Transferring Only the Gap-Protected Part of Mouse Perturbation Responses to Human Alzheimer's Cells

Most Alzheimer's drugs that work in mice fail in people, and part of the reason is that the same gene can be wired differently in the two species. This proposal borrows one idea from the physics of topological superconductors. In those materials, information stored in special "zero modes" survives local noise because it is separated from everything else by an energy gap. Here the same reasoning is applied to a learned model of how cells respond to genetic or drug perturbations. We fit a linear response operator on mouse microglia and neurons. We then transfer to human cells only the parts of that operator that are separated from the rest of its spectrum by a large gap. The rest, which matrix perturbation theory predicts should be scrambled by species divergence, is discarded.

The claim can be tested and it can fail. If species divergence acts like a bounded perturbation on the response operator, then subspaces with a large spectral gap should rotate little between mouse and human. Mouse-to-human transfer restricted to those subspaces should then beat both full-operator transfer and ordinary low-rank truncation. The application is the microglial inflammation axis and mitophagy in Alzheimer's disease. Both have been studied side by side in mice and in human cell systems, so we can check whether the "protected" subspace matches the biology that is known to be conserved.

Core Claim

Gene-Embedding Perturbation Operators for Zero-Shot and Transferable Prediction of Transcriptional Responses (DYNAMO) models each gene knockout as a linear operator acting in a lifted expression space. This is a Koopman representation, meaning a nonlinear system is described by linear dynamics on transformed features. DYNAMO showed that this structure transfers across human cell types better than an unconstrained network does. It did not test transfer across species. We supply the missing ingredient for crossing species: a rule for deciding which parts of an operator should be trusted after the crossing.

Let $A_m$ be the mouse operator. Let $A_h = A_m + E$ be the unknown human operator, where $E$ is the species divergence. Let $U_k$ be a singular subspace of $A_m$ that is separated from the remaining singular values by a gap $\delta_k$. Wedin's extension of the Davis–Kahan theorem bounds how far that subspace moves under the perturbation:

$$ \sin\Theta(U_k^{m}, U_k^{h}) \le \frac{\lVert E \rVert}{\delta_k} $$

Here $\Theta$ is the largest principal angle between the mouse and human subspaces.

This bound is classical linear algebra. Topological superconductivity contributes no new mathematics here. What it contributes is the design principle that a gapped subspace is robust and the bulk is not. The Kitaev-chain metamaterial study Topological computation by non-Abelian braiding in classical metamaterials showed this concretely: information stored in midgap boundary modes survived injected Gaussian disorder in the spring constants. The same paper also showed the principle's failure mode. When the gap closed during mode fusion, the stored information was lost. That failure mode carries over directly. Our predictions should degrade as $\delta_k$ shrinks, and we will test exactly that.

The new element is selecting the subspace by gap size rather than by variance or by ortholog conservation, and validating that selection against measured mouse and human perturbation responses.

How This Sits Against Prior Work

Several papers motivate the problem. Neurotoxic Microglial Activation via IFNγ‐Induced Nrf2 Reduction Exacerbating Alzheimer's Disease traced a signaling chain in which IFNγ lowers Nrf2 through Keap1, which raises reactive oxygen species, then NFκB, then neurotoxic microglial markers. It compared human iPSC mini-brains, 5XFAD mice and human post-mortem tissue, and found that some features were reproduced in mice while the spatial pattern of Nrf2-low microglia around tau was captured faithfully only in human systems. That split between conserved and non-conserved features is exactly what our method should recover.

Amelioration of Alzheimer's disease pathology by mitophagy inducers identified via machine learning and a cross-species workflow showed that mitophagy induction carries over from worms to mice. It also showed that a mismatch in drug delivery across the blood-brain barrier broke translation for Rhapontigenin. Histamine H3 Receptor as a target for alcohol use disorder: challenging the predictability of animal models for clinical translation in drug development documents an outright failure of mouse-to-human translation even though the drug target looked conserved.

Other papers constrain the design. geneRNIB: a living benchmark for gene regulatory network inference found that Pearson correlation and GRNBoost2 (gradient-boosted regression trees) beat deep foundation models and multiomic pipelines. Any added structure therefore has to beat simple baselines, not just other complex models.

Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN groups genes into "macrogenes" using protein language models. It reports that strict one-to-one orthology discards over 50% of genes across distant species, and it detected functional drift in the glaucoma gene MYOC. We reuse SATURN's macrogenes as the shared coordinate system for comparing mouse and human operators. SATURN itself only aligns static cell states. It does not model responses to perturbation.

The method itself comes from DYNAMO's operator learning. DYNAMO also reports a limitation we inherit: 22% of human genes lack network coverage, and zero-shot accuracy for those genes drops to ρΔ = 0.108.

Alz-QNet: A quantum regression network for studying Alzheimer's gene interactions is the closest work that combines quantum methods with Alzheimer's disease. It fits an 8-gene human network on a simulated quantum circuit and uses a symmetry constraint to halve the gate count. It is human-only and does no cross-species work. We do not use quantum hardware. In our view, the symmetry-as-regularizer idea is more useful here than the circuit.

Research Program

The work has three stages.

The first stage is a calibration on an existing benchmark. Before any cross-species data are collected, we check whether gap-selected transfer helps within human cells. We use the Replogle K562 to RPE1 transfer task from DYNAMO, treating the cell-type difference as the perturbation $E$. We fit DYNAMO-style operators on K562 and compute their singular spectra. We then transfer to RPE1 under three restrictions: the full operator, the top-$k$ components ranked by variance, and the components with the largest gaps.

We also measure directly whether the gap predicts subspace stability. To do this, we fit operators separately on K562 and RPE1 and correlate each subspace's gap with its principal angle across the two cell types. If no such relationship appears in this cleaner setting, the cross-species program stops here.

The second stage builds a paired cross-species perturbation set focused on Alzheimer's disease. No such paired set exists in the research file. On one side are human iPSC-derived microglia–neuron–astrocyte cultures in the mini-brain format of Kang et al. On the other are primary microglia and neurons from 5XFAD and 3xTg mice. Both are profiled by single-cell RNA sequencing under the same set of perturbations: IFNγ, Nrf2 overexpression, Keap1 knockdown, the antioxidant intervention used by Kang et al., Kaempferol, and CRISPR interference against roughly 50 to 100 genes. Those genes come from the Alzheimer's knowledge graph The Alzheimer's Knowledge Base: A Knowledge Graph for Alzheimer Disease Research and from Alz-QNet's subnetwork (PLD3, SREBF2, YY1 and others).

Both species are projected onto SATURN macrogenes. Untreated human entorhinal cortex nuclei from GSE138852 are used to check that the human cultures resemble real tissue at baseline.

The third stage fits operators in mouse and predicts human responses. The fitting uses mouse data only, with gaps chosen on held-out mouse perturbations. We then predict human responses to perturbations that were held out in both species.

The biological readout asks a separate question: do the gap-protected subspaces load on the Nrf2/Keap1/NFκB axis and on mitophagy genes, which prior work suggests are conserved? Conversely, are the components that the method discards enriched for the features Kang et al. found to be human-specific?

Baselines

The naive reference point assumes no divergence at all. The human response is predicted to equal the mouse response mapped through one-to-one orthologs. We also include a mean-shift baseline that predicts the average human response across all training perturbations. Every method must beat both on ρΔ, the Pearson correlation between predicted and observed expression change. This is DYNAMO's metric.

The simple associative baselines follow geneRNIB. We infer a human network with Pearson correlation and with GRNBoost2, then propagate each perturbation through it. These baselines get access to the same unperturbed human cells our method sees. geneRNIB found that these simple methods beat complex ones, so they are the real bar for whether added structure helps at all.

The strongest existing method is DYNAMO's full operator transferred across species, with SATURN macrogenes standing in for orthologs. On the calibration task, DYNAMO reports a K562 to RPE1 transfer correlation of 0.576, which retains 51% of its native performance, compared with 30% retention for an unconstrained DirectMLP. We rerun DYNAMO with the same embeddings, data splits, tuning budget and single-GPU compute as our variant. On the calibration task, beating it means a statistically reliable gain over our own reproduction of DYNAMO, evaluated across perturbation splits. For cross-species transfer there is no reported result on Lacuna for mouse-to-human perturbation operator transfer, so our reproduction of DYNAMO defines the target.

The key ablations hold rank fixed and change only how components are chosen. In one, we keep the same number of components ranked by variance. In another, we keep a random subspace of matched dimension. In a third, we select by gap after permuting gene identities within each macrogene, which tests whether the gain depends on real biology or on spectral artifacts. Gap selection must beat variance selection at matched rank. If it does not, the physics-inspired ingredient adds nothing.

Why not just…?

Why not just use low-rank truncation, which also discards noisy directions? This is the most serious threat. A large singular value and a large gap often coincide. The matched-rank ablation is designed to settle this. We expect gap selection to matter most when a weak but isolated component carries conserved biology. A small Nrf2 antioxidant program is a plausible example. If variance truncation does equally well across both stages, we will report that the topological framing reduces to standard truncation and drop it.

Why not just use Pearson or GRNBoost2 networks on human data directly, given geneRNIB's results? These methods describe associations and do not predict the effects of perturbations that were never measured. geneRNIB's benchmark is about inferring networks within one context, not about transfer. Even so, our benchmark gives these methods the same human baseline cells. If they match our method on held-out perturbations, the mouse data contribute nothing, and that would itself be a useful negative result for translational work.

Why not just fine-tune a single-cell foundation model on human data? DYNAMO reports that network-topology embeddings beat foundation-model and text embeddings by 19% on zero-shot prediction, and geneRNIB found that scPRINT trailed the simple baselines. With only tens of human perturbations, fine-tuning is likely to overfit. We include a fine-tuned scGPT arm with the same tuning budget so that this argument is settled by measurement rather than citation alone.

Why not just collect human perturbation data and skip mice? For a fixed panel of perturbations that is cheaper. The method is aimed at choosing which of many mouse findings deserve expensive human follow-up. Its value is measured by how much human data it saves, so we report accuracy as a function of the number of human perturbations used for training.

Decisive Checks

The idea survives only if three things hold. First, across subspaces, larger gaps must correlate with smaller mouse–human principal angles in the paired data. Second, gap-selected transfer must beat both DYNAMO full transfer and variance truncation at matched rank on ρΔ for held-out perturbations. Third, the retained subspace must be enriched for the IFNγ–Nrf2–NFκB and mitophagy programs that prior work identified as conserved. Failure on the first check falsifies the premise. Failure on the second alone means the selection rule is redundant with existing methods.

One limitation is fixed in advance. The method addresses divergence in gene regulation, not pharmacokinetic barriers. Rhapontigenin's failure to cross the blood-brain barrier, or differences in daily timing between nocturnal mice and diurnal humans, lie outside its scope.