TL;DR
This paper introduces Latent-aligned Restricted Boltzmann Machines (LaRBMs) to uncover a shared, interpretable "latent space" of brain activity across individual zebrafish larvae. This allows for the translation of brain activity patterns between fish, demonstrating that fundamental building blocks of neural dynamics are conserved across individuals, even without direct neuron-to-neuron correspondence.
What Problem Does This Paper Address?
Understanding how brains organize their activity, especially spontaneous activity (the brain's "idle" state), is crucial for deciphering neural function. A major challenge is identifying conserved patterns of this activity across different individuals. In vertebrates, this is particularly difficult because there isn't a perfect one-to-one mapping of neurons between individuals, unlike in simpler organisms like C. elegans. Previous methods often relied on anatomical registration or stimulus-driven activity, which don't fully capture the richness of spontaneous dynamics or work well at the single-neuron level across individuals. The authors aim to bridge this gap by developing a method to find a common, interpretable representation of spontaneous brain activity that generalizes across individual brains.
Methods
The core innovation is Latent-aligned Restricted Boltzmann Machines (LaRBMs). This is an unsupervised, generative approach that builds upon Restricted Boltzmann Machines (RBMs).
- RBMs: These are energy-based models with visible units (representing neurons or voxels) and hidden units (representing latent "cell assemblies" or co-activation motifs). They learn to model the probability distribution of neural activity.
- LaRBMs: This is a "teacher-student" framework. A "teacher" RBM is trained on one fish's data to define a latent space. Then, "student" RBMs are trained on other fish, but with a constraint: they must align their latent space with the teacher's. This alignment is achieved by training the student RBM not only on its own data but also on hidden configurations sampled from the teacher RBM. This forces the student to adopt a similar representational structure.
- Voxelization: For a coarser, but more robust, analysis across individuals, the authors also used a voxel-based approach, averaging neuronal activity within 20µm cubic voxels.
- Cross-individual Translation: Once a shared latent space is established, the authors demonstrate the ability to encode activity from one fish into this latent space and then decode it back into the neural space of another fish.
Datasets
The study uses whole-brain calcium imaging data from larval zebrafish. Specifically, they recorded from six larval zebrafish (nacre Tg(elavl3:H2B-GCaMP6f)). The data was pre-processed to obtain binarized spike trains for individual neurons.
Key Insights and Ideas
- Spontaneous activity is highly stereotyped: Despite individual differences, the fundamental "building blocks" of spontaneous brain activity (cell assemblies) are remarkably conserved across zebrafish larvae.
- A shared latent space is key: By forcing RBMs to share a latent space, the authors can extract these conserved motifs. This latent space acts as a common language for brain activity.
- "Fictive translation" is possible: The ability to translate brain activity from one fish to another demonstrates that the learned latent representation captures genuine functional relationships, not just statistical artifacts. This is like teaching one brain to "speak" the activity language of another.
- Cell assemblies are the conserved units: The conserved patterns are not at the level of individual neurons but at the level of functional groups of co-activating neurons (cell assemblies), which are represented by the hidden units in the RBMs.
Key Results
- Robust latent space alignment: LaRBMs successfully aligned the latent spaces of different fish, with hidden unit correlations being more conserved across individuals than raw voxelized activity.
- Successful cross-individual translation: Translated activity patterns were assigned high probability by the recipient model and retained the original spatial organization. This was quantitatively shown by preserved statistical features (mean activity, pairwise covariances) and spatial correlations, often outperforming within-animal reconstructions.
- Conserved spatial patterns of cell assemblies: Hidden units in the LaRBMs retained their spatially localized patterns, allowing for reliable matching between teacher and student RBMs.
- Improved training efficiency: LaRBM students converged significantly faster (requiring about one-tenth the training steps) compared to unconstrained models.
Limitations and Future Work
- Mesoscopic spatial assumption: The current approach assumes a spatial organization of brain function, which might not generalize to systems with less stereotyped or more complex architectures (e.g., mammalian cortex).
- Teacher-fish bias: The method requires selecting a reference "teacher" fish, which could introduce bias. The authors suggest future work could mitigate this with joint training or multi-teacher constraints for a more symmetric approach.
- Generalization to other systems: While successful in zebrafish, extending this framework to other species or more complex brain structures is a future direction.
- Comparative phenotyping: The authors envision this framework being used for comparative phenotyping, identifying deviations from typical activity patterns that could indicate developmental, genetic, or pathological variations.
Summary updated