Antiferromagnets are crystals whose atomic spins point in alternating opposite directions, so the material has no net magnetization and leaves no stray magnetic field. A recent theoretical result shows something surprising about them: if you only see noisy samples from such a "repelling" system, you can reconstruct its interaction graph almost exactly in near-quadratic time, because forbidden local configurations act as sharp constraints that similarities never provide. Current methods for reconstructing the two great graphs of molecular biology — cell family trees and protein interaction networks — do the exact opposite: they rely on similarity, pulling together things that look alike. This proposal asks whether the antiferromagnetic principle transfers: whether mutually exclusive biological states, such as sibling cells forced into alternative fates or proteins that can join complex A or complex B but never both, are abundant and learnable enough to serve as reconstruction constraints that beat similarity-based inference.
The claim is testable end to end. Computationally, we will adapt the repulsion-based structure-learning algorithms to two public benchmarks where similarity methods have documented ceilings. Experimentally, we will build a small paired benchmark in mice in which a conditional gene knockout, a heritable lineage mark, and clone-restricted co-immunoprecipitation (the standard biochemical assay for physical protein contact) are triggered by the same recombination event, so that a lineage tree and a physical interactome are measured on the very same cells. This is what no existing study does: the closest works either combine knockout with co-IP but no lineage record, or record lineage without physical interaction data.
Core Claim
Lineage trees and protein interaction networks are currently inferred with homophilic bias. Reconstructing Cell Lineage Trees from Phenotypic Features with Metric Learning builds cell family trees by learning a distance in which related cells map close together; PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs shows that even large sequence models, which implicitly encode "similar sequences interact," fail to reconstruct whole interaction networks. We claim the complementary signal is being wasted: exclusivity. Two daughter cells of one progenitor often adopt mutually exclusive fates (lateral inhibition and binary commitment make siblings dissimilar precisely because they are close in the tree); a protein recruited into one holoenzyme is excluded from a competing one; a nuclear factor and a cytosolic enzyme never meet in a genuine complex. These are "repelling" pairs in exactly the sense of Structure learning of antiferromagnetic Ising models, which proves that repelling models, unlike general attractive ones, evade a $\Omega(p^{d/2})$ statistical-query lower bound and admit reconstruction in $O(p^2)$ time in the strong-repulsion limit — but never applies those algorithms to empirical biological data.
Formally, given $p$ nodes (cells or proteins) and a repulsion graph $R$ of pairs believed mutually exclusive, we seek a tree or network $G$ such that
$$(i,j) \in R \;\Rightarrow\; d_G(i,j) \ge D,$$
where $d_G$ is path length in the reconstructed graph and $D$ is a problem-dependent exclusion distance. The repulsion graph is itself learned: from fate-marker exclusivity in single-cell data, from compartment annotations, and from never-co-captured pairs across co-IP pull-downs.
Research Program
Three components. First, an algorithmic transfer: we implement the hard-core and strong-repelling reconstruction algorithms from the antiferromagnetic Ising paper as graph priors, then couple them to biological observation models — phenotype noise for cells, pull-down spectral counts for proteins. The open theoretical question, which the source paper does not address, is whether the $O(p^2)$ guarantee survives realistic noise. CRISPR barcode data are corrupted by "double-resection" events that delete adjacent target sites and destroy phylogenetic information (ConvexML: Fast and accurate branch length estimation under irreversible mutation models), so we will characterize the phase boundary between repulsion strength and noise level at which exact reconstruction breaks, using simulations before touching real data.
Second, benchmark evaluation. On lineage trees, we rerun the full pipeline of CellTreeQM — its C. elegans developmental atlas and CRISPR-barcoded mouse embryonic stem cell datasets, its Robinson-Foulds evaluation — with repulsion constraints replacing or augmenting the quartet metric loss. On interaction networks, we use the PRING benchmark of 21,484 proteins and 186,818 interactions across four species, adding exclusive-complex and compartment-disjointness repulsion edges to the reconstruction objective and measuring network-level Graph Similarity, including cross-species transfer, where current models degrade badly.
Third, the paired wet-lab benchmark, which also tests the biology. We will use a triple mouse line — $MED1^{fl/fl}$ crossed to $Lyz2$-Cre and a Cre-dependent $Rosa26$ reporter with a Cre-activated epitope-tagged bait — so that one tamoxifen dose simultaneously deletes MED1 in myeloid cells, permanently marks those cells and all descendants, and activates an affinity tag on a bait protein in the same cells. This exploits the fact that the Cre-LoxP architecture already unifies conditional knockout and lineage marking (Genetic Lineage Tracing Summary); we add the third element, clone-restricted pull-down, to solve the problem that standard co-IP averages over millions of cells and cannot resolve rare marked subclones (Validating Physical Protein Interactions Summary). The model system is deliberate: MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia via ROS Generation and Inflammation showed that macrophage-specific MED1 loss accelerates arterial lesion growth and named PGC1$\alpha$ antioxidant signaling as the pathway, but explicitly left the physical MED1–PGC1$\alpha$ complex unresolved and never distinguished resident from recruited macrophages — both gaps our design closes. After carotid ligation, we sort reporter-positive (knockout, lineage-marked) and reporter-negative macrophages, run co-IP–mass spectrometry on the tagged bait separately for each, and obtain ground-truth interactome remodeling tied to a known lineage.
Proposed Experiments
Experiment one is the simulation study: synthetic repulsion graphs of $10^3$–$10^5$ nodes under controlled noise (dropout, barcode double-resections, spurious pull-down edges), measuring reconstruction error and runtime against the theoretical predictions of $O(p^2 \log p)$ and $O(p^2)$. Experiment two is the lineage benchmark: same datasets, splits, and feature budgets as CellTreeQM, metric is relative improvement in Robinson-Foulds distance over raw phenotypic distance; the number to beat is 75.7%, which CellTreeQM with Gumbel-Softmax feature gating reports on the C. elegans Small dataset. Experiment three is the interactome benchmark: PRING's four species, human-trained and transferred, metric is Graph Similarity; the number to beat is the best reported model, PLM-interact at 650M parameters, which stays below 0.5 on human and drops 15.2% to Arabidopsis, 25.3% to yeast, and 35.2% to E. coli. Experiment four is the mouse study: predict, from the computational model with repulsion priors, which complexes remodel in knockout-marked macrophages (PGC1$\alpha$-associated loss is the positive control suggested by the MED1 paper, which reports no quantitative cytokine or ROS values on Lacuna) and which are unchanged, then validate a ranked subset by targeted co-IP in the style of NAT10 Promotes Tubular Epithelial Cell Senescence in Cisplatin-Induced Acute Kidney Injury by Regulating DDX17, which used exactly this targeted physical-binding confirmation for NAT10–DDX17 but, crucially, in bulk cell culture with no lineage record.
Baselines
The naive reference point is doing no inference: raw phenotypic distance for tree reconstruction (the zero point against which CellTreeQM's relative improvements are defined) and unpruned co-IP–MS pull-down lists for interactions, where post-lysis mixing of compartments generates false positives by construction. The strongest existing competitor on lineage trees is CellTreeQM itself, which reports a 75.7% relative improvement in Robinson-Foulds distance on the C. elegans Small benchmark and already embodies the best structural insight available — the phylogenetic four-point condition — so beating it requires more than adding any constraint; we will rerun it on identical data, splits, and feature budgets, and give our method the same tuning budget, because an unfair comparison against a method this close in spirit would prove nothing. On interaction networks the strongest competitor is the PLM-interact 650M model as evaluated in PRING, with its documented human Graph Similarity ceiling below 0.5; we rerun it under PRING's original protocol and evaluate our repulsion-augmented reconstruction on the same graphs. For branch lengths once a topology is fixed, ConvexML — which reports being 10–100× faster than TiDeTree and LAML at chronogram estimation — is both a baseline and a component we may reuse, since it solves a problem orthogonal to ours. Finally, the key ablation isolates the proposal's central ingredient: the same architecture with repulsion edges removed (an attractive-only variant, which should fall back toward the quadruplet-loss level of 45.4% relative improvement) and with repulsion edges replaced by random pairs of equal count, which distinguishes genuine exclusivity from the mere presence of extra constraints.
Why not just…?
Why not just use pseudotime trajectories and siRNA knockdowns? Simpler, faster, and a real threat for the computational parts — but the file documents where it fails. Static transcriptomic snapshots mistake compensatory hypertrophy for regeneration (podocytes enlarge without dividing) and lose track of fibroblasts once they downregulate their markers, while the lineage mark is permanent (Genetic Lineage Tracing Summary). And global knockdown of a pleiotropic gene confounds cell-autonomous effects with systemic ones: TLR signaling adaptor protein MyD88 in primary sensory neurons contributes to persistent inflammatory and neuropathic pain and neuroinflammation could only attribute chronic pain to nociceptors because deletion was restricted to them, dropping MyD88-positive sensory neurons from about 52% to about 24%; a systemic inhibitor would have suppressed immunity everywhere. Our paired benchmark exists precisely because no cheaper method yields a lineage-resolved interactome on the same cells.
Why not just predict interactions with protein language models? For pairwise prediction this is competitive and cheap, which is why it is our strongest baseline rather than something we dismiss. But PRING shows the ceiling: the best 650M-parameter sequence model reconstructs human interaction networks with Graph Similarity below 0.5 and loses 15.2–35.2% when transferred across species, and PEER: A Comprehensive and Multi-Task Benchmark for Protein Sequence Understanding, while confirming that ESM-1b leads on 13 of 17 static sequence tasks, evaluates exactly the context-free setting that misses phosphorylation- and cell-type-dependent binding — the very remodeling our mouse experiment measures. Our PRING experiment will settle this directly: if sequence models plus repulsion priors do not beat sequence models alone on network-level Graph Similarity, the claim is wrong and we will say so.
Why not just use standard metric learning? The file quantifies the gap: triplet loss achieves only 17.9% relative improvement in Robinson-Foulds distance over raw data and quadruplet loss 45.4%, versus 75.7% for CellTreeQM with the four-point condition, on the same C. elegans benchmark. Structural inductive bias demonstrably matters. The honest residual risk is that the four-point condition already captures most of what repulsion adds; our random-repulsion ablation is designed to detect exactly this, and if real exclusivity contributes nothing beyond the quartet structure, we will report that as the finding.
Why not just CRISPR barcodes and ConvexML? Barcoding gives excellent trees, and ConvexML turns them into time-resolved chronograms fast. But barcodes carry their own corruption (double-resections deleting adjacent targets), and they give a tree without any physical interaction information on the same cells. We treat ConvexML as a collaborator, not a rival: it can supply branch lengths for the topologies we reconstruct.
Why involve antiferromagnets at all? A fair challenge, and we answer it honestly: we propose no antiferromagnetic material in living tissue. The documented constraints make that premature — the optically switched metastable state in CaMn$_2$Bi$_2$ lasts only 150+ picoseconds at about 200 $\mu$J/cm² pump fluence (Light-induced reorientation transition in an antiferromagnetic semiconductor), the NiO heterostructure's 33% Hall modulation requires films of roughly 1.2 nm (Electrical modulation and probing of antiferromagnetism in hybrid multiferroic heterostructures), and epitaxial perfection is mandatory. What transfers today is the mathematics: the proof that repelling systems evade a $\Omega(p^{d/2})$ lower bound and reconstruct in $O(p^2)$. The zero-stray-field robustness of antiferromagnets — memory states inert under 9 Tesla sweeps in FeRh (Room-temperature antiferromagnetic memory resistor) — motivates a future biosensing direction, but this proposal stakes its claims only on the algorithmic transfer.
Decisive Checks
The proposal succeeds if repulsion-augmented reconstruction beats 75.7% relative Robinson-Foulds improvement on the CellTreeQM benchmark, pushes human network Graph Similarity above the 0.5 ceiling on PRING while shrinking the 15.2–35.2% cross-species degradation, and if a majority of its ranked interactome-remodeling predictions in knockout-marked macrophages are confirmed by targeted co-IP. It fails informatively if the random-repulsion ablation matches the real one — that would show exclusivity carries no signal beyond ordinary structural constraints, a negative result worth publishing given how heavily current methods lean the other way.
References
- Reconstructing Cell Lineage Trees from Phenotypic Features with Metric Learning
- PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs
- Structure learning of antiferromagnetic Ising models
- ConvexML: Fast and accurate branch length estimation under irreversible mutation models, illustrated through applications to CRISPR/Cas9-based lineage tracing
- Viral and Transgenic Reporters and Genetic Analysis of Adult Neurogenesis
- Proteomics Analysis of Cellular Proteins Co-Immunoprecipitated with Nucleoprotein of Influenza A Virus (H7N9)
- MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia via ROS Generation and Inflammation
- NAT10 Promotes Tubular Epithelial Cell Senescence in Cisplatin-Induced Acute Kidney Injury by Regulating DDX17
- mTOR-mediated podocyte hypertrophy regulates glomerular integrity in mice and humans
- TLR signaling adaptor protein MyD88 in primary sensory neurons contributes to persistent inflammatory and neuropathic pain and neuroinflammation
- PEER: A Comprehensive and Multi-Task Benchmark for Protein Sequence Understanding
- Light-Induced Reorientation Transition in an Antiferromagnetic Semiconductor
- Electrical modulation and probing of antiferromagnetism in hybrid multiferroic heterostructures
- Room-temperature antiferromagnetic memory resistor