Conserved heavy/light contacts and germline preferences revealed by a large-scale analysis of natively paired human antibody sequences and structural data.
Paweł DudziciD, Dawid ChomicziD, Weronika BielskaiD, Igor JaszczyszyniD, Michał ZielińskiiD, Bartosz JanusziD, Sonia WróbeliD, Marguerite-Marie Le PannéreriD, Andrew PhillipsiD, Ponraj PrabakaraniD, Sandeep KumariD, Konrad KrawczykiD
Abstract
Antibody next-generation sequencing (NGS) datasets have become crucial to develop computational models addressing this successful class of therapeutics. Although antibodies are composed of both heavy and light chains, most NGS sequencing depositions provide them in unpaired form, reducing their utility. Here we introduce PairedAbNGS, a novel database with paired heavy/light antibody chains. To the best of our knowledge, this is the largest resource for paired natural antibody sequences with 58 bioprojects and over 14 million assembled productive sequences. We make the database accessible at http://naturalantibody.com/paired-ngs as a valuable tool for biological and machine-learning applications. Using this dataset, we investigated heavy and light chain variable (V) gene pairing preferences and found significant biases beyond gene usage frequencies, possibly due to receptor editing favoring less autoreactive combinations. Analyzing the available antibody structures from the Protein Data Bank, we studied conserved contact residues between heavy and light chains, particularly interactions between the CDR3 region of one chain and the FWR2 region of the opposite chain. Examination of amino acid pairs at key contact sites revealed significant deviations of amino acids distributions compared to random pairings, in the heavy chains CDR3 region contacting the opposite chain, indicating specific interactions might be crucial for proper chain pairing. This observation is further reinforced by preferential IGHV-IGLJ and IGLV-IGHJ pairing preferences. We hope that both our resources and the findings would contribute to improving the engineering of biological drugs.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 15
- FWCI
- 8.51
- Field percentile
- top 1%
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (30 Dec 2024) · CC BY 4.0
PairedAbNGS pools 58 single-cell bioprojects of natural paired human and mouse antibodies, revealing heavy/light V gene pairing biases beyond usage frequencies and conserved CDR3-FWR2 inter-chain contacts.
Public antibody sequencing data are mostly unpaired, and the existing paired resource OAS has limited diversity, with a single study (Jaffe et al.) accounting for most of it. This work assembles a larger, more diverse paired set to test whether germline-level and residue-level heavy/light pairing preferences exist.
Key results
ShownThe curated paired database spans many bioprojects and millions of chains, mostly human, with light chain ratios matching expected kappa/lambda consensus in human but not mouse.
58 studies, 2482 single-cell experiments, 14,401,268 productive nucleotide sequences, 4,280,331 unique amino acid sequences; human about 90% of sequences; kappa:lambda about 60%:40% in human and 90%:10% in mouse, a skew the authors say could stem from PRJNA1082281. · Overview of database curation; Database statistics
ShownHuman heavy/light V gene subgroup pairing departs from what usage frequencies alone predict, and germline usage differs strongly from the OAS paired database. The receptor-editing explanation is inferred, not tested.
Chi-squared on V gene contingency tables: p << 0.01 for kappa and lambda; pairwise tests with Bonferroni correction stay significant. Versus OAS: chi-square 138,899 (df 55) heavy V genes, 19,965 (df 33) kappa, 17,588 (df 39) lambda. · Germline encoded pairing preferences; Germline-encoded pairing preferences; OAS database preprocessing
ShownIn PDB antibody structures, the CDR3 of each chain contacts FWR2 of the opposite chain. Key interface positions overlap little with Vernier positions, and their amino acid distributions are broadly similar across datasets but statistically distinct.
Contacts defined at 4.5 Å heavy-atom distance. Total Variation Distance of PairedAbNGS vs PDB, therapeutics and OAS mostly below 0.2; chi-square rejected equality at most positions, and no IMGT position matched OAS. · Structural analysis of contact residues on the…; IMGT key interface residues
SuggestedAmino acid pairs at contacting heavy/light positions deviate modestly from random pairing, mainly via heavy-chain CDR3. Inter-chain IGHV-IGLJ and IGLV-IGHJ biases are also reported, though IGHJ4 bias may reflect recombination signal sequences.
Total Variation Distance of residue pairs vs random chain pairing reached up to 0.06, with most significant differences attributed to heavy-chain CDR3; inter-chain V-J pairing bias with p ~ 0. · Structural analysis of contact residues on the…; Discussion
SuggestedA Chao2 richness estimate puts the unique paired antibody count far below theoretical diversity. The authors call it informative but not conclusive and say it could be a lower bound.
Sobserved 339,192, q1 3,390,785, q2 1139 gives about 5 billion unique antibodies, versus theoretical estimates of 10^16 to 10^18 cited by the authors. · Estimating the total number of antibody pairs
Methods, data and reproducibility
Gemma-assisted screening of SRA single-cell projects, TRUST4 assembly and RIOT annotation against OGRDB; chi-square tests on V gene contingency tables; PDB contact extraction at 4.5 Å with Total Variation Distance comparisons; Chao2 richness estimate.
- PairedAbNGS: 10x Chromium 5' single-cell antibody bioprojects from NCBI SRA (human and mouse) · human and mouse (about 90% human) · n = 58 studies, 2482 experiments, 14,401,268 productive sequences
- Mouse extrafollicular plasmablast bioproject with elevated lambda chain share · mouse · PRJNA1082281
- Antibody structures from RCSB Protein Data Bank used to define VH/VL contact residues
- OAS paired database, reprocessed with RIOT as a comparator for germline and residue distributions
- Manually collected therapeutic monoclonal antibodies (standard formats only)
- Methods & tools
- Gemma language model, sra-toolkit, TRUST4, RIOT, OGRDB, IMGT numbering, Chao2 estimator, Chi-squared test with Bonferroni correction, Total Variation Distance, OAS paired antibody database, RCSB Protein Data Bank
- Data & accessions
- https://naturalantibody.com/paired-ab-ngs
- Reproducibility gaps
- Database is licensed for non-commercial use by non-commercial organizations only. No code repository, software versions or compute details are given in the text.
Limitations (6)
- The Chao2 estimate may be a lower bound and is affected by a small sampling ratio; authors call it informative but not conclusive and expect intrinsic dataset biases to shrink as more paired data arrive. · stated by the authors
- Structural analysis uses static PDB structures, whereas CDR3 loops are flexible and length depends on V and J usage; molecular dynamics ensembles would be needed to test pairing and contact effects. · stated by the authors
- Random pairing is a poor proxy for incorrect pairing because shuffled pairs sometimes recreate real ones, so residue-pair differences from random expectation are small. · stated by the authors
- Many sequences have no assigned B-cell type, disease or constant region label, which restricts subgroup-level interpretation; the authors only flag possible difficulty in annotating cell types. · noted in this summary
- Germline pairing analysis covers human projects only, because mouse OGRDB germlines lack a subgroup hierarchy; mouse pairing preferences were not assessed. · stated by the authors
- Receptor editing is proposed to explain pairing biases, but no autoreactivity or binding data were measured. Pairing errors from droplet-based assembly (TRUST4 on top-coverage chains) were not assessed. · noted in this summary
Who should read:Read this if you build or train antibody language models and need paired heavy/light data, or evidence for germline and interface-residue pairing constraints in natural repertoires.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Deep generative modeling reveals maturation-linked pairing signatures in human antibodies
Lea Brönnimann, Thomas Lemmin, Chiara Rodella·bioRxiv·Aug 24, 2025
- Looking for Therapeutic Antibodies in Next Generation Sequencing Repositories
Konrad Krawczyk, Matthew I. J. Raybould, … Charlotte Mary Deane·mAbsPeer-reviewed·Jun 20, 2019
- NAStructuralDB : Structural database to facilitate computational studies of molecular modeling and recognition of proteins with special focus on antibody-antigen interactions.
Dawid Chomicz, Paweł Dudzic, … Konrad Krawczyk·mAbsPeer-reviewed·Mar 8, 2026
- Antibody-Antigen Affinity Prediction with Chain-Aware Protein Language Modeling
Harshit Singh, Aastha Malhotra, … Rohan Gorantla·bioRxiv·Oct 3, 2026
- ImmunoMatch learns and predicts cognate pairing of heavy and light immunoglobulin chains
Dongjun Guo, Deborah K. Dunn‐Walters, … Joseph Chi-Fung Ng·Nature MethodsPeer-reviewed·Nov 18, 2025
- LICHEN: Light-chain Immunoglobulin sequence generation Conditioned on the Heavy chain and Experimental Needs
Henriette Capel, Isaac Ellmen, … Charlotte Mary Deane·bioRxiv·Aug 6, 2025