic measures.

In the complete community structure (Fig 2B), an unidentified metabolite in the largest community, Community 1, was most central (Fig 2B). The accurate mass m/z 210.0499 of this metabolite matches the M+2Na⁺-H⁺ adduct of phenylalanine (Eigenvector Centrality Score, ECS = 1.0); however, in correlation analyses (Supplement 3), this metabolite did not correlate with other confirmed forms of phenylalanine. Additionally, this metabolite was unlike other amino acids in having positive correlations with breast cancer and associated phenotypic markers; therefore, despite the accurate mass match to a common adduct form of phenylalanine, this metabolite is considered unidentified.

Of the phenotypic measures in Community 1, >3 live births had the highest Eigenvector Centrality Score (0.667) followed by primiparous status (0.559) (Fig 2B). Community 1 also included BC15, with a relatively high Eigenvector Centrality Score (0.324) despite only having correlations at this threshold with histidine, threonine and the unidentified central metabolite (m/z 210.0499).

The next largest community, Community 2, contained lysine as a hub connected to 5-oxoproline, and three unidentified metabolites, one of which could be an adduct form of histidine (Fig 2B). Lysine was connected to multiple nodes in Community 1 and to the phenotypic measure, E3/(E1+E2), in Community 3 (Fig 2B). Community 3 also included proline, with edges connected to nodes in Community 1 (Fig 2B). Collectively, the evidence provided by this network analysis shows that risks of breast cancer and breast cancer within 15 years of pregnancy are negatively associated with third trimester serum amino acids. The amino acids are similarly negatively associated with pregnancy measures previously found to be protective against breast cancer risk. Notably, at the correlation threshold used, age was not present in any association, and breast cancer diagnosis within 15 years was selectively associated with threonine and histidine.

Because proline, threonine, and histidine have previously been associated with mechanisms relevant to oncogenesis, we quantified amino acids present in the central communities in Fig. 2, including threonine, histidine, proline and lysine, to allow direct comparisons of differences between individuals who subsequently developed breast cancer and those who did not. As is evident for these amino acids (Fig. 3), even though the amino acids were lower in breast cancer at p < 0.05, the populations had substantial overlap so that individual measures are not likely to be useful as biomarkers of risk.

A relatively high Eigenvector Centrality Score for BC15 in the network analysis (Fig 2) suggested important metabolic relationships for individuals with breast cancer diagnosis within 15 years of sample collection. The top associations with threonine, histidine and the unidentified m/z 210.0499, occurred without BC15 having similar associations with primiparous status, >3 live births, or E3/E1+E2. These findings indicate that BC15 may have different metabolic antecedents than breast cancer diagnosed after 15 y. To investigate this, we repeated statistical analysis using individuals with breast cancer within 15 years of pregnancy as our target and the remaining subjects (breast cancer after 15 y and non-cases) for reference. Limma was used to select discriminatory metabolites at p <0.05 (Supplement 6), and these were used with PLS-DA Score Plots to visualize separation of BC15 from non-breast cancer and other BC cases. Results using these selected metabolites showed that overall group-wise separation was superior for BC15 compared to non-breast cancer (Fig. 4A) compared to all breast cancer compared to non-breast cancer (Fig 1C). Results also showed relatively good separation for BC15 compared to BC cases diagnosed after 15 years (BCother, Fig. 4B). Pathway enrichment analysis of the selected metabolites showed a smaller number of pathways (Fig. 4C), including lysine, histidine, tryptophan and methionine and cysteine. Consequently, the data show that increased risk of breast cancer within 15 y is associated with a modest but important interplay of multiple amino acid pathways in late pregnancy.