Time cells differentially populate trace and post-trace epochs, but do not remap for different trace intervals

Using two-photon calcium imaging in mice during trace-eyeblink conditioning, this study reveals that hippocampal time cells form a fixed sequence triggered by the conditioned stimulus that persists for over five seconds and is actively suppressed during extinction, rather than remapping or rescaling to accommodate different trace intervals.

The Gist

The Big Picture: The Brain's Internal Stopwatch

Imagine your brain is a busy train station. Usually, we know where we are (spatial maps) because we see landmarks like trees or buildings. But what about time? How does your brain know when something is going to happen if there are no visual landmarks?

This paper studies "Time Cells" in the hippocampus (a part of the brain famous for memory). Think of Time Cells as a row of dominoes that fall one after another. When a specific event happens, the first domino falls, then the second, then the third. By watching which domino is currently falling, the brain knows exactly how much time has passed.

The researchers wanted to see how these "dominoes" work when the timing changes, and what happens when the brain decides to stop paying attention.

---

The Experiment: The "Eye-Blink" Game

The researchers taught mice a simple game called Trace Eyeblink Conditioning (TEC).

• The Setup: A blue light flashes (the CS or "Warning Signal").
• The Wait: There is a short pause (the "Trace").
• The Surprise: A tiny puff of air hits the mouse's eye (the US or "Punishment").
• The Goal: The mouse learns to blink before the air puff hits, right when it expects it.

The researchers tested the mice with two different wait times: a short wait (250ms) and a long wait (550ms). They used a special microscope to watch the neurons in the mouse's brain light up like a city skyline at night.

Key Findings: What They Discovered

1. The Dominoes Don't Stretch (No "Remapping")

The Question: If you change the wait time from 250ms to 550ms, do the brain's dominoes stretch out to fill the extra space? Or do they rearrange themselves?
The Discovery: They don't change at all.

• Analogy: Imagine a song playing on a loop. If you ask the song to play for 5 seconds instead of 3, a normal tape recorder would stretch the audio, making it sound slow and weird. But the mouse's brain is like a digital clock. It plays the exact same sequence of "ticks" (Time Cells) regardless of how long the wait is.
• The sequence of neurons firing is identical for both the short and long waits. The brain doesn't "remap" or stretch the timeline; it just keeps running the same internal clock sequence.

2. The "Ghost" Sequence (It Lasts Longer Than You Think)

The Discovery: The sequence of Time Cells keeps going for a long time—up to 9 seconds after the light flashes—even though the air puff only comes 0.25 or 0.5 seconds later.

• Analogy: It's like a firework display. The main explosion (the air puff) happens quickly, but the sparks and trails (the Time Cell activity) continue to light up the sky for a long time afterward.
• The brain creates a "template" of time that extends far beyond the actual event. This suggests the brain is always ready, keeping a long timeline open just in case.

3. Learning vs. Unlearning (The "Active Eraser")

The Discovery: When the researchers stopped giving the air puff (a process called "Extinction"), the mice stopped blinking. But something interesting happened in the brain: The Time Cells didn't just go back to how they were before learning. They were actively removed.

• Analogy: Imagine you learn a new dance routine. If you stop dancing, you might just forget the steps (passive forgetting). But here, the brain is like a stage manager who actively turns off the lights and removes the dancers from the stage.
• The study found that even before the mouse learned the game, some Time Cells were already there (latent). When the mouse learned, they got "gated on" (turned on). When the game ended, the brain actively "gated them off" (turned them off). It's an active process of deletion, not just fading away.

4. The "Fast" vs. "Slow" Clocks

The researchers noticed that the brain handles time differently depending on how fast the event is.

• Fast Clock (This study): For very short times (fractions of a second), the brain uses a rigid, fast sequence of dominoes that doesn't stretch.
• Slow Clock (Other studies): For longer tasks (like waiting 10 seconds for food), the brain seems to stretch and rearrange its dominoes.
• Analogy: It's like the difference between a metronome (strict, fast, unchangeable beats) and a rubber band (stretchy, slow, adaptable). The mouse's brain uses the metronome for quick eye-blink reactions.

The Conclusion: A "Template" for Time

The authors propose a new model:

1. Salience Triggers the Sequence: As soon as the brain notices an important signal (the light), it immediately starts running a pre-programmed "Time Sequence."
2. The Sequence is Fixed: This sequence runs its course regardless of what happens next. It doesn't wait to see if the air puff comes; it just runs.
3. The Brain "Gates" the Connection: Learning happens when the brain realizes, "Oh, the air puff usually happens at this specific point in the sequence." It connects the puff to that specific domino.
4. Extinction is a Switch: When the puff stops coming, the brain doesn't just forget; it actively flips a switch to stop the sequence from running, effectively "unlearning" the connection.

Why This Matters

This study changes how we think about memory. It suggests that for quick, precise reactions, our brains don't constantly recalculate time. Instead, they have a fixed, internal movie that plays out every time a signal appears. We learn by watching the movie and predicting the ending. If the ending never comes, we stop the movie entirely.

This helps us understand how animals (and humans) can react so quickly to precise timing, and how we can "unlearn" fears or habits by actively turning off these internal clocks.

Technical Summary

1. Problem Statement and Motivation

The study investigates the neural mechanisms of hippocampal time cells, which are neurons that fire at specific moments during a delay period, serving as a substrate for linking discontiguous events in time. While time cells have been extensively characterized in working memory tasks (delays of ~15 seconds) and Trace Eyeblink Conditioning (TEC) (delays of ~300 ms), significant gaps remain in understanding:

• Rescaling: Do time cells stretch or remap when the trace interval changes (e.g., from 250 ms to 550 ms), or do they maintain a fixed temporal code?
• Learning Dependence: Are time cells generated de novo during learning, or do they exist as latent sequences triggered by salient stimuli before association is formed?
• Extinction: What happens to time cell representations when a learned behavior is extinguished? Does the network revert to a pre-learning state, or is there an active suppression mechanism?
• Temporal Scope: How do time cell sequences evolve over periods significantly longer than the trace interval itself?

The authors hypothesize that TEC, with its sub-second timescale and requirement for precise, pre-emptive timing, may utilize distinct mechanisms compared to multi-second working memory tasks.

2. Methodology

The study employed in vivo two-photon calcium imaging in the CA1 region of the dorsal hippocampus in mice.

• Subjects: 7 mice trained on the task; 4 mice underwent cranial window surgery for imaging (expressing GCaMP6f in pyramidal neurons).
• Task Protocol (Trace Eyeblink Conditioning - TEC):
  • Conditioned Stimulus (CS): 50 ms blue LED flash.
  • Unconditioned Stimulus (US): 50 ms air puff to the eye.
  • Trace Intervals: Mice were trained through staged intervals (250 ms, 350 ms, 450 ms, 550 ms).
  • Interleaved Block Design: A critical phase involved alternating blocks of 5 trials with 250 ms intervals and 5 trials with 550 ms intervals within the same session. This design was intended to eliminate learning-dependent rescaling and test for immediate remapping.
  • Extinction: Sessions where the CS was presented without the US reinforcement.
  • Probe Trials: ~10% of trials were CS-only (no US) to test sequence stability without reinforcement.
• Imaging & Analysis:
  • Data Acquisition: 2-photon imaging at ~12.8 Hz frame rate.
  • Time Cell Identification: Used a fast C++ implementation of the Ridge-to-Background (R2B) algorithm (Modi et al., 2014) and Temporal Information (TI) metrics. Criteria included hit rate, precision, and R2B ratio > 2.
  • Epoch Definition: Trials were divided into: Pre-CS, E1 (0–1s), E2 (1–3s), E3 (3–6s), and E4 (>6s).
  • Statistical Analysis: Used Friedman tests, Wilcoxon signed-rank tests, Spearman correlations, and session-level bootstrapping (10,000 iterations) to assess significance.

3. Key Results

A. Behavioral Performance

• Mice successfully learned the TEC task across all intervals, achieving >60% conditioned responses (CR).
• Tuning: The peak of the eyeblink response shifted to match the trace interval (250 ms vs. 550 ms), demonstrating precise temporal tuning.
• Extinction: Removal of the US led to a rapid decline in CRs.

B. Population Activity and Time Cell Density

• Post-Stimulus Activity: CS triggered a sharp rise in population calcium events, peaking at Epoch 2 (1–3s) and stabilizing at a plateau for up to 9 seconds post-CS.
• Learning Independence: The proportion of time cells and overall population event rates did not change significantly between pre-learning, learning, and post-learning stages. Time cells were present even before the animal learned the association.
• Extinction Effect: During extinction, the proportion of time cells dropped drastically (from ~7.3% to ~1.1%), despite overall calcium event rates remaining slightly elevated. This suggests an active removal of time cell representations rather than simple forgetting.

C. No Remapping for Trace Intervals (The Core Finding)

• Stability: When comparing 250 ms and 550 ms trials in the interleaved blocks, time cells did not remap.
• Correlation: The activity time-courses of identified time cells were strongly correlated (Spearman's $\rho \approx 0.5$) across both intervals and probe trials.
• Peak Timing: The distribution of peak firing times did not significantly differ between 250 ms and 550 ms conditions. Cells did not "stretch" to fill the longer interval; instead, they maintained their specific temporal firing preferences (e.g., firing at "2 seconds" regardless of whether the US arrived at 250 ms or 550 ms).
• Contrast with Controls: Non-time-tuned active cells showed no such correlation, confirming the specificity of the time cell sequence.

D. Epoch-Specific Dynamics

• Density Gradient: Time cell tuning density was highest in the first second (E1) and declined monotonically over time.
• Interval Differences: The 250 ms condition had a higher proportion of cells in E1, while the 550 ms condition showed enrichment in E2, suggesting a shift in coverage toward the reinforcement time, but the underlying sequence remained stable.
• Tuning Width: Cells peaking in the 1–3s epoch (E2) had significantly broader tuning widths (FWHM) compared to those in E1 or E3.

4. Key Contributions

1. Invariance of Time Cells in TEC: Demonstrated that in sub-second trace conditioning, time cells encode absolute time (specific durations) rather than relative time (rescaling to the interval). This contrasts with working memory tasks where remapping/stretching is common.
2. Pre-Learning Existence: Showed that time cell sequences are initiated by the salient CS before behavioral learning occurs, challenging the view that they are solely a product of associative learning.
3. Active Extinction Mechanism: Provided evidence that extinction involves an active suppression of time cell sequences, distinct from a reversion to a pre-learning baseline.
4. Extended Temporal Scope: Revealed that time cell sequences in TEC extend well beyond the trace interval (up to 9 seconds), suggesting a bridge between sub-second conditioning and multi-second working memory mechanisms.
5. Epoch Structure: Identified distinct epochs of time cell activity with varying densities and tuning widths, proposing a hierarchical temporal encoding structure.

5. Significance and Hypothesis

The authors propose a "Latent Sequence Gating" model:

• Salience Gating: A salient stimulus (CS) triggers a pre-existing, latent sequence of time cells.
• Association: Learning involves mapping the US onto this existing sequence.
• Stability: The sequence is robust and does not remap for small changes in interval (250–550 ms), acting as a fixed "template."
• Extinction: The network actively "gates off" or suppresses this sequence when the association is broken, rather than erasing it.

Significance: This work suggests that the hippocampus utilizes different computational strategies for fast (sub-second) vs. slow (multi-second) timing. The findings support the idea that time cells provide a stable temporal scaffold that can be flexibly linked to events, with the flexibility lying in the gating of the sequence rather than the remapping of the sequence itself. This has implications for understanding how the brain handles rapid associative learning and the neural basis of extinction disorders.