q-bio.QMOct 8, 2026
SaveCross-species representation learning aligns mouse and human neural dynamics and tracks clinical drug efficacy
Organizations: Exin Therapeutics, Inc., 2261 Market Street STE 22614, San Francisco, California 94114, United States
Abstract
Preclinical models poorly predict human drug efficacy, particularly in neurological disorders. Neural activity offers a uniquely rich source of translational information because it captures high-dimensional variation in nervous-system function that can be measured in both animals and humans. However, its high dimensionality makes it difficult to distinguish conserved disease-related features from variation arising from species, recording modality and experimental context. Here, we test whether shared neural dynamics can be identified directly from electrophysiology data by learning representations organized by biological state rather than species. We develop a dual-rule contrastive learning framework that aligns corresponding mouse and human states while preserving separation between distinct phenotypes. This framework recovered conserved sensory-response structure across species and, in epilepsy, resolved distinct relationships between three mouse models and heterogeneous human patient populations. When treated animals were projected into a frozen cross-species representation, drug-induced movement towards the human-aligned healthy state retrospectively tracked known clinical efficacy across ten model-drug combinations including a disease-specific detrimental effect. The framework also identified shared disease-associated neural dynamics between Fmr1-knockout mice and human 16p11.2 copy-number variant carriers despite differences in genetic aetiology and recording modality. Together, these findings show the potential of cross-species neural representation learning to map heterogeneous human disease onto experimentally tractable preclinical states and assess whether interventions restore human-relevant circuit function.
Figures & tables
Figure 1: Dual-rule contrastive neural representation learning for cross-species biomarker discovery. Schematic of the framework used to identify neural features that generalize between mice and humans.
Figure 2: Dual-rule contrastive learning recovers conserved sensory-response axes across humans and mice (A) Schematic of the experimental and computational framework. (B,C) Representative two-dimensional projections of the learned latent space showing baseline and stimulus-evoked activity in humans and mice for a visual paradigm (B, optic flow) and an auditory paradigm (C, ASSR). Large triangles and circles denote the corresponding human and mouse centroids. (D) Stimulus-versus-baseline response vectors for individual subjects across six sensory paradigms. Each point represents an individual subject projected onto the plane defined by the human- and mouse-derived mean response directions. Arrows indicate the species means. Cosine similarity and angular separation between the human and mouse mean directions are reported above each paradigm. (E) Joint embedding of auditory and visual stimulus responses in the coprojected mouse-human latent space. (F) Individual-subject projections onto the plane defined by the human- and mouse-derived mean sensory-response directions, pooled across sensory paradigms. Arrows indicate the species means. Cosine similarity and angular separation between the mean directions are shown above the plot. (G) Cross-species decoding of sensory state using an axis derived from one species and evaluated in held-out subjects from the other species. Points indicate balanced accuracy for individual destination subjects, and vertical bars indicate group summaries. The dashed line denotes chance performance (balanced accuracy = 0.5). Human = 9; mouse = 12.
Figure 3: Cross-species representation learning resolves conserved epilepsy signatures and model-specific patient heterogeneity. A, Schematic of the cross-species framework, in which human EEG from TEMPLE/TUH is aligned independently with three mechanistically distinct mouse epilepsy models (AY9944, 4-AP and PTZ). B, Schematic of the seizure-specific alignment objective, jointly organizing mouse and human EEG ictal periods to compare shared features. C, Heatmap of cross-species associations between the three mouse epilepsy models and clinically annotated human seizure types. Values indicate standardized model-specific associations, with positive and negative values shown in red and blue, respectively. D, Frequency-dependent changes in spectral power during ictal relative to non-ictal activity in PTZ, AY9944, and 4-AP mice. Pairwise Pearson correlation coefficients between the model-specific spectral profiles are shown above the plot. E, Unified cross-species disease space showing the relative positions of healthy and epileptic human EEG and the three mouse models, illustrating a conserved disease direction together with model-specific structure. F, Heatmap of patient-level affinities to AY9944, PTZ, and 4-AP across 96 human epilepsy patients. Each column represents an individual patient, and values indicate affinity relative to patient controls. G, Ternary plot of the relative affinities of individual human epilepsy patients to the three mouse models. Each point represents one patient, with the three vertices corresponding to AY9944, PTZ, and 4-AP.
Figure 5: Cross-species alignment of Fmr1 mouse electrophysiology with human 16p11.2 EEG. A, Schematic of the dual-rule contrastive learning framework aligning high-density electrophysiological recordings from Fmr1 knockout (KO) and wild-type (WT) mice with scalp EEG from individuals carrying 16p11.2 deletions or duplications and neurotypical controls. B, Participant-level alignment scores with the Fmr1 -associated disease representation across five independently trained models, grouped by human genotype. C, Distribution of mean alignment scores for neurotypical controls, 16p11.2 deletion carriers and duplication carriers. Alignment differed across groups (Kruskal-Wallis = 7.37, = 0.025). Points represent individual participants, and horizontal bars indicate group means. D, Individual participant alignment scores ranked by Fmr1 -associated alignment, grouped by human genotype. Dashed horizontal lines indicate group means (control, 0.43; deletion, +0.07; duplication, +0.30), with error bars indicating uncertainty in participant-level estimates.
| Task | Subjects tested | Projection axis | Mean projection | Sign | Sample size | P -value | FDR-adjusted P (BH) |
|---|---|---|---|---|---|---|---|
| Audio vs Visual | Human | Human | 0.604 | 9 | 0.03516 | 0.04280 | |
| Audio vs Visual | Human | Mouse | 0.606 | 9 | 0.03516 | 0.04280 | |
| Audio vs Visual | Mouse | Mouse | 0.815 | 12 | 0.0004883 | 0.002279 | |
| Audio vs Visual | Mouse | Human | 0.821 | 12 | 0.0004883 | 0.002279 | |
| ASSR | Human | Human | 0.703 | 9 | 0.01953 | 0.02734 | |
| ASSR | Human | Mouse | 0.658 | 9 | 0.01953 | 0.02734 |
Table 1: Subject-level validation of within- and cross-species sensory-response axes. Mean subject-level projections onto species-derived sensory-response axes for each sensory paradigm and for the pooled analysis. For each task, responses from human or mouse subjects were projected onto either the axis derived from the same species (own axis) or the axis independently derived from the other species (cross-species axis). Positive projection indicates that the subject-level stimulus response is oriented in the predicted direction of the corresponding sensory-response axis. The table reports the mean projection, projection sign, number of subjects ( ), raw P -value and Benjamini-Hochberg false-discovery-rate-adjusted P -value. Multiple-comparison correction was applied across all 28 reported subject-level tests. Human = 9; mouse = 12. Bold adjusted P -values indicate FDR 0.05
| Attribute | Mouse PTZ | Mouse 4-AP | Mouse AY9944 |
|---|---|---|---|
| Mouse contrast | Discharge vs duration-matched non-seizure (within animal) | Detected ictal vs non-seizure (within animal) | SWD vs the same detector’s positives in saline animals (between animal) |
| Human comparator | TUSZ, by seizure type vs healthy | TUSZ, by seizure type vs healthy | TUSZ, by seizure type vs healthy |
| Mouse N | 13 animals / 57 sessions (Control arm) | 23 animals | 32 untreated + 32 saline animals / 233 recordings |
| Spectral peak (Hz) | 3 | 29 | 3 |
| Gamma 30-70 Hz enhancement | -20.4 | +22.2 | -4.0 (suppressed) |
| Best human match | ABSZ +0.94 | TCSZ +0.64 | ABSZ +0.91 |
Table 2: Summary of cross-species EEG alignment across mouse seizure models and human epilepsy phenotypes. Comparison of mouse contrast definitions, human reference cohorts, sample sizes, spectral characteristics, human seizure-type matches, bidirectional transfer performance, and seizure-type specificity across PTZ, 4-AP, and AY9944 analyses.
| Unified epilepsy model | Fmr1 –16p11.2 model | |
|---|---|---|
| Data | ||
| Human units | TUEP: 96 patients with epilepsy, 82 without; each patient’s recordings ( 10) pooled into one unit | 14 controls, 25 deletion and 15 duplication carriers; carriers pooled as one class in training |
| Mouse units | PTZ, 13 mice: 0–30 min post-PTZ vs 0–30 min post-saline, same animal 4-AP, 14 mice: 0–60 min post-4-AP vs pre-injection epoch, same animal AY9944: 43 untreated or control-vector mice vs 32 saline controls | 53 wild-type, 17 Fmr1 -KO |
| Channels | Mouse: EEG1, EEG2 Human: 2 drawn at random per window from 19 referential 10–20 channels | 117 human and 32 mouse candidates; 97–117 and 25–32 remain after bad-channel rejection 1 ; 16 drawn at random per window in both species |
| Resampling and filtering | 250 Hz; 0.5–70 Hz band-pass (4th-order zero-phase Butterworth); no notch | Mouse: 30 kHz 1.5 kHz (FIR), notch at 60 Hz and harmonics, 1–95 Hz band-pass, 250 Hz Human: notches at 60–240 Hz, 1–95 Hz band-pass, 500 250 Hz Both: 56–64 Hz band-stop; all filters zero-phase, 4th-order Butterworth |
| Windows | 8 s (2,000 samples), evenly spaced; 400 per human recording, 240 per mouse epoch | 8 s, 240 per recording (step 2–8 s); a window is rejected if 25% of its channels are outliers 2 |
Table 3: Training and evaluation configuration for the two cross-species models. Dataset composition, preprocessing, architecture, loss terms with their weights, optimization, cross-validation and selection, and evaluation, for the unified epilepsy model and for the Fmr1 –16p11.2 model.
Figure 1: Clinical features associated with patient-specific alignment to mouse epilepsy models. A, Association between clinical metadata and each patient’s position in the cross-species latent space, quantified as affinity to a generic mouse epilepsy signature or to the AY9944, PTZ and 4-AP models. Colours indicate standardized effect size; points denote nominal significance ( 0.05; large points, 0.005). Variables directly reflecting overall treatment burden were excluded from the generic disease analysis where indicated. B, PTZ-specific affinity was positively associated with multiple measures of seizure and antiseizure-medication burden, with particularly consistent associations for benzodiazepine and GABAergic drug exposure. Points and horizontal bars show effect estimates and uncertainty for individual clinical variables. C, Across patients, PTZ-specific affinity increased with the number of distinct maintenance antiseizure medications (Spearman = 0.411, = 3.2 10 -5 , = 96). D, In held-out patients, benzodiazepine exposure was associated with greater PTZ-specific affinity (no benzodiazepine, = 59; benzodiazepine, = 37), consistent with the opposing modulation of GABA A receptor signaling by benzodiazepines and the PTZ seizure model. These analyses indicate that individual patient–model alignment captures clinically and pharmacologically meaningful heterogeneity rather than a uniform epilepsy axis.
Figure 2: Conventional EEG seizure readouts track human clinical efficacy worse than the cross-species biomarker. (A–E) Each panel plots one mouse EEG readout against frozen human efficacy (ordinal, 1 to 4) for the same 10 drug–model arms: 5 in 4-AP (triangles), 3 in AY9944 (circles) and 2 in PTZ (squares). Colours mark the drug, as in Figure 4 . Each readout was averaged per arm over treated animals and z-scored within model family (4-AP; AY9944 + PTZ). Signs are set so that higher values always mean greater predicted rescue. (A) Number of seizure episodes per mouse in the 30 min after the convulsant. Episodes were detected algorithmically as 4 above the animal’s own pre-drug baseline, lasting 2 s. (B) Fraction of 1 s epochs exceeding 4 of the animal’s own baseline. (C) Mean EEG amplitude shift relative to the animal’s own pre-drug baseline. (D) Mouse-only control: health-directed displacement computed from mouse recordings alone, with no human data. (E) Percentage of the 30 min epoch spent in detected seizures. Grey lines are least-squares fits, shown only as a guide. Insets give the pooled Spearman correlation and from a two-sided exact blocked permutation test within the experimental family (all 14,400 relabellings). They also give the within-family correlations. (F) Pooled Spearman correlation of each readout with human efficacy under this identical test, with the cross-species biomarker from Figure 4 shown for reference (blue). * 0.05. None of the EEG readouts reached significance. The best, seizure count, reached = +0.63 ( = 0.062). The cross-species biomarker reached = +0.87 ( = 0.0039). (G) Pooled Spearman correlation with human efficacy as a function of the number of animals per arm, computed on random subsets of animals within each arm. Agreement with human efficacy increases with group size for the cross-species biomarker but not for the conventional readouts.