🧬 biology

Fast probabilistic spike deconvolution using optimized Markov chain Monte Carlo

This paper introduces Optimized MCMC Spike Inference (OMSI), a method that significantly reduces the computational cost of continuous-time probabilistic spike deconvolution in calcium imaging by over 90% while maintaining accuracy comparable to traditional MCMC and non-probabilistic alternatives.

Original authors: Dylan Martins, Michael Goard

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dylan Martins, Michael Goard

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Neuroscientists have long possessed a powerful tool for watching the brain at work: a technique that uses light to see inside living tissue and track the activity of thousands of neurons simultaneously. By injecting a special fluorescent dye that glows when calcium enters a cell, researchers can watch neurons light up as they fire. However, this glow is a slow, blurry reaction. A single electrical spark from a neuron, which happens in a fraction of a millisecond, triggers a calcium wave that lingers for seconds. The resulting video looks like a slow, rolling wave of light rather than a rapid series of distinct flashes. To understand how the brain thinks, scientists need to reverse-engineer this slow glow to find the exact moments the original electrical sparks occurred. This process, known as spike deconvolution, is like trying to guess the precise timing of individual drumbeats by listening only to the long, fading echo they create in a large hall.

For years, the most accurate way to solve this puzzle has been a method called Markov chain Monte Carlo. This approach treats the problem as a game of chance, where a computer iteratively guesses, adds, removes, and shifts potential spikes in a continuous timeline to see which arrangement best explains the observed light. While this method is highly accurate and provides a measure of confidence for every single guess, it is incredibly slow. Processing a standard twenty-minute recording of five hundred neurons could take hours, making it impractical for the massive datasets modern neuroscience now generates. A new study by Dylan Martins and Michael Goard at the University of California, Santa Barbara, introduces a refined version of this method that cuts the required computing time by more than ninety percent without sacrificing accuracy. They call their approach Optimized MCMC Spike Inference, or OMSI.

The researchers achieved this dramatic speedup not by changing the fundamental logic of the problem, but by making the computer work smarter. In the traditional method, every time the computer proposed a change to the timeline of spikes, it would recalculate the entire twenty-minute recording to see if the new guess was better. The new method realizes that a single spike only affects the calcium glow for a few seconds. Therefore, the computer only needs to check a small, local window of time around the proposed change. This simple shift means the computer stops wasting energy recalculating minutes of data that have nothing to do with the specific spike being tested.

Further refinements accelerated the process even more. The team replaced the slow, initial guess used in older software with a faster mathematical shortcut that gets the computer to a good starting point almost immediately. They also taught the system to recognize when it has found a stable answer and stop working, rather than running a fixed number of rounds for every single cell. For neurons that are quiet or very noisy, the system stops early, saving time. For complex, active neurons, it continues until it is sure. Additionally, the software was rewritten to run at the speed of the computer's hardware, utilizing parallel processing to handle many calculations at once and using a lighter numerical format that requires less memory.

When the team tested OMSI against existing methods using simulated data, the results were striking. The new method produced spike trains that were just as accurate as the slow, traditional approach, matching the ground truth of when spikes actually occurred. In terms of speed, OMSI completed the analysis of a twenty-minute, five-hundred-neuron session in about five minutes on a standard computer processor. In contrast, the traditional method took roughly two and a half hours to do the same job. The new method was also faster than popular deep-learning tools that require expensive graphics cards, running at comparable speeds on standard hardware.

The researchers did not stop at simulations; they validated the method on real biological data where they knew the exact timing of the electrical spikes because they recorded them with electrodes alongside the calcium imaging. In these tests, OMSI performed as well as the best existing probabilistic methods and significantly better than older, faster algorithms that tended to guess too many spikes. The method proved robust across different types of light-sensitive dyes and various recording speeds. While the new approach did show slightly more difficulty with a specific, very fast dye that produces extremely brief flashes, it still outperformed other methods in most conditions and provided a level of detail that older, faster tools could not match.

The significance of this work lies in its ability to make high-precision analysis practical for large-scale studies. For decades, researchers had to choose between speed and accuracy: they could get a quick, rough estimate of neural activity or a slow, precise map of every single spike. OMSI bridges this gap, allowing scientists to process massive datasets with the precision of the most rigorous statistical methods in a fraction of the time. By making continuous-time probabilistic inference efficient, the tool opens the door to analyzing the complex, millisecond-scale conversations of large neural populations in ways that were previously computationally impossible.

Technical Summary: Fast Probabilistic Spike Deconvolution using Optimized Markov chain Monte Carlo

Problem Statement
Two-photon calcium imaging enables the monitoring of large neural populations, but the slow kinetics of calcium indicators often obscure individual action potentials within the observed fluorescence traces ( $\Delta F/F$ ). While algorithms exist to infer underlying spike trains, they face significant trade-offs. Deterministic methods (e.g., OASIS) are fast but lack per-spike uncertainty quantification, output binned activity rather than continuous-time spikes, and predict rates in arbitrary units. Deep learning approaches (e.g., CASCADE) offer speed but require extensive training data, GPUs for optimal performance, and lack interpretability or per-spike probability distributions. Conversely, Markov chain Monte Carlo (MCMC) methods provide a rigorous probabilistic framework that infers continuous-time spikes, quantifies uncertainty, and requires no training data. However, traditional MCMC implementations are computationally prohibitive, often requiring hours of compute time for standard datasets, rendering them impractical for large-scale pipelines.

Methodology: Optimized MCMC Spike Inference (OMSI)
The authors introduce OMSI, a method that preserves the full probabilistic framework of MCMC spike inference while achieving a >90% reduction in compute time through ten targeted optimizations:

Local Support Windowing: Instead of re-evaluating the log-likelihood ratio over the entire recording duration ( $T$ ) for every proposal, the algorithm restricts calculations to a local support window ( $T_{supp}$ ). This window is adaptively determined per cell by truncating the calcium kernel where its amplitude falls below 1% of its peak (typically ~6 seconds). This reduces the cost of proposal evaluation from $O(T)$ to $O(T_{supp})$ without sacrificing accuracy.
Adaptive Proposal Counting: The number of spike addition and removal proposals per sweep is scaled linearly with recording duration ( $\lceil T/500 \rceil$ ), ensuring proportional coverage for longer recordings.
NNLS Initialization: The method replaces the FOOPSI-based initialization with Non-Negative Least Squares (NNLS). NNLS is approximately 1.5 times faster than FOOPSI initialization while providing a sufficiently accurate starting spike train.
Adaptive Convergence Detection: Rather than running a fixed number of sweeps, OMSI monitors two sequential criteria checked every 50 sweeps. "Burn-in" ends when the amplitude trace stabilizes (means of two 50-sweep halves agree within 0.01%). Post-burn-in sampling terminates when the spike-count distribution stabilizes (difference < 3.16%), with a hard ceiling of 2,000 sweeps. This focuses compute time on complex cells and avoids waste on low-firing or high-noise cells.
Sub-Frame Precision via Accumulators: To support continuous-time spike positions with sub-frame precision, spike contributions are binned into two accumulator arrays weighted by rise and decay exponentials. This allows the Infinite Impulse Response (IIR) filter to be run in a single causal sweep, reducing reconstruction complexity from $O(T \times K)$ to $O(T)$ , where $K$ is the number of spikes.
Kernel Energy Caching: The self-energy term (squared-norm) required for likelihood evaluation is pre-computed and cached after updates to time constants. This allows $O(1)$ lookup for energy terms during proposal evaluation, avoiding redundant summation.
SNR Pre-screening: Cells with a Signal-to-Noise Ratio (SNR) < 2.0 are screened out and skipped entirely to prevent wasting compute resources on traces where inference is unlikely to succeed.
Sensor-Specific Kernel Adaptation: For fast indicators (e.g., GCaMP8) at high frame rates where rise and decay poles are numerically indistinguishable, the algorithm automatically switches from a double-exponential AR(2) kernel to a single-exponential AR(1) kernel.
Dynamic Firing Rate Estimation: For fast indicators, the empirical firing rate $\lambda$ is re-estimated from the current spike count at the end of every sweep, allowing the sampler to self-correct initial over-estimations.
Implementation Efficiency: The code is executed at C-speed using Numba, parallelized across CPU cores using Ray, utilizes float32 arithmetic, and minimizes memory allocation by overwriting arrays rather than re-initializing them.

Results
The performance of OMSI was evaluated on both simulated data and ground-truth datasets featuring simultaneous electrophysiology and calcium imaging.

Simulated Data: On a 20-minute, 500-neuron dataset, OMSI achieved accuracy comparable to the standard CaImAn-MCMC implementation (Median $F_\beta$ : 0.992 vs. 0.979; Mean CosMIC: 0.621 vs. 0.598) while reducing wall-clock time by 97.1% (from ~2.5 hours to ~5 minutes). OMSI's speed matched or exceeded deep learning approaches (CASCADE), running ~17.9% faster than CASCADE on CPU and only ~2.4% slower than CASCADE on GPU, all while maintaining the probabilistic advantages of MCMC.
Ground-Truth Data: On a public dataset of mouse visual cortical cells (GCaMP6f/6s), OMSI consistently matched or exceeded the performance of CaImAn-MCMC across various zoom levels and decay time constants. OMSI demonstrated high precision and balanced recall, outperforming OASIS (which tended to over-predict spikes) and matching CASCADE in accuracy metrics.
Scalability: OMSI showed linear scaling with cell count and recording duration. Unlike other methods, OMSI's compute time did not appreciably scale with spike train density.
Robustness: The method maintained high performance across a range of frame rates (7.5–100 Hz) and SNR levels (>2.0). Performance degraded significantly for all methods at SNR < 2.0.

Significance and Claims
The paper claims that OMSI establishes continuous-time probabilistic inference as a practical, high-throughput tool for large-scale calcium imaging pipelines. By reducing the computational burden of MCMC by more than an order of magnitude without sacrificing the statistical rigor of the full probabilistic framework, OMSI addresses the historical bottleneck that prevented the widespread adoption of probabilistic spike deconvolution.

The authors position OMSI as the appropriate choice when experimental aims require:

Accurate continuous-time action potential onsets.
Per-spike probability distributions (uncertainty quantification).
Generalization across experimental conditions without the need for matched training data.

While acknowledging that OASIS remains superior for near-instantaneous speed on standard hardware and that CASCADE offers rapid GPU-accelerated deconvolution, the paper argues that OMSI fills the critical gap for researchers needing the interpretability and precision of probabilistic methods at a scalable computational cost. The authors note that performance on GCaMP8 sensors was slightly lower than on other sensors, potentially due to the high noise levels in the available ground-truth datasets rather than a fundamental flaw in the method.

Technical Summary: Fast Probabilistic Spike Deconvolution using Optimized Markov chain Monte Carlo

More like this