Dataset DOI: 10.5061/dryad.fqz612k24

Preprint: https://www.medrxiv.org/content/10.1101/2024.10.03.24314850

Description of cohorts utilized in study

Please refer to our manuscript for detailed description of cohorts and data structure. Specifically, we refer you to Supplementary Table 1.

Cohort descriptions: 

1. Cohort I: cross-sectional cohort of term and preterm births
2. Cohort II: cross-sectional cohort of term and preterm births
3. Cohort III: longitudinal cohort of term births
4. Cohort IV: cross-sectional cohort of term and preterm births
5. Cohort V: cross-sectional cohort of term and preterm births
6. Cohort VI: longitudinal cohort of term births
7. Cohort VII: longitudinal cohort of term and preterm births
8. Cohort VIII: longitudinal cohort of term and preterm births
9. Healthy Baseline 1: cross-sectional cohort of non-pregnant individuals
10. Healthy Baseline 2: cross-sectional cohort of non-pregnant individuals
11. GFAP: immunoprecipitations with a positive control anti-GFAP antibody
12. Mock IP: immunoprecipitations without antibody

Items 1-8 in the list above, refer to serum samples from women prior to, during, or after pregnancy. Items 9-10 in the list above, refer to serum samples from men and non-pregnant women. Items 11-12 refer to technical controls of the PhIP-seq method.

Experimental groups:

Column "ptb" refers to the outcome of the study (1 - preterm birth; 0- term birth)

Pregnant versus non-pregnant comparisons include Cohorts I-VIII versus Healthy Baseline 1 and Healthy Baseline 2. Additionally some cohorts have pre-pregnancy samples, which are indicated by negative 'GA_test' values (see dictionary below).

Technical performance evaluation can be completed by comparing any of the human samples versus GFAP or Mock IP. 

Abbreviations

PTB preterm birth

IP  immunoprecipitation

PhIP-seq Phage display immunoprecipitation and sequencing

RPK reads per kilobase (our standard is to normalize to 100,000)

All other terms within metadata are defined below

Description of the data and file structure

I. PhIP-seq Data

Aligned RPK files. Note in some cases samples were technically replicated. The replicates are not averaged in these files:

File name structure:
Uncollapsed_RPK_{Cohort}_Samples.csv where Cohort == ["CohortII","CohortIII","CohortIV","CohortV","CohortVI","CohortVII","CohortVIII","HealthyBaseline1", "HealthyBaseline2", "GFAPantibody", "Mock_IP"]

Collapsed Aligned RPK files. The replicates are averaged in these files:

File name structure:
Collapsed_RPK_{Cohort}_Samples.csv where Cohort == ["CohortII","CohortIII","CohortIV","CohortV","CohortVI","CohortVII","CohortVIII","HealthyBaseline1", "HealthyBaseline2", "GFAPantibody", "Mock_IP"]

• Collapsed_RPK_CohortII_Samples.csv
• Collapsed_RPK_CohortIII_Samples.csv
• Collapsed_RPK_CohortIV_Samples.csv
• Collapsed_RPK_CohortV_Samples.csv
• Collapsed_RPK_CohortVI_Samples.csv
• Collapsed_RPK_CohortVII_Samples.csv
• Collapsed_RPK_CohortVIII_Samples.csv
• Collapsed_RPK_GFAP_antibody_Samples.csv
• Collapsed_RPK_HealthyBaseline1_Samples.csv
• Collapsed_RPK_HealthyBaseline2_Samples.csv
• Collapsed_RPK_Mock_IP_Samples.csv
• Uncollapsed_RPK_CohortII_Samples.csv
• Uncollapsed_RPK_CohortIII_Samples.csv
• Uncollapsed_RPK_CohortIV_Samples.csv
• Uncollapsed_RPK_CohortV_Samples.csv
• Uncollapsed_RPK_CohortVI_Samples.csv
• Uncollapsed_RPK_CohortVII_Samples.csv
• Uncollapsed_RPK_CohortVIII_Samples.csv
• Uncollapsed_RPK_GFAP_antibody_Samples.csv
• Uncollapsed_RPK_HealthyBaseline1_Samples.csv
• Uncollapsed_RPK_HealthyBaseline2_Samples.csv
• Uncollapsed_RPK_Mock_IP_Samples.csv

Metadata file includes outcome (described for each cohort). The identifier is "unique_sample_identifier" that links for RPK file columns. There are Uncollapsed (not averaged) and Collapsed metadata files for each cohort.

Metadata dictionary:

'GA_test': numerical; gestational age at test in weeks, decimal format. If negative, collection was performed before pregnancy. If 'GA_test' > 'GA_del' collection was performed postpartum. 

'GA_del': numerical; gestational age at delivery in weeks, decimal format.

'ptb': dummy; 1=preterm; GA_del<37, 0=term; GA_del>=37

'Maternal_Fetal': categorical; "Maternal" sample derived from maternal serum, "Fetal" sample derived from cord blood at time of delivery.

'nulligravida': dummy; For non-pregnant female subjects indicates whether they have been pregnant before. 1=never pregnant (nulligravida) 0=has had previous pregnancies. Not applicable for male subjects or pregnant subjects (left blank).

'non_preg_group': categorical; Grouping for non-pregnant subjects. "nulli": nulligravida (see above), "male": male subjects, "unk_f": unknown gravidity status female, "multi": previous pregnancy. Not applicable for pregnant subjects (left blank).

'barcode_pub': categorical; Individual identifier formatted as follows {Cohort}_sample{number}. Example identifier: "CohortII_sample0001". This indicates that the sample0001 was from CohortI. Samples that were technically replicated will have identical value. Longitudinal samples will also have an identical value. Refer to GA_test for timepoint information.

'unique_sample_identifier': categorical; Sample identifier formatted as follows {Cohort}sample{number}{unique_identifier}. This column contains unique values for every PhIP-seq experiment and corresponds to fastq filenames.

File name structure:
Uncollapsed_Metadata_{Cohort}_Samples.csv where Cohort == ["CohortII","CohortIII","CohortIV","CohortV","CohortVI","CohortVII","CohortVIII","HealthyBaseline1", "HealthyBaseline2", "GFAPantibody", "Mock_IP"]

Collapsed_Metadata_{Cohort}_Samples.csv where Cohort == ["CohortII","CohortIII","CohortIV","CohortV","CohortVI","CohortVII","CohortVIII","HealthyBaseline1", "HealthyBaseline2", "GFAPantibody", "Mock_IP"]

• Collapsed_Metadata_CohortII_Samples.csv

• Collapsed_Metadata_CohortIII_Samples.csv

• Collapsed_Metadata_CohortIV_Samples.csv

• Collapsed_Metadata_CohortV_Samples.csv

• Collapsed_Metadata_CohortVI_Samples.csv

• Collapsed_Metadata_CohortVII_Samples.csv

• Collapsed_Metadata_CohortVIII_Samples.csv

• Uncollapsed_Metadata_CohortII_Samples.csv

• Uncollapsed_Metadata_CohortIII_Samples.csv

• Uncollapsed_Metadata_CohortIV_Samples.csv

• Uncollapsed_Metadata_CohortV_Samples.csv

• Uncollapsed_Metadata_CohortVI_Samples.csv

• Uncollapsed_Metadata_CohortVII_Samples.csv

• Uncollapsed_Metadata_CohortVIII_Samples.csv

• Uncollapsed_Metadata_GFAP_antibody_Samples.csv

• Uncollapsed_Metadata_HealthyBaseline1_Samples.csv

• Uncollapsed_Metadata_HealthyBaseline2_Samples.csv

• Uncollapsed_Metadata_Mock_IP_Samples.csv

  Note: Collapsed metadata is not provided for "Collapsed_RPK_GFAP_antibody_Samples.csv", "Collapsed_RPK_HealthyBaseline1_Samples.csv", "Collapsed_RPK_HealthyBaseline2_Samples.csv", or "Collapsed_RPK_Mock_IP_Samples.csv" as these are technical or biological control samples. All required information can be obtained from the uncollapsed metadata file. 

II. IP-MS Data from Cohort II

Aligned Peptide Counts: Cohort2_IPMS_protein_abundances.csv, empty cells in this file indicate this protein was not detected. They were not infilled with NA or 0, as this is up to the user interpretation.

Metadata file includes preterm versus term outcomes: Cohort2_IPMS_metadata.csv

• First column corresponds to name of sample in Cohort2_IPMS_protein_abundances.csv file.

• The following metadata are provided: 

• 'GA_test': numerical; gestational age at test in weeks, decimal format

  'GA_del': numerical; gestational age at delivery in weeks

  'ptb': dummy; 1=preterm; GA_del<37, 0=term; GA_del>=37

  'name': text; corresponds to the sample ID found in the protein abundances chart.

Machine learning models to predict preterm outcomes

I. Logistic regression model to predict pregnant versus non-pregnant

• Pregnancy_model.zip

  Regularised logistic regression predicting pregnant (1) vs non-pregnant (0) serum samples from PhIP-seq peptide enrichment (adata.obsm["log_FC_over_AG"]) using the 3096 pregnancy-associated peptides.

• The zip directory contains the following files: 

  - final_classifier.joblib The fitted scikit-learn LogisticRegression (penalty=l1, solver=saga, C=1.0, max_iter=10000, random_state=666) trained on all labelled samples, with 593 of 3096 non-zero coefficients and intercept ≈ -0.0005; load with joblib.load.

  -feature_scaler.joblib — The fitted StandardScaler(with_mean=False) that divides each feature by its training-set standard deviation without centering, which must be applied to new data (columns in the order of model_features.csv) before calling the classifier.

  - model_features.csv — The ordered list of the 3096 peptide identifiers that define the input column order expected by both the scaler and the classifier.

  - coefficients.csv— The classifier coefficient for each peptide (column coef_preg_non), in the same order as model_features.csv; the intercept is not in this file and lives only in final_classifier.joblib.

  - cv_metrics.csv— Per-fold fit accuracy and ROC-AUC from 5-fold stratified cross-validation (mean ROC-AUC 0.936, mean accuracy 0.970).

  - ROC.pdf— ROC plot showing the mean ± SD cross-validated curve (blue), a label-shuffled baseline (red dashed), and the final model's pregnancy probability scored against the unrelated preterm-birth label ptb as a negative control (green).

  - requirements.txt— Pinned package versions of the environment that trained the model

II. Logistic regression model to preterm

• Preterm_model.zip
• Compressed directory contains the following files: 
  • model_stage1_elasticnet.joblib — The fitted scikit-learn LogisticRegression (penalty=elasticnet, solver=saga, C=0.18, l1_ratio=0.9, max_iter=10000, tol=1e-4, class_weight=balanced) over 12,072 features (12,071 peptides plus one gene), with 463 non-zero coefficients, saved under scikit-learn 1.5.2 and loaded with joblib.load.
  • scaler_maxabs.joblib — The fitted MaxAbsScaler (12,072 features, fit on the 2,607 training samples) that divides each feature by its training-set maximum absolute value and must be applied to new data before calling the classifier.
  • selected_peptide_indices.npy — An int64 array of 12,071 column positions in the 731,724-peptide library that defines which peptides, and in what order, form the first 12,071 model features.
  • selected_gene_indices.npy — A one-element int64 array (value 16481) giving the position of SPATA6 in all_gene_names.npy, the single gene-level feature that occupies the 12,072nd column.
  • all_gene_names.npy — An object array of 19,997 sorted gene symbols derived from the peptide-to-gene annotation, needed only to resolve selected_gene_indices.npy to a gene name; load with allow_pickle=True.
  • var_names.npy — An object array of all 731,724 peptide identifiers in library column order, used to align a recipient's hit matrix to the indices in selected_peptide_indices.npy; load with allow_pickle=True
  • requirements.txt — Pinned package versions of the environment that trained the model (Python 3.10.17, scikit-learn 1.5.2, numpy 2.2.4, scipy 1.15.2, joblib 1.4.2, threadpoolctl 3.6.0)
  • all_coefficients.csv — Table of ranked coefficients in the final model. 'rank': rank of feature in model by coefficient, 'feature_type': whether the feature was peptide or gene, 'feature_name': feature name corresponding to RPK file peptide name or respective gene, 'gene': gene from which the peptide was derived, 'coefficient': trained model coefficient, 'abs_coefficient': absolute value of 'coefficient', 'direction': whether the coefficient is more useful for predicting preterm ("PTB") or term ("Term")

Missing or NA values

With the exception of Cohort2_IPMS_protein_abundances.csv as defined above, NA values indicate that data was not available for the participant. 

Access to Cohort I Data

Individual-level data for Cohort I were obtained via agreement with the California State Committee for the Protection of Human Subjects which prohibits the depositing of any individual level data into any external data repository where it could be used by unapproved investigators even if that data could not be used to identify individual subjects. Requests for access to the Cohort I individual-level data should be directed to Laura L. Jelliffe-Pawlowski. Access is subject to approval by the California State Committee for the Protection of Human Subjects and completion of any required data use agreements, which restrict use of the data to approved investigators and approved research purposes. Requests will be reviewed and responded to within a reasonable timeframe following receipt of the required documentation.

Laura L. Jelliffe-Pawlowski

email: laura.jelliffe.pawlowski@nyu.edu

phone number: 212 998 9020

Code/software

PhIP-seq analysis code is available on GitHub https://github.com/h-s-miller/phagepy

Human subjects data

Specimens received by the researchers were de-indentified prior to analysis in the laboratory. Thus, specimen identifiers were never linked to any personally identifiable information (PII).

As a second layer of anonymization, data were further stripped of cohort-specific identifiers and a random label was assigned. Cohorts are anonymized and labeled as Cohort I-VIII which do not identify specimen origin. Sample_IDs and corresponding sequencing files have been likewise de-identified. For example "CohortI_sample0001" is a completely random sample name.

Limited patient metadata is made publicly available that does not allow identification of the subject (gestational age at test and gestational age at delivery).

Sequencing data provided contains sequences to our synthetic human proteome library only. No DNA sequences derived from patients were generated, nor are they deposited here.

Informed consent was obtained to publish de-identified data.