Can cross-species neural representation learning align mouse and human brain dynamics?

Species variance hides the biology that matters

Preclinical models predict human drug efficacy poorly, especially for neurological disorders. The paper's framing is that neural activity is an unusually rich translational signal because it captures high-dimensional variation in nervous-system function measurable in both animals and humans — but that same dimensionality is the problem. Species, recording modality and experimental context each add variance, so a sufficiently expressive model can score high accuracy by exploiting features that separate groups within one species while carrying little information about the corresponding biological state in another.

The authors therefore treat translation as a representation problem rather than a classification problem. The target representation should keep neural features tied to the biological phenotype and suppress differences that arise only from species or recording conditions. They also reject full domain invariance: because species identity is entangled with genuine biological differences, forcing complete invariance could delete phenotype-relevant signal. The constraint has to specify which structure should be shared.

Two rules align states and keep phenotypes apart

The framework is a dual-rule supervised contrastive objective. Species-specific input modules map human and mouse measurements into a shared encoder that produces a latent representation. The first rule, alignment, pulls observations sharing a biological label together across species, with cross-species positive pairs upweighted. The second rule, separation, pushes biologically distinct states apart regardless of species. In the sensory model, a gradient-reversal species classifier on the shared latent adds regularization against species identity dominating the space.

Once trained on matched biological states, the space serves as a fixed reference: unseen perturbations, including drug treatments, can be projected into it to ask whether a manipulation moves neural activity along a conserved, human-relevant axis. The unified epilepsy model added private representations for acquisition-specific information and anti-collapse constraints based on latent variance, covariance and participation ratio. Checkpoints were selected on validation subjects only, and held-out patient-alignment and all drug-response metrics were excluded from model selection.

Sensory-response axes transfer across species

The proof of concept used matched audiovisual stimulation: nine healthy humans recorded with an eight-channel wet EEG cap at 250 Hz, and twelve mice with intracranial EEG in a virtual-reality arena, resampled to 250 Hz. The design deliberately preserved large differences in species, recording geometry and environment while holding the sensory manipulation constant.

After training, five of six paradigms showed strong cross-species directional alignment, measured as cosine similarity between independently estimated stimulus-versus-baseline response directions: ASSR 0.99 (8° separation), chirp 0.83 (34°), optic flow 1.00 (5°), checkerboard 0.97 (15°) and gratings 0.99 (6°). The oddball comparison produced no consistent cross-species axis, which the authors read as evidence that the framework does not impose alignment where no consistent shared response exists. Pooled across paradigms, the human- and mouse-derived sensory axes were nearly collinear (cosine similarity 1.00, 4°).

This was not only a group-centroid effect. Projecting individuals from one species onto an axis estimated independently in the other gave positive, significant projections after false-discovery-rate correction for ASSR, checkerboard, gratings and optic flow, and for the pooled auditory-versus-visual comparison. Mouse responses projected onto the human axis at roughly 0.64 for ASSR, 0.82 for checkerboards, 0.97 for gratings and 0.94 for optic flow, with human responses projecting positively onto mouse-derived axes. Axes learned in one species also decoded sensory state above chance in held-out subjects of the other.

Epilepsy models map onto distinct patient subsets

The harder test used human EEG from the Temple University Hospital datasets and three mechanistically distinct mouse epilepsy models: PTZ, which induces mixed myoclonic, clonic and tonic-clonic seizures through GABAA receptor antagonism; AY9944, which produces chronic atypical absence seizures with slow spike-and-wave discharges; and 4-AP, which induces predominantly convulsive seizures by blocking voltage-gated potassium channels. A separate spectral analysis compared ictal and non-ictal recordings from each model with seizure-type-annotated human EEG, matching event durations across species and summarizing power into 34 contiguous 2-Hz bins from 2 to 70 Hz.

The result the authors emphasize is not that one model wins. In the learned disease space, the three mouse models occupied distinguishable regions, and individual patients varied substantially in their affinity to each model. The paper argues this reframes preclinical validity: for heterogeneous disorders, the useful question is which patients, or which dimensions of their disease, a given model represents, rather than whether a model reproduces the disease as a whole.

Representation quality bounds how far this can be pushed. Balanced accuracy on held-out subjects was 0.62 [0.54, 0.70] for human epilepsy, 0.55 [0.41, 0.67] for AY9944, 0.99 for PTZ and 1.00 for 4-AP. The authors state plainly that the epilepsy representation showed modest human disease discrimination and that AY9944 was weakly separated from control in the unified space — cross-species alignment cannot recover biological information that is poorly represented in the source data.

Drug-induced displacement tracks clinical efficacy

The central translational claim concerns intervention, not classification. Treated animals were projected into the frozen cross-species representation, and efficacy was expressed as movement toward the human-aligned healthy state. Drug-induced movement along this axis retrospectively tracked known clinical efficacy across ten model-drug combinations, including a disease-specific detrimental effect. No drug-response data were used for training or recalibration.

The tiagabine result illustrates why direction matters. The same compound produced opposite effects in the PTZ model and the absence-like AY9944 model, which the authors attribute to the different disease dynamics in which the intervention acts. The latent space also described response in more than one dimension: movement along the conserved disease-to-healthy trajectory tracked therapeutic efficacy, whereas orthogonal displacement showed a weaker relationship with clinical secondary effects. The authors label that second result exploratory rather than established.

The benchmark itself limits the inference. The pharmacological analysis covered ten model-drug combinations and relied on retrospective clinical efficacy estimates rather than prospectively collected human responses, and the benchmark combined interventions evaluated across different epilepsy syndromes and clinical endpoints. This is retrospective tracking of known efficacy, not a prospective prediction of how an untested drug will perform in patients.

Convergence appears without shared genetic cause

The final test crossed both recording modality and disease aetiology: in-vivo electrophysiology from Fmr1 knockout mice aligned with scalp EEG from people carrying 16p11.2 copy-number variants and neurotypical controls. Across five independently trained models, mean alignment scores with the Fmr1-associated disease representation were −0.43 in controls, +0.07 in deletion carriers and +0.30 in duplication carriers, with a significant overall group difference (Kruskal-Wallis H = 7.37, P = 0.025).

Alignment was strongest in duplication carriers, which the authors connect to evidence that 16p11.2 duplication can produce impaired inhibitory transmission, cortical hyperexcitability and hypersynchronous network activity — features that converge with circuit abnormalities reported in Fragile X syndrome. Individual distributions overlapped broadly within the deletion and duplication groups, with some carriers showing little correspondence to the mouse phenotype and others strong alignment. The authors interpret this as a continuous dimension of neural dysfunction rather than a uniform signature attached to each genetic diagnosis.

Held-out balanced accuracy for this comparison was 0.59 [0.38, 0.78] for human CNV versus control and 0.80 [0.66, 0.92] for mouse knockout versus wild type, so the human side of the alignment is the weaker link. The claim supported here is that shared disease-associated dynamics can be recovered despite different genetic aetiologies and measurement geometries, not that the two conditions are equivalent.

The healthy-axis assumption sets the next test

The framework's practical value depends on a specific geometric assumption: that successful therapy moves neural activity toward an average healthy state. The authors acknowledge that effective interventions may instead produce compensatory states that restore function without reproducing untreated healthy dynamics, which would require richer representations of the healthy and disease manifolds rather than a single restorative axis. External validation across recording systems, additional disease models and substantially larger intervention panels is also needed.

Given those bounds, the paper's most decision-relevant contribution is a change in what preclinical evidence is asked to show. Instead of asking whether a drug rescues an animal phenotype, the framework asks which components of its effect move mouse neural dynamics along a human-anchored disease-to-health axis — and it does so with a representation frozen before any drug data are seen. The ten-combination retrospective panel, including the opposite-direction tiagabine result across PTZ and AY9944, is the concrete demonstration that this displacement carries information conventional within-species endpoints would not expose.

Sources used in this answer