Accurate identification of communication between multiple interacting neural populations

TL;DR

This work tackles the challenge of identifying how multiple neural populations communicate across distributed brain regions using simultaneous recordings. It introduces MR-LFADS, a sequential variational autoencoder in which each region is modeled by a region-specific dynamical generator and inter-regional communication and inputs from unobserved areas are disentangled into separate latent streams. Through extensive synthetic benchmarks and real electrophysiology in mice, MR-LFADS outperforms static and dynamic baselines in recovering both the pathways (effectomes) and content of inter-regional communication, and it successfully predicted brain-wide effects of perturbations held out during training. The approach rests on three design pillars—automatic inference of unobserved inputs, data-constrained communication anchored to observed firing rates, and structured KL bottlenecks to prevent misattribution—yielding robust, interpretable inferences about distributed brain information processing with potential utility for guiding causal perturbations.

Abstract

Neural recording technologies now enable simultaneous recording of population activity across many brain regions, motivating the development of data-driven models of communication between brain regions. However, existing models can struggle to disentangle the sources that influence recorded neural populations, leading to inaccurate portraits of inter-regional communication. Here, we introduce Multi-Region Latent Factor Analysis via Dynamical Systems (MR-LFADS), a sequential variational autoencoder designed to disentangle inter-regional communication, inputs from unobserved regions, and local neural population dynamics. We show that MR-LFADS outperforms existing approaches at identifying communication across dozens of simulations of task-trained multi-region networks. When applied to large-scale electrophysiology, MR-LFADS predicts brain-wide effects of circuit perturbations that were held out during model fitting. These validations on synthetic and real neural data position MR-LFADS as a promising tool for discovering principles of brain-wide information processing.

Figures (10)

• Figure 1: MR-LFADS architecture. (a) Single-region LFADS, as adapted for this work. (b) MR-LFADS with $N$ regions. KL penalties from SR-LFADS in panel (a) are included in MR-LFADS, but are omitted in the diagram for clarity.
• Figure 2: Experiment 1. (a) Left: DGN setup. Each region, area $A^i$, receives a private stimulus $s_{t}^{i}$ (red) and communicates a two-step delayed version $s_{t-2}^{i}$ (blue) to a downstream region. Each region is trained to recall (green) the last five time steps of its private stimulus and its received communication. Right: Potential incorrect communication model capable of accurately reconstructing these synthetic neural data. (b) $R^2$ scores for data reconstruction. Box plots describe distributions of values across $10$ models fit from distinct random initializations (seeds). Boxes represent the interquartile range (IQR), and whiskers extend to the most extreme points within $1.5$ IQR from the quartiles. (c) Left: Cosine similarity between inferred effectomes and ground truth. Right: Example fitted models. Color intensity indicates the relative message norm, computed by concatenating all multidimensional messages across trials and time, taking the 2-norm, then normalizing across communication channels. $s^{1:3}$ indicates the ground truth input. (d) $R^2$ of linear prediction of ground truth stimulus inputs (with time lag $\in \{0, 2\}$) from inferred inputs. (e) $R^2$ of linear prediction of ground truth messages (with time lag $\in \{2, 4\}$) from inferred messages.
• Figure 3: Experiment 2. (a) Left: DGN setup, with stimulus (red), communication (blue), trained readouts (green), and computations (I: identity; $\int$: integration). Right: Potential failure mode for a learned model. (b) $R^2$ scores for data reconstruction. (c) Left: Cosine similarity between inferred effectomes and ground truth. Right: Example fitted models. Color intensity indicates relative message norm. (d) $R^2$ of linear prediction of ground truth input $s$ to region $P$, from inferred inputs. (e) Left:$R^2$ when predicting ground truth messages $m_{}^{P \rightarrow D}$$=$$s_{}^{}$ from inferred messages, $\mu_m^{P \rightarrow D}$. Right:$R^2$ when predicting the decision variable $d_{}$ from $\mu_m^{P \rightarrow D}$, indicating mislocalization of integration.
• Figure 4: Experiment 3. (a) Top: Each DGN is configured with a random number of areas ($3$ or $4$), random inter-regional connectivity, and is trained on a randomly selected task. Bottom: Example task setup. On each trial, the DGNs received a noisy fixation stimulus $s_{\text{fix}, t}^{}$ and two noisy task stimuli, represented in polar coordinates as $s_{t}^{k} = (a_t^{(k)}, \theta_t^{(k)})$ for $k \in \{1, 2\}$. DGNs were trained to process these inputs into task-dependent outputs: a response angle $\theta^\text{resp}$ and a task-dependent saccadic eye movement. Here, multisensory decision-making is depicted as an example task. When the fixation cue disappears, the output area $A^N$ must saccade and report the $\theta^{(i)}$ value corresponding to the stimulus with larger $a^{(i)}$. (b) $R^2$ scores for data reconstruction. Box plots describe distributions of values across $35$ DGN fits. (c) Cosine similarity ($S_\text{cos}$) of inferred effectomes relative to ground truth, compared to that of MR-LFADS(R), with $\Delta S_\text{cos} = S_\text{cos}^\text{model} - S_\text{cos}^\text{MR-LFADS(R)\xspace}$. One-tailed t-test p-values: $p = 0.00016$, $0.12$, $0.006$, $0.0$, $0.01$. NS: not significant. (d) Left:$R^2$ scores, relative to MR-LFADS(R), for linear prediction of ground truth messages from inferred messages. One-tailed t-test p-values: $p = 0.0$, $0.0$, $0.001$, $0.0$, $0.14$. Right:$R^2$ scores, relative to MR-LFADS(R), for linear prediction of inferred messages from ground truth messages. One-tailed t-test p-values: $p = 0.0004$, $0.0$, $0.08$, $0.0$, $0.0$. (e) Inferred effectomes from models fit to datasets from a three-region DGN (top) and a four-region DGN (bottom). We chose the datasets on which MR-LFADS(R) achieved its median $S_\text{cos}$ scores (across the 35 generated datasets). Color intensity indicates relative message norm.
• Figure 5: MR-LFADS(R) applied to multi-region, high-density electrophysiology. (a) Mice receive a high- or low-tone auditory stimulus ("sample") and respond by licking left or right ("response"). (b) MR-LFADS single-trial predicted firing rates in held-out control trials (blue), recorded spike times (black vertical ticks), and smoothed, binned spike counts (black; causal, exponential filter) for example neurons in each modeled brain region. (c) Condition-averaged smoothed spike counts of example neurons in control and photoinhibition trials (left-hit condition). Shaded regions indicate the standard deviation across trials. (d) Photoinhibition-related changes in population recordings, as observed experimentally (left) and as predicted by MR-LFADS (right). (e) Cosine similarity of inferred effectomes across models with different random initializations (seeds). (f) Message norms inferred by MR-LFADS(R) for all connections across example seeds $k$. (g) Correlation of inferred message norms across pairs of seeds $(k,l)$. Box plots describe distributions of values across $5$ models fit from distinct random initializations.