Evidence for Maternal Autoantibodies in the Pathogenesis of Preterm Birth

431 432 Certain pathogenic alloantibodies are known to drive pregnancy complications in humans6,12,13, 433 motivating a systematic examination of maternal autoreactivity during gestation. By profiling more 434 than two thousand maternal sera across eight pregnancy cohorts, we defined a comprehensive 435 landscape of humoral autoreactivity in term and preterm pregnancies. Overall autoreactivity increased 436 with advancing gestational age and gravidity, underscoring the dynamic and evolving nature of the 437 maternal immune repertoire. After adjusting for gestational age at sampling, we identified a preterm 438 birth-associated autoantibody signature that was consistently enriched across geographically and 439 ethnically diverse cohorts. Notably, these reactivities were detected an average of fifteen weeks 440 before preterm delivery, indicating that serologic changes precede clinical onset rather than result 441 from it. Because preterm birth arises from diverse pathophysiological causes, including inflammatory, 442 vascular, and idiopathic etiologies, the associated autoantibody landscape is correspondingly 443 heterogeneous, likely contributing to the large feature set required for classification. The 444 autoreactivies that drive the classification model may be mechanistically pathogenic and/or reflect 445 broader immune dysregulation. 446 447 448 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint One limitation of this work is that the model does not reveal the underlying triggers or events that lead 449 to defects in self-tolerance associated with preterm birth. However, within 450 this broader signature, multiple proteins comprise plausible autoreactive targets of functional 451 significance. Among these, a rare subset of preterm pregnancies harbored functionally inhibitory 452 antibodies against the IL-1 receptor antagonist (IL1RA) that exacerbated IL-1β–driven reproductive 453 defects in mice and were capable of binding IL1RA in human placenta. Their presence may reflect 454 post-infectious or autoimmune immune priming, as anti-IL1RA antibodies have been described after 455 COVID infection52, and in IgG451 and Still’s disease53. Together, these findings support a two-hit 456 model, in which pre-existing or infection-induced dysregulation of IL-1β signaling is amplified by anti-457 IL1RA antibodies, tipping the balance toward inflammation, placental injury, and pregnancy loss. 458 Although the pathogenic role of these antibodies requires further investigation, recombinant IL1RA 459 (anakinra; SOBI, Inc.) has been used clinically for over two decades with established long-term safety 460 in non-pregnant populations56,57 and limited case reports of use in pregnancy58. These data suggest 461 that anakinra use in pregnancy to prevent preterm labor merits further evaluation, particularly in 462 stratified cohorts defined by anti-IL1RA antibody status. 463 464 Beyond an individual protein, this collective proteome-wide data, representing thousands of 465 individuals, are a rich compendium of human humoral autoreactivity in normal and preterm 466 pregnancy, making it ideal for emerging machine learning approaches, especially with respect to 467 classification tasks. As human antibody and antigenic repertoire techniques continue to also evolve, it 468 is likely that clinically useful descriptors of immune dysregulation will also emerge, as will targeted 469 therapeutic interventions to limit the effects of autoantibodies, which in turn may yield improved 470 management of pregnancies at risk for preterm delivery and reductions in neonatal morbidity and 471 mortality. 472 473 474

475 476 Human cohorts and samples 477 Sera from the Cohort I was obtained from a California-wide biobank (Committee for the Protection of 478 Human Subjects within the Health and Human Services Agency of the State of California protocol# 479 12-09-0702). Sera from the Cohort II, Cohort III, Cohort VI, and Cohort VII (UCSF IRB# 10-00505, 20-480 31171, 20-32077, 20-32779, 16-20474; San Francisco, CA) was obtained from deliveries at UCSF. 481 Sera from Cohort IV (UCSF IRB# 10-00350) was obtained from UCL Hospital and Homerton 482 University Hospital of the National Health Services of the United Kingdom. Sera from Cohort V was 483 obtained from ‘Fondazione IRCCS Policlinico San Matteo’ (San Matteo Research Hospital) in Pavia, 484 Italy, under RC08061819 and RC08061821 approved IRB protocols. Sera from Cohort VIII (IRB# 485 2009P000557 and 2014P001109, Boston, MA) was obtained from Brigham and Women’s Hospital in 486 Boston, MA, USA from the previously described VDAART clinical trial59. De-identified healthy control 487 (non-pregnant) plasma was collected from two sources: courtesy of New York Blood Center (New 488 York, NY) and donors from a UCSF community drive (UCSF IRB# 22-3611; San Francisco, CA). All 489 samples were collected under the referenced Institutional Review Board protocols, and all patients 490 gave informed consent prior to sample collection. 491 492 Uncomplicated, term pregnancies were selected from Cohort I to define a baseline of autoreactivity. 493 An uncomplicated, term pregnancy for this cohort was defined as one resulting in delivery at greater 494 than or equal to 37 weeks of gestation and did not have a diagnosis of any of the following: placental 495 abruption, placenta previa, chorioamnionitis, oligohydraminos, polyhydraminos, premature rupture of 496 membranes, bacterial vaginosis, urinary tract infection, diabetes or gestational diabetes, preexisting 497 or gestational hypertension, preeclampsia, infection during pregnancy, any specified placental 498 condition, retained placenta with or without hemorrhage, unspecified hemorrhage, abnormal clotting, 499 or threatened abortion. Any pregnancy that required transfusion, rhesus isoimmunization, or cerclage 500 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint was excluded from this group. We also excluded mothers with any hypertension disorder, 501 endometriosis, Group B streptococcus positivity, any coagulation deficiency, malignancy, asthma, 502 allergic dermatitis, anaphylaxis, rheumatoid arthritis, systemic lupus erythematosus, autoimmune 503 thyroiditis, or mental disorder. Mothers who reported any smoking, drug or alcohol abuse or 504 dependency during pregnancy were excluded. Pregnancies where the infant exhibited convulsions, 505 abnormal neural imaging or exam, retinopathy of prematurity, respiratory distress syndrome, 506 intraventricular hemorrhage, necrotizing enterocolitis, bronchopulmonary dysplasia, periventricular 507 leukomalacia, hypoxic-ischemic encephalopathy, or known congenital heart disease were also 508 excluded from this group. Finally, we also required that women defined as having uncomplicated 509 pregnancies did not have a history of recurrent pregnancy loss, preterm delivery, or poor 510 obstetric/reproductive outcomes. 511 512 Term pregnancies for all cohorts were defined as deliveries at 37 or greater weeks of gestation. 513 514 Preterm pregnancies were defined as those delivering prior to 37 weeks of gestation and were further 515 subdivided into spontaneous or iatrogenic preterm delivery. Spontaneous preterm pregnancies were 516 defined as pregnancies with spontaneous labor with delivery prior to 37 weeks of gestation. Iatrogenic 517 preterm pregnancies were defined as pregnancies with delivery prior to 37 weeks of gestation and 518 without spontaneous labor. 519 520 PhIP-seq with human peptidome library 521 The human T7 phage display library used for immunoprecipitation and sequencing is previously 522 described and sera from the above cohorts was used in high-thruput protocols as previously 523 described. Detailed protocols are published on protocols.io DOI: 524 dx.doi.org/10.17504/protocols.io.4r3l229qxl1y/v1 525 526 Trapped Ion Mobility Time of Flight Mass Spectrometry of Immunoprecipitated Placental Protein 527 Lysate 528 529 Term placenta from a single donor was harvested within 2 hours of delivery. Implantation side of 530 placenta was dissected to 1 cm3 blocks and flash frozen in liquid nitrogen. Placenta protein lysate 531 was prepared by mechanically dissociating tissue using a glass Dounce homogenizer and an electric 532 tissue homogenizer in RIPA buffer with protease inhibitors (Roche) on ice. Protein was quantified 533 using Bradford Assay (Pierce), normalized to 500µg in TNP40 and incubated with 1µL of human sera 534 overnight at 4C with overhead mixing. Proteins were subsequently incubated with a 1:1 mix of protein 535 A and protein G Dynabeads (Thermo) for 1 hour and washed five times with RIPA and once with 536 TrisHCl. A NanoDrop reading was taken under TrisHCl to quantify protein. 537 538 Supernatant was removed and proteins were then resuspended in 8M urea, 50mM Tris (pH 8) before 539 being subject to a standard on-bead protein digestion. The protein digestion procedure included 540 disulfide reduction with 5 mM DTT, alkylation with 14 mM iodoacetamide, and an overnight digestion 541 with Lys/C at a protein to protease ratio of 1:50. In order to obtain an optimal tryptic digestion, the 542 urea concentration was then diluted to 2 M with 50 mM Tris (pH 8) followed by the addition of trypsin 543 to a protein to protease ratio of 1:50, and a 4 hour incubation at room temperature. The peptide 544 supernatant pH was then decreased to 2 with 7 µL formic acid (FA) and separated from the beads 545 with a magnet. An extraction was performed on the beads by resuspension with 0.1% FA, the 546 supernatant of which was added to the final sample. Peptide desalting was performed with the 547 AssayMAP Bravo (Agilent Technologies) using a standard RPS peptide cleanup cartridges and 548 protocol60. Approximately 25 ng of desalted peptides were analyzed on a TIMSTOF SCP mass 549 spectrometer (Bruker Corporation) coupled with an EASY-nLC 1200 LC system (Thermo Fisher 550 Scientific). Peptides were separated by reverse-phase chromatography on a 25 cm column (75-μm 551 inner diameter, packed with 1.6 μm C18 resin, AUR2-25075C18A-CSI; IonOpticks). Peptides were 552 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint introduced into the mass spectrometer using a gradient starting with 2% to 8% buffer B (0.1% (v/v) 553 formic acid in 80% acetonitrile) for 1 min followed by an increase to 25% buffer B for 25 min then an 554 increase to 40% buffer B for 5 min at a flow rate of 100 nL/min and were ionized by CSI (captive 555 spray ionization). 556 557 Samples were first analyzed by PASEF-dda with a duty cycle of 1.03 sec comprising one MS1 survey 558 for every 5 PASEF MS2 ramps and a mass range between 100 and 1700 m/z. The TIMS device 559 ramp/accumulation time was set to 166 ms with a 100% duty cycle and a mobility range of 0.7 to 1.3 560 1/K0. Precursor ions from the survey scan were selected using an adjusted isolation width of 2 m/z 561 below 700 m/z and 3 m/z above 800 m/z precursor mass-to-charge value in the quadrupole that 562 aligns with the anticipated TIMS elution time relative to the ion mobility value. Fragmentation was 563 performed by Collision Induced Dissociation (CID) with CE (collisional energy) interpolation between 564 42 eV at 0.65 1/K0, 31.92 eV at 0.8 1/K0, 36.96 eV at 1 1/K0, 42 eV at 1.2 1/K0, 47.04 eV at 1.4 1/K0 565 and 51.24 eV at 1.6 1/K0. Active exclusion of precursors were released after 0.2 minutes. Precursor 566 repetitions were set to a target intensity of 20000 and a threshold of 500. 567 568 Real-time searching was performed with PaSER (Parallel Search Engine in Real-Time) against a 569 human Uniprot database (downloaded on 30 July 2022) using a reverse-decoy method with default 570 settings. A spectral library was generated from the searched DDA data using PaSER. Samples were 571 then analyzed by PASEF-dia using a high speed acquisition scheme from Meier F et al and 572 searched in real-time against the DDA generated spectral library using the TIMS DIA-NN algorithm 573 with default settings. 574 575 Peptide counts were collapsed to the corresponding protein, log-normalized, and fold changes and z-576 scores were calculated over mean signal in term pregnancies. Preterm-specific hits were identified by 577 requiring that at least one preterm pregnancy and no term pregnancies met the threshold of z>=5 for 578 a given protein. 579 580 Placental Immunofluorescence 581 Placental tissue was harvested within 2 hours of delivery and embedded, unfixed, in O.C.T compound 582 (Tissue-Tek) and flash frozen in an isopentanol bath submerged in liquid nitrogen. Ten-micron 583 cryosections were cut from frozen blocks and adhered to glass slides. Slides were thawed and briefly 584 post-fixed with acetone, rehydrated in 1x PBS, blocked with 10% (v/v) goat serum in 1x PBS for 2 585 hours, and incubated with 1:1000 rabbit anti-human anti-IL1RA (Millipore Sigma Cat No. HPA001482) 586 and 1:500 rat anti-human CD31 (Abcam Cat. No. ab9498) overnight at 4C in a humidified chamber. 587 Primary antibodies were washed using standard protocols and incubated with anti-rabbit and anti-rat 588 secondaries using standard protocols. Secondary only controls were stained identically but omitting 589 primary antibody. Stained tissue was imaged using the Crest LFOV Spinning Disk/ C2 Confocal at 590 400x magnification under identical camera exposure and laser settings for secondary only and 591 experimental samples. Micrograph exposure was normalized to secondary only negative control in 592 FIJI and applied to all images at once. In peptide blocking experiments, IL1RA PrEST Antigen 593 (Millipore Sigma APREST83081) was incubated with anti-IL1RA antibody per manufacturer’s 594 instructions at 4C overnight and used to stain tissue as above. In serum blocking experiments, 595 placental sections were blocked with 1% (v/v) goat serum and 10% human serum in PBS-T overnight 596 at 4C. Controls were blocked without human serum. Placenta was subsequently probed with the 597 commercial anti-IL1RA antibody (Millipore Sigma Cat No. HPA001482) and imaged as above. 598 599 In vitro IL1RA blocking assay 600 Diluted patient sera, anti-IL1RA (R&D AF-280-NA), or anti-GFAP were pre-incubated with human IL1-601 RA (Peprotech) overnight and applied to the HEK-Blue IL1b cells (InvivoGen) for 2 hours. Cells were 602 subsequently stimulated with human IL1b (Peprotech) for 72 hours at 37C. Cells treated with IL1b 603 alone or TNFa alone (Peprotech) were used as positive and negative controls, respectively. 604 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint Supernatants were assayed using the QANTI-Blue assay (InvivoGen), as previously described51,52. 605 IL1b activity index was calculated by measuring SEAP activity at OD655 every 15 minutes for 2 hours 606 and determining the rate of enzymatic activity. Slopes were background subtracted (cell treated with 607 TNFa) and normalized to mean signal in IL1b treated cells. IL1b activity index was defined as the fold 608 change of signal in serum or antibody treated cells over signal in cells incubated with IL1RA and IL1b 609 only. Each experiment was carried out in triplicate for a total of three biological replicates. 610 611 Murine Il1ra dot blot 612 500 or 250 nanograms of recombinant mouse Il1ra (R&D 480-RM-050) were blotted onto 613 nitrocellulose using a custom 3D printed dot blot device under vacuum. Membranes were blocked for 614 1 hour at room temperature in 5% (w/v) milk and subsequently probed overnight at 4C with either 615 anti-human IL1RA antibodies (R&D AF-280-NA or Sigma Prestige HPA001482) or anti-mouse Il1ra 616 antibodies (Invitrogen PA5-21776 or R&D MAB4801) or no primary antibody. Following standard 617 washing procedures, membranes were probed with relevant secondaries conjugated to Licor-IR dyes 618 and blots were imaged using the Licor Odyssey machine. Membranes were imaged together in one 619 scan under identical exposure settings. 620 621 Animal Husbandry and Injections 622 All mice were housed, bred, and maintained in a pathogen-free facility at the University of California 623 San Francisco (UCSF). All procedures were performed in concordance with UCSF Institutional 624 Animal Care and Use Committee (IACUC) regulations and approved protocol. Timed-pregnant 625 C57BL/6 females mated with C57BL/6 males were purchased from Jackson Laboratories (strain 626 #000664). Retro-orbital (RO) injections were done following with UCSF IACUC procedural guidelines. 627 Briefly, pregnant dams were anesthetized at E13.5-E15.5 with isoflurane and injected in the RO sinus 628 using 0.5mL insulin syringes, with a maximum volume of 150µL containing 0 or 5µg of human IL1b 629 (Peprotech 200-01B) and 50µg of polyclonal goat IgG (isotype control; R&D AB-108-C) or of goat 630 anti-human IL1RA antibody (R&D AF-280-NA) in sterile PBS. Reagents were mixed immediately prior 631 to injection. 632 633 Mouse harvesting 634 Mice were euthanized at E18.5 with CO2 and a sample of serum was obtained via transcardial 635 puncture. Resorbed and malpefrused fetuses were quantified. Whole fetuses and their placentas 636 were harvested and weighed. Murine Il1b concentrations were measured in maternal serum by a 637 commercial ELISA (Abcam Cat No. ab197742). 638 639 Histological analysis of murine placenta 640 Whole placentas were fixed in 4% paraformaldehyde and embedded in paraffin using standard 641 protocols. Embedding, grossing, sectioning, immunohistochemistry (IHC) with anti-human antibodies 642 to anti-CD68, anti-cleaved Caspase-3, or anti-CD31, and scanning was performed at HistoWiz. 643 Blinded H&E slides were assessed by a licensed pathologist. IHC was quantified by defining 5 non-644 overlapping regions in each of the two placental sections for a total of 10 regions for each placenta in 645 QuPath software. Regions were randomly drawn and covered most of the placental area, avoiding 646 anomalous regions (e.g. small tears in tissue). For anti-CD68 IHC, regions of interest were drawn in 647 the labyrinth region of the placenta. A pixel thresholding classifier was built for each of the antibody 648 stains using a Laplacian of Gaussian prefilter setting. DAB positive area was calculated as a 649 percentage of total area within the defined region. Each individual region of interest was reported for 650 isotype or anti-IL1RA treated conditions. 651 652 Luminex assay to detect anti-IL1RA antibodies in patient sera 653 Recombinant IL1RA (Peprotech) or BSA (Thermo) were conjugated to spectrally-distinct Luminex 654 beads in separate 1.5mL protein LoBind tubes. Each bead conjugation was performed as previously 655 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint described62 using the Antibody Coupling Kit following manufacturer’s instructions (Luminex, 40-656 50016). All serological analyses were performed exactly as previously described62 in technical 657 duplicate on two separate experimental days. Positive samples (with greater than 20 net MFI 658 IL1RA/BSA) were repeated on an additional experimental day in technical duplicate. All net MFI 659 IL1RA/BSA values were averaged and z-scores above mean in term were calculated. 660 661 Bioinformatic Analyses 662 To analyze peptide enrichment after PhIP-seq, reads were aligned at the protein level using 663 RAPsearch63, as previously described . Aligned reads were normalized to 100,000 reads per k-mer 664 (RPK) to account for varying read-depth. All downstream analyses were performed by using an 665 implementation of PhagePy python package (https://github.com/h-s-miller/phagepy). 666 667 All data was initially filtered to samples with fewer than recovered 100,000 reads. Data was 668 pseudocounted and fold change over mock IP controls was calculated. Co-correlations of the fold 669 change over mock IP matrix were used to determine technical replicate consistency. Subsequently, 670 technically replicated samples (with the exception of positive and mock IP controls) were averaged. 671 672 Fold change over mock IP was re-calculated with averaged values. Additional fold changes were 673 calculated over mean in healthy term pregnancies from Cohort I only (see above), preterm 674 pregnancies (downsampled to n=204, five times, then averaged), or over all term pregnancies. Z-675 scores were calculated for each of the fold-change matrixes after log-10 transformation relative to 676 mock IP, healthy term pregnancies, preterm pregnancies, and over all term pregnancies. Peptides 677 with a z-score greater than 3 in a minimum of 4 experimental samples and 0 control samples were 678 considered specific for that comparison, unless otherwise indicated. Analysis was performed on all 679 cohorts together, except when performing comparisons for Cohort II PhIP-seq data with Cohort II IP-680 MS data, where the same analysis was performed using samples only from Cohort II. Autoreactivities 681 were counted per person by summing up binarized per peptide hits within the preterm- or term-682 specific PhIP-seq signature. Where indicated, cumulative enrichment was calculated by summing log-683 transformed enrichments over mock IPs for peptides identified as preterm- or term-specific. 684 685 Positional enrichment analysis across IL1RA was performed as previously described65 for Cohort I 686 samples. GSEA analysis was performed by manually annotating genes with previously reported roles 687 in placental function and pregnancy. GO term enrichment analysis supplemented manual grouping 688 and was performed utilizing the gseapy package. 689 690 Machine Learning Predictive Modelling 691 Data and preprocessing 692 Gestational age at sampling was normalized to the average term pregnancy (40 weeks for all 693 samples) and included as a co-variate. Clinical outcome labels were encoded as binary preterm birth 694 (ptb: 1 = preterm, 0 = term). One auxiliary grouping variable was constructed to support fold-change 695 estimation and stratified splitting: hc_term (1 = healthy term controls from Cohort I; 0 = all preterm). 696 697 Fold-change matrices 698 Peptide-level fold change matrices were computed as above over healthy term controls. 699 Unless otherwise noted, downstream predictive modeling used this log-transformed matrix as 700 features, and a sparse “selection-only” matrix of binarized peptides with a z-score greater than 701 3 over healthy term controls to drive in-fold feature selection (see below). 702 703 Train/holdout partitioning 704 Samples were split once into train and holdout sets. Splits were stratified by outcome and 705 constrained by cohort and gestational-age bins ( bins = 12, 16, 20, 24, 28, 32, 36, 40 weeks). 706 When available for longitudinal cohorts, subject barcode column was used to prevent leakage 707 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint by keeping all samples from a subject within a single partition. All model development and 708 cross-validation used only the training partition; final performance was assessed on the 709 untouched holdout. 710 711 Feature construction and covariates 712 For modeling, the feature design matrix was built log10-transformed fold change over healthy 713 term controls and augmented with the prespecified normalized gestational age at time of 714 sampling, which was always included. The feature selection matrix was binarized peptides with 715 a z-score greater than 3 over healthy term. 716 717 In-fold supervised feature selection (dual-matrix) 718 We implemented a fast, deterministic, in-fold selection that evaluates binary peptide “hits” from 719 the selection matrix while training models on the continuous log10-transformed fold change 720 matrix. Prior to cross-validation, the selection matrix was binarized once (hit_cutoff = 0.0). 721 Within each training fold, features were retained if they met: 722 minimum hit count in preterm = 4, and 723 maximum hit count in term (any term sample) = 12. 724 725 If no features met criteria, a fallback chose the top fallback_k = 50 features ranked by 726 (case_hits − control_hits). A hard cap of max_features_per_fold = 5000 limited the per-fold 727 dimensionality. Selection yielded names that were mapped to column indices of the FC matrix; 728 GA_test_norm was appended post-selection. All selection occurred inside each training fold to 729 avoid information leakage. The selection logic was encapsulated in a scikit-learn–compatible 730 transformer (DualMatrixFoldSelector). 731 732 Classifier and hyperparameters 733 The core classifier was logistic regression with L2 penalty (solver="lbfgs", C=1.0, 734 max_iter=2000) and class_weight="balanced". Features were imputed with a constant fill (0.0) 735 and standardized (z-score) within each fold. All preprocessing steps (selection → imputation → 736 scaling → classifier) were wrapped in a scikit-learn Pipeline to ensure proper cross-validation 737 hygiene. 738 739 Cross-validation, out-of-fold (OOF) predictions, and model selection 740 To perform feature selection, we ran 1,000 outer iterations of 5-fold stratified CV on the training 741 set (StratifiedKFold with shuffling; per-iteration seed i + 3). For each iteration and fold: 742 The fold-specific feature set was selected using the training portion only. 743 The pipeline was fitted and used to produce OOF probabilities for the held-out fold. 744 Across iterations, we recorded: 745 per-fold ROC-AUC and PR-AUC, 746 per-iteration mean ROC-AUC (averaged over the 5 folds), and 747 the full OOF probability matrix (iterations × samples). 748 749 We also retained the per-fold selected feature lists and summarized their stability (see below). 750 For qualitative reference, we saved the best and worst single fold models by ROC-AUC across 751 all iterations. 752 753 Feature stability, intersections, and frequency panels 754 From all selection outputs (iteration × fold), we computed: 755 (i) the global intersection (features present in every fold of every iteration), 756 (ii) per-iteration intersections (features present in all 5 folds of an iteration), and 757 (iii) frequency of selection across all folds. We defined stability panels at thresholds of 758 ≥80%, ≥50%, and ≥25% of folds. We also constructed a frequency-top-N panel (e.g., 759 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint top 500 by fold count). All panels included normalized gestational age at sampling at 760 training time. 761 762 Panel training, coefficients, and OOF performance 763 For each panel, we trained a fixed-column logistic-regression pipeline (column locking → 764 imputation → scaling → LR) and generated 5-fold OOF predictions on the training set (one 765 repeat). We reported OOF ROC-AUC and PR-AUC and plotted OOF ROC curves. Final panel 766 models were then refit on all training samples and saved with their locked column order. We 767 exported standardized coefficients (per-SD) and approximate raw-scale coefficients together 768 with the intercept to facilitate interpretation. 769 770 Holdout evaluation and subgroup analyses 771 Each saved panel model was evaluated on the held-out test set created by Step1_split. We 772 computed ROC-AUC and rendered holdout ROC curves. To examine performance 773 heterogeneity, we generated stratified ROCs contrasting preterm subtypes versus term (PTB 774 type: iatrogenic, spontaneous) and timepoint (First, Second, Third, Cord_blood). For subgroup 775 plots, we applied the same prediction vectors but stratified truth labels and sample indices per 776 category. For all ROC curves, the trained model generated a predicted probability of preterm 777 birth for every sample, and the ROC-AUC was calculated directly from these predicted 778 probabilities against the true binary outcome. 779 780 Reproducibility and implementation details 781 All analyses were performed in Python using AnnData for matrix management and scikit-learn 782 for model building. Randomness entered only through the CV shuffling seeds and the initial 783 train/holdout split; fixed seeds are stated above. No global prefiltering was applied to the FC 784 feature space beyond per-fold selection. 785 786 Reporting 787 All reported cross-validated metrics are computed strictly within the training partition using 788 OOF predictions to avoid optimistic bias. Holdout metrics are computed once on the untouched 789 test set. Subgroup ROCs are descriptive and based on the same prediction vectors, stratified 790 by the indicated metadata fields. 791 792 Data Availability 793 All raw and processed data, the forecasting model and feature weights, as well as the associated 794 code are available for download at Dryad. Link for peer review: 795 http://datadryad.org/stash/share/gOsb-l9_flhFstHmz1-CuX7BpqmGOf1DF1qQEX-UVK0 . PhIP-seq 796 analytical code is freely available on GitHub: https://github.com/h-s-miller/phagepy 797 798 799 800

801 802 1. Goldenberg, R. L., Culhane, J. F., Iams, J. D. & Romero, R. Epidemiology and causes of preterm 803 birth. Lancet (London, England) 371, 75–84 (2008). 804 2. Liu, L. et al. Global, regional, and national causes of under-5 mortality in 2000–15: an updated 805 systematic analysis with implications for the Sustainable Development Goals. The Lancet 388, 806 3027–3035 (2016). 807 3. Törnblom, S. A. et al. Non-infected preterm parturition is related to increased concentrations of IL-808 6, IL-8 and MCP-1 in human cervix. Reproductive Biology and Endocrinology 3, 1–10 (2005). 809 4. Frascoli, M. et al. Alloreactive fetal T cells promote uterine contractility in preterm labor via IFN-γ 810 and TNF-α. Science Translational Medicine 10, (2018). 811 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint 5. PrabhuDas, M. et al. Immune mechanisms at the maternal-fetal interface: Perspectives and 812 challenges. Nature Immunology 16, 328–334 (2015). 813 6. Urbaniak, S. J. & Greiss, M. A. RhD haemolytic disease of the fetus and the newborn. Blood 814 Reviews 14, 44–61 (2000). 815 7. Lee, J. et al. Maternal HLA Panel-Reactive Antibodies in Early Gestation Positively Correlate with 816 Chronic Chorioamnionitis: Evidence in Support of the Chronic Nature of Maternal Anti-fetal 817 Rejection. American Journal of Reproductive Immunology 66, 510–526 (2011). 818 8. Bell, S. C. & Billington, W. D. Major anti-paternal alloantibody induced by murine pregnancy is 819 non-complement-fixing IgG1. Nature 288, 387–388 (1980). 820 9. Suah, A. N. et al. Pregnancy-induced humoral sensitization overrides T cell tolerance to fetus-821 matched allografts in mice. Journal of Clinical Investigation 131, (2021). 822 10. Hönger, G. et al. Frequency and determinants of pregnancy-induced child-specific sensitization. 823 American Journal of Transplantation 13, 746–753 (2013). 824 11. Küssel, L. et al. Longitudinal assessment of HLA and MIC-A antibodies in uneventful pregnancies 825 and pregnancies complicated by preeclampsia or gestational diabetes. Scientific Reports 7, 1–10 826 (2017). 827 12. Thangaratinam, S. et al. Association between thyroid autoantibodies and miscarriage and preterm 828 birth: Meta-analysis of evidence. Bmj 342, 1–8 (2011). 829 13. Wallukat, G. et al. Patients with preeclampsia develop agonistic autoantibodies against the 830 angiotensin AT1 receptor. Journal of Clinical Investigation 103, 945–952 (1999). 831 14. Zhou, C. C. et al. Angiotensin receptor agonistic autoantibodies induce pre-eclampsia in pregnant 832 mice. Nature Medicine 14, 855–862 (2008). 833 15. Yoshimura, K. & Hirsch, E. Effect of stimulation and antagonism of interleukin-1 signaling on 834 preterm delivery in mice. J Soc Gynecol Investig 12, 533–538 (2005). 835 16. Romero, R. et al. A fetal systemic inflammatory response is followed by the spontaneous onset of 836 preterm parturition. American Journal of Obstetrics and Gynecology 179, 186–193 (1998). 837 17. Romero, R., Mazor, M. & Tartakovsky, B. Systemic administration of interleukin-1 induces preterm 838 parturition in mice. Am J Obstet Gynecol 165, 969–971 (1991). 839 18. Nadeau-Vallée, M. et al. Antenatal Suppression of IL-1 Protects against Inflammation-Induced 840 Fetal Injury and Improves Neonatal and Developmental Outcomes in Mice. The Journal of 841 Immunology 198, 2047–2062 (2017). 842 19. Girard, S., Tremblay, L., Lepage, M. & Sébire, G. IL-1 Receptor Antagonist Protects against 843 Placental and Neurodevelopmental Defects Induced by Maternal Inflammation. The Journal of 844 Immunology 184, 3997–4005 (2010). 845 20. Sadowsky, D. W., Adams, K. M., Gravett, M. G., Witkin, S. S. & Novy, M. J. Preterm labor is 846 induced by intraamniotic infusions of interleukin-1β and tumor necrosis factor-α but not by 847 interleukin-6 or interleukin-8 in a nonhuman primate model. American Journal of Obstetrics and 848 Gynecology 195, 1578–1589 (2006). 849 21. Sadowsky, D. W., Haluska, G. J., Gravett, M. G., Witkin, S. S. & Novy, M. J. Indomethacin blocks 850 interleukin 1β–induced myometrial contractions in pregnant rhesus monkeys. American Journal of 851 Obstetrics and Gynecology 183, 173–180 (2000). 852 22. Baggia, S., Gravett, M. G., Witkin, S. S., Haluska, G. J. & Novy, M. J. Interleukin-1 beta intra-853 amniotic infusion induces tumor necrosis factor-alpha, prostaglandin production, and preterm 854 contractions in pregnant rhesus monkeys. J Soc Gynecol Investig 3, 121–126 (1996). 855 23. Witkin, S. S., Gravett, M. G., Haluska, G. J. & Novy, M. J. Induction of interleukin-1 receptor 856 antagonist in rhesus monkeys after intraamniotic infection with group B streptococci or interleukin-857 1 infusion. Am J Obstet Gynecol 171, 1668–1672 (1994). 858 24. Gravett, M. G. et al. An experimental model for intraamniotic infection and preterm labor in rhesus 859 monkeys. Am J Obstet Gynecol 171, 1660–1667 (1994). 860 25. Fidel Jr, P. l. et al. Interleukin-1 Receptor Antagonist (IL-1ra) Production by Human Amnion, 861 Chorion, and Decidua. American Journal of Reproductive Immunology 32, 1–7 (1994). 862 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint 26. Genc, M. R. et al. A disproportionate increase in IL-1β over IL-1ra in the cervicovaginal secretions 863 of pregnant women with alteredvaginal microflora correlates with preterm birth. American Journal 864 of Obstetrics and Gynecology 190, 1191–1197 (2004). 865 27. Romero, R. et al. Interlukin-1 stimulates prostaglandin biosynthesis by human amnion. 866 Prostaglandins 37, 13–22 (1989). 867 28. Heng, Y. J. et al. The interplay of the interleukin 1 system in pregnancy and labor. Reproductive 868 Sciences 21, 122–130 (2014). 869 29. Yockey, L. J. & Iwasaki, A. Interferons and Proinflammatory Cytokines in Pregnancy and Fetal 870 Development. Immunity 49, 397–412 (2018). 871 30. Stock, S. J. & Aiken, C. E. Barriers to progress in pregnancy research: How can we break 872 through? Science 380, 150–153 (2023). 873 31. Chang, C. Y. et al. Withdrawing Approval of Makena — A Proposal from the FDA Center for Drug 874 Evaluation and Research. New England Journal of Medicine 383, e131 (2020). 875 32. Zhang, G. et al. Genetic Associations with Gestational Duration and Spontaneous Preterm Birth. 876 N Engl J Med 377, 1156–1167 (2017). 877 33. Presicce, P. et al. Amnion responses to intrauterine inflammation and effects of inhibition of TNF 878 signaling in preterm Rhesus macaque. iScience 26, 108118 (2023). 879 34. Halkias, J. et al. CD161 contributes to prenatal immune suppression of IFNγ-producing PLZF+ T 880 cells. The Journal of Clinical Investigation https://doi.org/10.1172/JCI125957 (2019) 881 doi:10.1172/JCI125957. 882 35. Padula, A. M. et al. Environmental pollution and social factors as contributors to preterm birth in 883 Fresno County. Environ Health 17, 70 (2018). 884 36. York, T. P., Eaves, L. J., Neale, M. C. & Strauss, J. F. The contribution of genetic and 885 environmental factors to the duration of pregnancy. Am J Obstet Gynecol 210, 398–405 (2014). 886 37. Jain, V. G. et al. IRAK1 Is a Critical Mediator of Inflammation-Induced Preterm Birth. J Immunol 887 204, 2651–2660 (2020). 888 38. Kallapur, S. G. et al. Intra-amniotic IL-1β induces fetal inflammation in rhesus monkeys and alters 889 the regulatory T cell/IL-17 balance. J Immunol 191, 1102–1109 (2013). 890 39. Presicce, P. et al. IL-1 and TNF mediates IL-6 signaling at the maternal-fetal interface during 891 intrauterine inflammation. Front Immunol 15, 1416162 (2024). 892 40. Busse, M. et al. Regulatory B Cells Are Decreased and Impaired in Their Function in Peripheral 893 Maternal Blood in Pre-term Birth. Frontiers in Immunology 11, 1–10 (2020). 894 41. Rizzuto, G. et al. Establishment of fetomaternal tolerance through glycan-mediated B cell 895 suppression. Nature https://doi.org/10.1038/s41586-022-04471-0 (2022) doi:10.1038/s41586-022-896 04471-0. 897 42. Yougbaré, I. et al. Maternal anti-platelet β3 integrins impair angiogenesis and cause intracranial 898 hemorrhage. J Clin Invest 125, 1545–1556 (2015). 899 43. O’Donovan, B. et al. High-resolution epitope mapping of anti-Hu and anti-Yo autoimmunity by 900 programmable phage display. Brain Communications 2, 1–16 (2020). 901 44. Larman, H. B. et al. Autoantigen discovery with a synthetic human peptidome. Nature 902 Biotechnology 29, 535–541 (2011). 903 45. Bodansky, A. et al. Unveiling the proteome-wide autoreactome enables enhanced evaluation of 904 emerging CAR T cell therapies in autoimmunity. J Clin Invest 134, (2024). 905 46. Inoue, N., Satouh, Y. & Wada, I. IZUMO family member 3, IZUMO3, is involved in male fertility 906 through the acrosome formation. Molecular Reproduction and Development 88, 479–481 (2021). 907 47. Pavličev, M. et al. A common allele increases endometrial Wnt4 expression, with antagonistic 908 implications for pregnancy, reproductive cancers, and endometriosis. Nat Commun 15, 1152 909 (2024). 910 48. Lee, J. et al. Detection of anti-HLA antibodies in maternal blood in the second trimester to identify 911 patients at risk for antibody-mediated maternal anti-fetal rejection and spontaneous preterm 912 delivery. Am J Reprod Immunol 70, 162–175 (2013). 913 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint 49. Bastard, P. et al. Autoantibodies against type I IFNs in patients with life-threatening COVID-19. 914 Science 370, (2020). 915 50. Thurner, L. et al. IL-1RA Antibodies in Myocarditis after SARS-CoV-2 Vaccination. N Engl J Med 916 387, 1524–1527 (2022). 917 51. Jarrell, J. A. et al. Neutralizing anti-IL-1 receptor antagonist autoantibodies induce inflammatory 918 and fibrotic mediators in IgG4-related disease. J Allergy Clin Immunol 149, 358–368 (2022). 919 52. Pfeifer, J. et al. Autoantibodies against interleukin-1 receptor antagonist in multisystem 920 inflammatory syndrome in children: a multicentre, retrospective, cohort study. The Lancet 921 Rheumatology 4, e329–e337 (2022). 922 53. Hoffmann, M.-C. et al. Autoantibody-Mediated Depletion of IL-1RA in Still’s Disease and Potential 923 Impact of IL-1 Targeting Therapies. J Clin Immunol 44, 45 (2024). 924 54. Schreuder, H. et al. A new cytokine-receptor binding mode revealed by the crystal structure of the 925 IL-1 receptor with an antagonist. Nature 386, 194–200 (1997). 926 55. Krishnan, S. & Raibekas, A. A. Multistep Aggregation Pathway of Human Interleukin-1 Receptor 927 Antagonist : Kinetic , Structural , and Morphological Characterization. Biophysj 96, 199–208 928 (2009). 929 56. Ikonomidis, I. et al. Inhibition of Interleukin-1 by Anakinra Improves Vascular and Left Ventricular 930 Function in Patients With Rheumatoid Arthritis. Circulation 117, 2662–2669 (2008). 931 57. Nuki, G., Bresnihan, B., Bear, M. B., McCabe, D., & European Group Of Clinical Investigators. 932 Long-term safety and maintenance of clinical improvement following treatment with anakinra 933 (recombinant human interleukin-1 receptor antagonist) in patients with rheumatoid arthritis: 934 extension phase of a randomized, double-blind, placebo-controlled trial. Arthritis Rheum 46, 935 2838–2846 (2002). 936 58. Brien, M.-E. et al. A Systematic Review of the Safety of Blocking the IL-1 System in Human 937 Pregnancy. Journal of Clinical Medicine 11, 225 (2022). 938 59. Litonjua, A. A. et al. Six-Year Follow-up of a Trial of Antenatal Vitamin D for Asthma Reduction. 939 New England Journal of Medicine 382, 525–533 (2020). 940 60. Zhang, L., McAlpine, P. L., Heberling, M. L. & Elias, J. E. Automated Ligand Purification Platform 941 Accelerates Immunopeptidome Analysis by Mass Spectrometry. J. Proteome Res. 20, 393–408 942 (2021). 943 61. Meier, F. et al. diaPASEF: parallel accumulation–serial fragmentation combined with data-944 independent acquisition. Nat Methods 17, 1229–1236 (2020). 945 62. Zamecnik, C. R. et al. An autoantibody signature predictive for multiple sclerosis. Nat Med 30, 946 1300–1308 (2024). 947 63. Zhao, Y., Tang, H. & Ye, Y. RAPSearch2: a fast and memory-efficient protein similarity search 948 tool for next-generation sequencing data. Bioinformatics 28, 125–126 (2012). 949 64. Vazquez, S. E. et al. Identification of novel, clinically correlated autoantigens in the monogenic 950 autoimmune syndrome APS1 by proteome-wide PhIP-Seq. eLife 9, 1–25 (2020). 951 65. Rajan, J. V. et al. Phage display demonstrates durable differences in serological profile by route of 952 inoculation in primary infections of non-human primates with Dengue Virus 1. Scientific Reports 953 11, 1–12 (2021). 954 955

956 We thank members of the DeRisi Lab for helpful discussions and JM Rackaitis for support during 957 these studies. We thank Joelle Ostroff for inspiring the studies herein. We also acknowledge the New 958 York Blood Center for contribution of healthy control plasma. We thank the UCSF CALM-NIC 959 microscopy core for use of the CREST LFOV Spinning Disk/C2 confocal funded by the UCSF 960 Program for Breakthrough Biomedical Research, the Sandler Foundation, Strategic Advisory 961 Committee, and the EVCP Office Research Resource Program Institutional Matching Instrumentation 962 Award. ER is funded by 2022 Next Gen Pregnancy Initiative Research Grant from the Burroughs 963 Wellcome Fund, the Eunice Kennedy Shriver National Institute of Child Health and Human 964 Development award 5T32HD098057, and National Institute of Allergy and Infectious Diseases award 965 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint K99AI182451. CMB is funded by The Emiko Terasaki Foundation (project 7027742/fund B73335) and 966 by the National Institute of Neurological Disorders and Stroke of the NIH award K99NS117800. GR is 967 funded by National Institute of Allergy and Infectious Diseases award K08AI137209 and 2023 Next 968 Gen Pregnancy Initiative Research Grant from the Burroughs Wellcome Fund. HK is supported by 969 NICHD F30HD117526. SLG is supported by National Institute of Allergy and Infectious Diseases 970 award K08AI141728 and R01HD111582. The Host-pathogen group (AB) at Fondazione IRCCS 971 Policlinico San Matteo, Pavia, Italy, is supported by grants from the Italian Ministry of Health, 972 RC08061819 and RC08061822, and by 5X1000 grant 08061821from the San Matteo Hospital. NP 973 and JLS are supported by National Heart Lung and Blood Institute award R01HL169300. JLD is 974 supported through funding from the Chan Zuckerberg Biohub. Contents herein are the sole 975 responsibility of the authors and do not necessarily represent the official views of the NIH or other 976 funding agencies. 977 978 AUTHOR CONTRIBUTIONS 979 ER and JLD designed the research. ER, BB, HMK, HSM, SJS, KCZ, RW, FM, JSC, MM, EK, RP, AK, 980 DJLY, CC, SG, AD, QK, GW, AS, SAM, AM, GR performed the research. ER, HMK, HSM, AFK, 981 CMB, JSC, FM, JEE, and JLD contributed to new reagents and analytic tools. SO, RW, DH, RJB, 982 KKR, SLG, SLH, LLJP, JH, NCS, FT, CA, TB, RA, CB, BG, AB, NP, JLS, MC, provided clinical 983 samples, metadata, and/or contributed to clinical interpretation of data. ER, JLD, TCM, MSA, MRW 984 provided supervision of the work. JSC, FM, and JEE generated and analyzed mass spectrometry 985 datasets. ER, BB, and TCM provided significant contributions to mouse model development. ER and 986 JLD analyzed data and produced figures. ER and JLD wrote the manuscript. All authors reviewed the 987 manuscript and provided feedback. 988 989 COMPETING INTERESTS 990 ER and JLD are inventors on a patent application submitted by the Regents of the University of 991 California and the Chan Zuckerberg Biohub San Francisco. JLD reports being a founder and paid 992 consultant for Delve Bio, Inc., and a paid consultant for PHC Global, Inc. JLD, MRW and CMB 993 receive licensing fees from CDI Labs. MRW reports being a founder and board member for Delve Bio, 994 Inc., a paid consultant for Vertex Pharmaceuticals, Ouro Medicines, Indapta Therapeutics and Pfizer, 995 and the recipient of unrelated research grant support from Genentech / Roche, Novartis, and Kyverna 996 Therapeutics. JLS is a scientific consultant to TruDiagnostic and Antipode. 997 998

AND CORRESPONDENCE 999 Correspondence and requests for materials should be addressed to Joseph L. DeRisi, 1000 [email protected]. 1001 1002 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint Cohorts I, II, III, IV, V, VII, VIII n=2,719 Male or Nulligravida n=130 Pregnancy PhIP-seq SignatureFeature selectionTest 3,096 peptides Cohort VI: IVF timecourse 2,244 proteins log10 FCindividual pre-pregnancy baseline Pregnant Non-pregnant Logistic regression classifier z≥3 over baseline in n≥10 pregnant samples & n=0 baselines n=4,512 PhIP-seq enrichments Quality Control maternal autoreactivity enrichment atlas n=2,194 nt=3,101 V n=143 Pregnancy Cohorts cross-sectional longitudinal I n=858 IV n=69 VI VII III n=20 nt=89 III n=20 nt=89 II n=80 nt=96 VIII n=13 nt=97 n=214 nt=341 n=797 nt=1,408 a b c ed Figure 1 Figure 1. Maternal autoantibody signature induced by becoming pregnant a. Schematic of study design integrating data from eight human pregnancy cohorts of maternal sera human analyzed by PhIP-seq with the human peptidome (n indicates human subject counts, nt indicates human samples across multiple timepoints). b. Histogram of gestational age at time of sampling in weeks for women that delivered term and preterm across all cohorts tested. c. Schematic for identifying the PhIP-seq signature utilizing PhIP-seq data. Feature selection was performed on Cohort VI, where log10 fold changes were taken over individual pre-pregnancy baselines and z-scores were calculated over this baseline. Peptides were considered pregnancy-specific if they were detected in 10 pregnant samples and zero baselines. These features were used to train a five-fold cross validated logistic regression classifier to discriminate pregnant (Cohorts I-V,VII-VIII) and non-pregnant samples. d. Five-fold cross-validated performance of features selected from Cohort VI on pregnant (Cohorts I-V,VII-VIII) and non-pregnant samples. e. Sum fold change over mock immunoprecipitation of selected proteins in the PhIP-seq pregnancy signature in pregnant or never pregnant samples. Mann-Whitney U test for significance in b. Pregnant Σ log10(FCmock IP) Histocompatibility Pregnancy Sperm Previously reported autoreactivities in pregnancy Yes Never 0 5000 ≥10000 -20 0 20 40 60 Gestational age at test (weeks) 0.02 0.040 0.06 Density p=0.3 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Delivered term Delivered preterm False Positive Rate True Positive Rate 5-fold CV AUC=0.93 ± 0.01 Mean ROC ±1 std. dev. Shuffled Labels Chance HLA-C HMHA1 UTY KDM5D CGB7 TG TPO TRIM21 ACRV1 SPAG17 SPAG9 ODF2 TEX15 TCTEX1D4 . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint . CC-BY-NC 4.0 International licenseIt is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted December 5, 2025. ; https://doi.org/10.1101/2024.10.03.24314850doi: medRxiv preprint