🌌 VBASE3 • BCRbase

VBASE3 PaCMAP Manifold — Stratum 4 (2,048 Points Across 32 Benchmark Taxa)

1. Why PaCMAP Outperforms t-SNE and UMAP for Repertoires

Antibody repertoires operate simultaneously across three distinct biological scales: micro-scale somatic hypermutation (SHM) clouds differing by single amino acids; meso-scale V(D)J combinatorial usage and CDR3 loop diversity; and macro-scale ancestral framework constraints. While t-SNE destroys global phylogeny (causing arbitrary cluster scattering) and UMAP introduces artificial cluster boundaries, PaCMAP optimizes a balanced three-tier triplet loss: Loss = w_NB · Σ [d² / (10 + d²)] + w_MN · Σ [d² / (10000 + d²)] + w_FP · Σ [1 / (1 + d²)] • Nearest Neighbors (k-NN): Preserves fine clonal micro-structure and SHM variants.
• Mid-Near Pairs (MN): Preserves intermediate structural relatedness, grouping V families and ancestral Kabat/Dildrop clans.
• Further Pairs (FP): Preserves global repulsion across the embedding space, cleanly separating Heavy, Light (κ, λ, ρ, σ), and species lineages.

2. Complete Variable Regions & Paired Heavy-Light Chains

• Complete Rearranged Variable Region: VBASE3 embeds the full mature variable domain exon (FR1 → CDR1 → FR2 → CDR2 → FR3 → CDR3 → FR4), strictly terminating at canonical synthesis motifs (e.g. TVSS/TVSA for Heavy, VEIK/LEIK/TVL for Light) with zero constant-region contamination.
• Native Paired Chains (VH ⊕ VL): For single-cell BCR-seq, heavy and light chain vectors are concatenated to directly map complete paratope geometries, pairing preferences, and public clonotypes.

🎯 Future Horizon: Antigen Specificity via AIRR TSV Standards

When antibody specificity or target antigen is experimentally determined, this functional metadata can be seamlessly integrated into the PaCMAP manifold using AIRR Community TSV standards (fields: antigen, target_substrate, K_D, IC₅₀). This dual sequence–affinity manifold exposes:

• Convergent Paratopes Across Disparate V-Genes: Antibodies from completely distinct germline families converging onto the identical structural epitope.
• Public Neutralizing Footprints: Mapping shared immune responses across diverse patient cohorts against pathogens (e.g., SARS-CoV-2, HIV-1, Influenza).
• Off-Target & Polyreactivity Screening: Detecting biophysical clusters associated with non-specific binding prior to recombinant synthesis.

3. References & Precedents Across Biological Domains

1. Wang, Y., Huang, H., Rudin, C., & Shaposhnik, Y. (2021). Understanding How Dimension Reduction Tools Work: An Empirical Approach to Deciphering t-SNE, UMAP, TriMAP, and PaCMAP for Data Visualization. Journal of Machine Learning Research (JMLR), 22(201): 1–73.
2. Sommer, M., Frank, R., & Heinz, S. (2023). ProtSpace: Continuous visual embedding and interactive exploration of protein language model representations across multigene families (Kinases, GPCRs, Olfactory Receptors, Globins). Bioinformatics Advances, 3(1), vbad120.
3. Cabello-Aguilar, S., et al. (2022). TopoMetry: A framework for discovering the latent geometry of single-cell data and continuous immune trajectories. eLife, 11, e79294.