To investigate whether an AI model can detect stochastic synaptic events in noisy single-trial time-series data, we designed a deep neural network consisting of CNN, long short-term memory (LSTM), and fully connected dense layers (’miniML’, Figure 1A). The CNN-LSTM model takes a section of a univariate time-series recording as input and outputs a label for that section of data (Figure 1B). The miniML model is trained to classify short segments of electrophysiological data as either positive or negative for a synaptic event using supervised learning. The trained classifier can then be applied to unseen time-series data to localize events.

To train the miniML model, we first extracted a large number of synaptic events and corresponding event-free sections from previous voltage-clamp recordings of cerebellar mossy fiber to granule cell (MF–GC) miniature excitatory postsynaptic currents (mEPSCs) (Delvendahl et al., 2019). All samples were then visually inspected and labeled to generate the training dataset. We applied data augmentation techniques to include examples of typical false positives in the training data (Materials and methods). In total, the training data comprised ∼30,000 samples that were split into training and validation sets (0.75/0.25). Across training epochs, loss decreased and accuracy increased, stabilizing after ∼30 epochs (Figure 1C). The model with the highest validation accuracy was selected for further use, achieving 98.4% (SD 0.1, fivefold cross-validation). Saliency map analysis (Simonyan et al., 2013) indicated that the AI model mainly relied on the data sections around the peak of synaptic events to discriminate with respect to the labels (Figure 1—figure supplement 1). The trained miniML model achieved an area under the receiver operating characteristic (ROC) curve close to 1 (Figure 1D), indicating almost perfect separability of the classes (Figure 1E, Figure 1—figure supplement 1). Deep learning typically requires large datasets for training (van der Ploeg et al., 2014). To investigate how miniML’s classification performance depended on the dataset size, we systematically varied the number of training samples. As expected, the accuracy increased with larger datasets. However, the performance gain was marginal when exceeding 5000 samples (<0.2%, Figure 1—figure supplement 2), indicating that relatively small datasets suffice for effective model training (Bailly et al., 2022). These results demonstrate the efficacy of supervised deep learning in accurately classifying segments of neurophysiological data containing synaptic events.

Synaptic events are typically recorded as continuous time-series data of either membrane voltage or current. To apply the trained classifier to detect events in arbitrarily long data, we used a sliding window approach (Figure 2A). Time-series data are divided into overlapping sections that correspond to the input shape of the CNN-LSTM classifier. We used a stride for the sliding window, which reduces the number of inferences needed and speeds up the computation time while maintaining high detection performance (Figure 2—figure supplement 1). By reshaping the data into short sections, model inference can be run in batches and employ parallel processing techniques, including graphics processing unit (GPU) computing, resulting in analysis times of a few seconds for minute-long recordings (Figure 2—figure supplement 1). The miniML detection method is thus time-efficient and can be easily integrated into (high-throughput) data analysis pipelines. To facilitate the use of the method, miniML also includes a graphical user interface (Figure 2—figure supplement 2).

The output of the miniML model predicts the label—no event or event—for each time step (i.e. stride) with a numerical value ranging from zero to one (Figure 1B). These output values can be interpreted as the confidence that the corresponding data segment contains a synaptic event. Model inference thus outputs a prediction trace that has a slightly lower temporal resolution compared to the original input according to the stride of the sliding window. Using peak finding on the prediction trace allows extracting data segments with individual events from a recording. While 0.5 represents a reasonable minimum peak value, the exact choice is not critical to detection performance (see below). To quantify events, extracted sections are checked for overlapping events, and detected events are aligned by the steepest slope, allowing calculation of individual event statistics. Figure 2B illustrates the sliding window approach to detect spontaneous events in time-series data, such as a continuous voltage-clamp recording of spontaneous mEPSCs in a cerebellar GC. miniML provided a clear peak for all synaptic events present, without false positives (Figure 2C, Figure 2—figure supplement 3), allowing fast and reliable event quantification (Figure 2D). These data demonstrate that a deep learning model can be applied to detect synaptic events in electrophysiological time-series data.

To benchmark the AI model’s event detection performance, we compared it with commonly used template-based approaches (template-matching Clements and Bekkers, 1997 and deconvolution Pernía-Andrade et al., 2012), and a finite-threshold-based method (Kudoh and Taguchi, 2002). Some of these—or similar—algorithms are implemented in proprietary software solutions to record and analyze electrophysiological data. We also included a Bayesian event detection approach (Merel et al., 2016), SimplyFire (Mori et al., 2024), and the automated event detection routine of MiniAnalysis software (Synaptosoft Inc). To compare synaptic event detection methods, we developed a standardized benchmarking pipeline (Figure 3A). We first performed event-free voltage-clamp recordings from mouse cerebellar GCs (in the presence of blockers of inhibitory and excitatory transmission; see Materials and methods and Figure 2—figure supplement 3). We then generated synthetic events with a two-exponential time course and superimposed these on the raw recording to produce ground-truth data (Figure 3B–C). Event amplitudes were drawn from a log-normal distribution (Figure 3—figure supplement 1) with varying means to cover the range of SNRs typically observed in recordings of miniature events (2–15 dB, data from n = 170 GC recordings, Figure 3D). We generated events with kinetics that closely matched the template used for the matched-filtering approaches, thus ensuring a conservative comparison of miniML with other methods applied under optimal conditions (i.e. using the exact average event shape as template). To measure detection performance, we calculated recall (i.e. sensitivity), precision (fraction of correct identifications) and the F1 score. Recall depended on SNR for all methods, with miniML and deconvolution showing the highest values (Figure 3E). The precision was highest for miniML, which detected no false positives at any SNR, in contrast to all other methods (Figure 3F). When assessing overall performance, miniML provided the highest F1 scores across SNRs (Figure 3G). In addition, miniML showed superior results when changing event kinetics, indicating higher robustness to variable event shapes (Figure 3H, Figure 3—figure supplement 1), which may be particularly important in neurons with diverse synaptic inputs due to mixed receptor populations (Lesperance et al., 2020), or during pharmacological experiments (Ishii et al., 2020).

Conventional event detection methods typically produce a detection trace with a shape identical to the input data and values in an arbitrary range (Figure 3C). In contrast, miniML generates output in the interval [0, 1], which can be interpreted as the confidence of event occurrence. The output of event detection methods—the detection trace—significantly increases the SNR with respect to the original data. In our benchmark scenario, miniML and the Bayesian method provided the greatest discrimination from background noise (Figure 3C). To locate the actual event positions, a threshold must be applied to the detection trace, which can drastically affect the results, as peaks in the detection trace may depend on the event amplitude. Intriguingly, miniML’s detection trace peaks did not depend on event amplitudes, setting it apart from other methods (Figure 3—figure supplement 1). For template-based methods, recommendations on threshold selection are provided (Clements and Bekkers, 1997; Pernía-Andrade et al., 2012), but users usually need to adjust this parameter according to their specific data. The choice of the threshold strongly influences the detection performance, as even small changes can lead to marked increases in false positives or false negatives. To investigate the threshold dependence of different methods, we systematically varied the threshold and analyzed the number of detected events and the F1 score. Notably, miniML’s detection performance remained consistent over a wide range (5–195%, Figure 3I-J, Figure 3—figure supplement 1), with false positives occurring only at the lower threshold limit (5%, corresponding to a cutoff value of 0.025 in the detection trace). Conversely, the other detection methods were very sensitive to threshold changes (Figure 3J, Figure 3—figure supplement 1) due to the comparatively low SNR ratio of the output traces (but note that the Bayesian method is only slightly threshold dependent). In addition, the miniML threshold (i.e. minimum peak) value is bounded with an intuitive meaning. These comparisons underscore that miniML requires no prior knowledge of the exact event shape and is virtually threshold independent, thus enabling reliable event detection.

We trained the miniML model using data from cerebellar MF–GC synapses, which enabled reliable analysis of mEPSC recordings from this preparation (Figures 2 and 4A-C). Comparison with several previous detection methods revealed that miniML detected more events with a waveform that is consistent with mEPSCs (Figure 4D, Figure 4—figure supplement 1). Synaptic event properties, such as kinetics, SNR, or frequency, are variable between preparations. This heterogeneity is caused by, for example, differences in the number and type of postsynaptic neurotransmitter receptors, the number of synaptic inputs, postsynaptic membrane properties, or presynaptic release properties. To test miniML’s generalizability, we evaluated the detection performance on out-of-sample data. We recorded and analyzed data from the mouse calyx of Held synapse (Figure 4E), a large axosomatic synapse in the auditory brainstem that relays rate-coded information over a large bandwidth and with high temporal precision. Due to the large number of active zones, synaptic events are quite frequent. Despite being trained on cerebellar MF–GC data, miniML accurately detected mEPSCs in these recordings (Figure 4F–H). This result may be facilitated by the fast event kinetics at the calyx of Held, which approach those of MF–GC mEPSCs. To test whether miniML could also detect events with slower kinetics, we applied it to recordings from cerebellar Golgi cells (Figure 4I; Kita et al., 2021). These interneurons in the cerebellar cortex provide surround inhibition to GCs and receive input from parallel fiber and MF synapses. miniML reliably detected synaptic events in these recordings (Figure 4J–L), although event decay kinetics were slower compared to the training data (Golgi cell, 1.83 ms [SD 0.44 ms, n = 10 neurons], GC training data, 0.9 ms). Synaptic events are also commonly recorded from neuronal culture preparations. We determined whether miniML could be applied to recordings from cultured human induced pluripotent stem cell (iPSC)-derived neurons. We patch-clamped neurons in 8-week-old cultures of predominantly cortical glutamatergic identity (Asadollahi et al., 2023) and recorded spontaneous synaptic events in voltage-clamp (Figure 4M). Using miniML on these human iPSC-derived neuron data showed robust detection of synaptic events (Figure 4N–P), which had an average frequency of 0.15 Hz (SD 0.24 Hz, n = 56 neurons). Taken together, the consistent performance across different synaptic preparations indicates that the miniML model can be directly applied to out-of-sample data with similar kinetics. miniML also provided higher event detection accuracy than template-based or finite-threshold-based detection methods, which are prone to false positives (Figure 4D, H, L, P, Figure 4—figure supplement 1).

While applicable to out-of-sample data (Figure 4), simulations indicated that larger differences in event kinetics and waveform may ultimately hinder detection when using the MF–GC mEPSC model (Figure 5—figure supplement 1). Not only distinct event waveforms, but also different hardware and recording conditions may affect the characteristics of the recorded data. Scalability and generalizability are important for robust event detection. To facilitate the application of miniML to different preparations, recording conditions, and data types, we employed a transfer learning (TL) strategy. TL is a powerful technique in machine learning that allows for the transfer of knowledge learned from one task or domain to another (Yosinski et al., 2014; Caruana, 1994). TL is widely used with CNNs to take advantage of large pre-trained models and repurpose them to solve new, unseen tasks (Theodoris et al., 2023). Importantly, only a part of the network needs to be trained for the novel task, which significantly reduces the number of training samples needed and speeds up training while avoiding overfitting (Yosinski et al., 2014). We therefore reasoned that TL based on freezing the convolutional layers during training of our pre-trained network could be used to train a new model to detect events with different shapes and/or kinetics, using a lower number of training samples (Figure 5A).

We tested the use of TL for miniML with recordings of miniature excitatory postsynaptic potentials (mEPSPs) in mouse cerebellar GCs. These events have opposite polarity and slower kinetics compared to the original mEPSC training data. We compared TL-based model training to full training (with all layers trainable and reinitialized weights), varying the training sample size. Whereas accuracy increased and loss decreased with the number of samples (Figure 5B), TL models performed well with as few as 400 samples. Under these conditions, accuracy was only slightly lower than for full training with almost 10 times the sample size (median accuracy, 95.4 vs 96.1%; Figure 5C). This suggests that TL can significantly reduce the sample size needed for training, minimizing the time-consuming process of event extraction and labeling. Across different datasets, TL-trained models performed comparably to those trained from scratch (Figure 5—figure supplement 2). Taken together, these data demonstrate that TL allows model training with an order of magnitude smaller amount of training data, making miniML easily transferable to new datasets with diverse characteristics.

To investigate how a TL-trained model performs in event detection, we analyzed data in different recording modes (voltage-clamp vs. current-clamp). The different noise conditions typically encountered in current-clamp and the difference between EPSP and EPSC waveforms necessitate distinct detection schemes for common detection methods. Recording miniature EPSPs and EPSCs in the same cerebellar GCs (Figure 5D–E) enabled us to compare detection performance via event frequency. mEPSPs could be approximated by a two-exponential time course, similar to postsynaptic currents. However, their kinetics were considerably slower due to the charging of the plasma membrane (Figure 5E). We used a TL-trained model for mEPSP detection and the standard miniML model for mEPSCs. Remarkably, the average event frequencies were very similar in the two different recording modes (voltage-clamp: 0.49 Hz, SD 0.53 Hz, current-clamp: 0.54 Hz, SD 0.6 Hz, n = 15 for both) and highly correlated across neurons (Figure 5F). This highlights the reliability of TL with small training sets for the consistent detection of synaptic events in datasets with varying characteristics.

The presence of overlapping and highly variable event shapes pose additional challenges for event detection methods. To evaluate miniML’s performance in such complex scenarios, we analyzed a dataset recorded from principal neurons of the adult zebrafish telencephalon (Rupprecht and Friedrich, 2018). We focused on spontaneous excitatory inputs to these neurons, characterized by diverse event shapes and frequencies (Rupprecht and Friedrich, 2018). Training via TL (see Materials and methods) yielded a model that enabled the reliable detection of spontaneous excitatory currents (Figure 6A). Analysis of event properties across cells revealed broad distributions of event statistics (Figure 6B), including a large diversity of event rise and decay kinetics (Figure 6B–C). Notably, miniML consistently identified synaptic events with diverse kinetics and shapes.

We next used the extracted event kinetics features of individual neurons (Figure 6D–H) to demonstrate miniML’s utility in better understanding the diversity of an existing dataset. First, we explored whether the anatomical location of each neuron could predict event decay times. We mapped the recorded neurons to an anatomical reference and plotted decay times as a function of their position but did not find a strong relationship (Figure 6I; correlation with position p > 0.05 in all three dimensions). Second, we tested the hypothesis that slower event kinetics are associated with larger cells. In large cells, EPSCs may undergo stronger filtering as they propagate from synaptic sites to the soma. Consistent with this idea, we observed correlations between decay and rise times across neurons (Figure 6J). Furthermore, the distribution of decay times (examples shown in Figure 6H) was broader for neurons with longer decay times (Figure 6K), suggesting a broader distribution of distances from synapses to the cell body. Finally, for a subset of neurons, we recorded input resistance, which approximates the cell membrane resistance and is therefore a proxy for cell size. Input resistance was negatively correlated with decay times across neurons (Figure 6L), consistent with the hypothesis that diverse event kinetics across neurons are determined by the conditions of synaptic event propagation to the soma and, more specifically, cell size. Taken together, this analysis underscores the versatility of miniML, as it can be successfully applied to new datasets with varying recording conditions. miniML consistently extracted synaptic events across a spectrum of event kinetics, enabling the identification and investigation of key factors determining event kinetics and other event-related properties across neurons.

We next applied miniML to analyze data obtained from the Drosophila melanogaster larval neuromuscular junction (NMJ) (Baccino-Calace et al., 2022). This synapse is characterized by a higher frequency of spontaneous synaptic events with a slower time course compared with typical brain slice recordings. In addition, mEPSC recordings are performed using two-electrode voltage-clamp, which can be associated with large leak currents and rather low SNR. Because of the different event shapes in these data, we used a small dataset of manually extracted NMJ events to train a TL model. The TL model was able to reliably detect synaptic events in wild-type (WT) NMJ recordings (Figure 7A–C). We next assessed event detection upon deletion of the non-essential glutamate receptor subunit GluRIIA, which causes a strong reduction in mEPSC amplitude and faster kinetics (DiAntonio et al., 1999). A separate TL model allowed reliable synaptic event detection in recordings from GluRIIA mutant larvae (Figure 7D–F). We observed a 54% reduction in mEPSC amplitude compared to WT (Figure 7G), consistent with previous reports (DiAntonio et al., 1999; Petersen et al., 1997). In addition, the event frequency was reduced by 64% (Figure 7H). Although event amplitude distributions had a similar shape in both genotypes (Figure 7C and F), small events below the detection limit in GluRIIA synapses may contribute to the observed frequency difference. Half decay and rise times were also shorter at GluRIIA than at WT NMJs (−58% and −18%, respectively; Figure 7I–J), which can be explained by the faster desensitization of the remaining GluRIIB receptors (DiAntonio et al., 1999). Thus, miniML can be applied to two-electrode voltage clamp recordings at the Drosophila NMJ and robustly resolves group differences upon genetic receptor perturbation.

To assess miniML’s generalizability to non-electrophysiological data, we applied it to analyze events in time-series data from live fluorescence imaging experiments. Recent developments of highly sensitive fluorescent probes have enabled the recording of synaptic events using various imaging techniques (Hao et al., 2024; Aggarwal et al., 2023; Abdelfattah et al., 2023; Ralowicz et al., 2024). These technological advances offer exciting new possibilities in the study of synaptic function, but also generate novel challenges for the detection and analysis of synaptic events. Live imaging datasets typically feature a lower sampling rate, lower SNR, and distinct noise profile compared with electrophysiological recordings. Nevertheless, the waveforms of imaged synaptic release events, voltage changes, or Ca²⁺ transients closely resemble those used to train miniML. Thus, we hypothesized that miniML could also be employed for event detection in fluorescence imaging data.

We first used miniML to analyze a previously published dataset Aggarwal et al., 2023 from rat neuronal cultures expressing the glutamate sensor iGluSnFR3 (Figure 8A–B). These data, recorded in the presence of TTX, contain spontaneous transient increases in fluorescence intensity representing individual release events (‘optical minis’). Initially, we selected a small subset of events from the data to train a TL model. Given the low sampling rate of the imaging (100 Hz), we upsampled the data by a factor of 10 to match the model’s input shape. The TL model was subsequently applied to all detected sites (n = 1524) within the widefield recording (Aggarwal et al., 2023). A qualitative assessment of the imaging traces showed excellent event detection, with miniML consistently localizing the iGluSnFR3 fluorescence transients at varying SNRs (Figure 8C). The detected optical minis had similar kinetics to those reported in Aggarwal et al., 2023 (10–90% rise time, median 21.8 ms; half decay time, median 48.7 ms, Figure 8D). In addition, analysis of event frequencies across sites revealed a power-law distribution (Figure 8E), consistent with fractal behavior of glutamate release (Lowen et al., 1997). Thus, miniML can reliably detect synaptic release events in time-series data from iGluSnFR3 recordings.

To further investigate miniML’s performance in imaging data, we performed simultaneous electrophysiological and fluorescence recordings of mEPSPs in cultured rat hippocampal neurons expressing ASAP5-Kv (Hao et al., 2024) at physiological temperature (Figure 9A–B). ASAP5 allows resolving small voltage changes in the mV-range, as illustrated by the close correlation of optical and current-clamp signals (Figure 9C). However, event detection is more challenging in voltage imaging data due to the lower SNR compared to electrophysiological recordings (ASAP5: 2.26, SD 0.46; electrophysiology: 4.79, SD 0.74; n = 5 neurons; Figure 9—figure supplement 1). Photon shot noise and other sources of noise in the imaging setup not only limit the SNR, but can also lead to spurious event detection (Wilt et al., 2013; Sjulson and Miesenböck, 2007).

We trained separate miniML TL models to detect mEPSPs in optical and current-clamp data (Figure 9C–D). Optical detection yielded events that closely resembled mEPSPs, with slightly slower decay and rise kinetics (Figure 9E, Figure 9—figure supplement 1). Analysis of the corresponding amplitudes of mEPSPs and optically detected events confirmed that ASAP5 linearly reports subthreshold voltage changes in neurons (Figure 9E), in line with previous reports (Hao et al., 2024). To provide a more quantitative assessment of detection performance, we compared miniML to established template-based analysis methods (Figure 9F). The simultaneous recording of membrane voltage through both optical and electrophysiological means allowed us to establish ground truth data for evaluating detection performance. We quantified the detection of events by different methods in imaging data relative to electrophysiology, with threshold settings of the matched-filtering approaches adjusted to achieve ∼90% precision. High precision of event detection is crucial for robust data analysis. At comparable levels of precision, miniML demonstrated superior recall of mEPSPs compared with template matching or deconvolution (Figure 9G–H). Notably, miniML detected ∼40% of mEPSPs in optical data, a substantial improvement over template-based methods (Figure 9H, Figure 9—figure supplement 2). These voltage imaging experiments suggest that miniML facilitates the optical detection of mEPSPs. More extensive training datasets may further increase the recall performance of miniML in voltage imaging datasets.

Together, the analysis of different optical recordings containing synaptic events demonstrates that miniML generalizes beyond electrophysiological data and can support the evaluation of fluorescence imaging datasets. miniML may thus pave the way for high-throughput analyses of spontaneous events in diverse imaging experiments.

Animals were treated according to national and institutional guidelines. All experiments were approved by the Cantonal Veterinary Office of Zurich (approval number no. ZH206/2016 and ZH009/2020). Experiments were performed in male and female C57BL/6 J mice (Janvier Labs, France, RRID:MGI:2670020). Mice were 1–5 months-old, but for recordings from the Calyx of Held, which were performed in P9 animals. Animals were housed in groups of 3–5 in standard cages under a 12h-light/12h-dark cycle with food and water ad libitum. Mice were sacrificed by rapid decapitation after isoflurane anesthesia. The cerebellar vermis was quickly removed and mounted in a chamber filled with chilled extracellular solution. Parasagittal 300-µm-thick slices were cut with a Leica VT1200S vibratome (Leica Microsystems, Germany), transferred to an incubation chamber at 35 °C for 30 min, and then stored at room temperature until experiments. The extracellular solution (artificial cerebrospinal fluid, ACSF) for slicing and storage contained (in mM): 125 NaCl, 25 NaHCO₃, 20 D-glucose, 2.5 KCl, 2 CaCl₂, 1.25 NaH₂PO₄, 1 MgCl₂, aerated with 95% O₂ and 5% CO₂.

Slices were visualized using an upright microscope with a 60×, 1 NA water immersion objective, infrared optics, and differential interference contrast (Scientifica, UK). The recording chamber was continuously perfused with ACSF. For event-free recordings, we blocked excitatory and inhibitory transmission using ACSF supplemented with 50 µM D-APV, 10 µM NBQX, 10 µM bicuculline, and 1 µM strychnine. Patch pipettes (open-tip resistances of 3–8 MΩ) were filled with solution containing (in mM): 150 K-D-gluconate, 10 NaCl, 10 HEPES, 3 MgATP, 0.3 NaGTP, 0.05 ethyleneglycol-bis(2-aminoethylether)-N,N,N’,N’-tetraacetic acid (EGTA), pH adjusted to 7.3 with KOH. Voltage-clamp and current-clamp recordings were made using a HEKA EPC10 amplifier controlled by Patchmaster software (HEKA Elektronik GmbH, Germany, RRID:SCR_000034). Voltages were corrected for a liquid junction potential of +13 mV. Experiments were performed at room temperature (21–25 °C). Miniature EPSCs (mEPSCs) were recorded at a holding potential of −100 mV or −80 mV, and miniature EPSPs (mEPSPs) at the resting membrane potential. Data were filtered at 2.9 kHz and digitized at 50 kHz. Synaptic event recording periods typically lasted 120 s.

Voltage-clamp recordings were performed on human iPSC-derived neurons as described in Asadollahi et al., 2023. We recorded spontaneous EPSCs in 8-week-old neuronal cultures with cortical glutamatergic identity at a holding potential of −80 mV and with ACSF consisting of (in mM): 135 NaCl, 10 HEPES, 10 D-glucose, 5 KCl, 2 CaCl₂, 1 MgCl₂. Synaptic events were observed in ∼51% of neurons. The experimental procedures for recordings of spontaneous or miniature EPSCs in zebrafish and Drosophila are described in Rupprecht and Friedrich, 2018 and Baccino-Calace et al., 2022, respectively.

Cultured rat hippocampal neurons at 17–20 days in vitro were patched and imaged simultaneously at physiological temperature (30–34 °C) following the procedure described in Hao et al., 2024. Briefly, for electrophysiological recordings, extracellular solution containing 145 mM NaCl, 3 mM KCl, 2 mM CaCl₂, 2 mM MgCl₂, 10 mM HEPES, and 10 mM glucose (pH adjusted to 7.4, Osmolarity 310 mOsm/kg) supplemented with TTX (1 µM) and PTX (50 µM), and a pipette solution containing 123 mM K-gluconate, 10 mM KCl, 8 mM NaCl, 1 mM MgCl₂, 10 mM HEPES, 1 mM EGTA, 0.1 mM CaCl₂, 1.5 mM MgATP, 0.2 mM Na₄GTP, and 4 mM glucose (pH adjusted to 7.2, Osmolarity 295–300 mOsm/kg) were used. Whole-cell recordings were performed with a Multiclamp 700B amplifier, Digidata 1440 A digitizer, and pClamp software (all from Molecular Devices). Patch pipettes (open-tip resistances of 3–5 MΩ) were produced from glass capillaries with filament (outer diameter, 1.5 mm; inner diameter, 1.0 mm; King Precision Glass). Electrophysiological recordings were digitized at 10 kHz for acquisition and then lowpass filtered with 1 kHz cut-off frequency in post-processing in Clampfit.

ASAP5-Kv was excited using a SOLIS-470C LED light source (Thorlabs) with a 482/18 nm bandpass filter (Semrock); excitation light was set to provide an irradiance of 46 mW/mm². Imaging data were acquired at 400 fps using an iXon 860 EMCCD camera (Andor – Oxford Instruments) and a 525/50 nm bandpass filter (Semrock). Neurons requiring more than ±100 pA holding current to maintain the resting membrane potential near −70 mV were excluded from the analysis.

We used synaptic event recordings from a previous publication (Delvendahl et al., 2019) to generate the training dataset. mEPSCs were extracted based on low-threshold template-matching. Corresponding sections of data without events were randomly selected from the recordings. The extracted windows had a length of 600 data points. Given our sampling rate of 50 kHz, this corresponds to 12 ms. We subsequently manually scored data sections as event-containing (label = 1) or not event-containing (label = 0). The ratio of events to non-events was kept close to one to ensure efficient training. Based on empirical observations of model performance, we included relatively small amplitude events that are often missed by other methods. Similarly, including negative examples that are commonly picked up as false positives, resulted in more accurate prediction traces. We used data augmentation techniques to further improve model discrimination. We simulated waveforms of non-synaptic origin, which are occasionally encountered during recordings, and superimposed them onto noise recordings. Examples included rapid current transients that can be caused by peristaltic perfusion pump systems often used in brain slices recordings, or slow currents with a symmetric rise and decay time course. A total of 4500 segments were created and labeled as 0. To maintain the ratio of negative to positive examples, we added an equivalent number of synthetic synaptic events. The biexponential waveform described in the section Benchmarking detection methods was used for event simulation. The final training dataset contained 30,140 samples (21,140 from recorded traces and 9,000 simulated samples).

We built a discriminative end-to-end deep learning model for one-dimensional time-series classification (Ismail Fawaz et al., 2019). The neural network architecture comprised multiple convolutional layers, an LSTM layer, and a fully connected layer. The approach is related to networks designed for event detection in audio (Passricha and Aggarwal, 2019) and image data (Islam et al., 2020; Donahue et al., 2017), or classification of genomic data (Tasdelen and Sen, 2021). The combination of convolutional and recurrent neural network layers helps to improve the classification performance for time-series data. In particular, LSTM layers allow learning temporal features. An initial comparison of different types of network architectures showed superior performance of a CNN-LSTM model over CNN-Dense, Residual Neural Network (ResNet), and multi-layer perceptron (MLP) architectures (Figure 1—figure supplement 2).

The miniML model was built using TensorFlow, an open-source machine learning library for Python (Abadi et al., 2015). To optimize the model architecture, we performed a Bayesian optimization of hyperparameters. Hyperparameter ranges were chosen for the free parameters of all layers. Optimization was then performed with a maximum number of trials of 50. Models were evaluated using the validation dataset. Because a higher number of free parameters tended to increase inference times, we then empirically tuned the chosen hyperparameter combination to achieve a trade-off between number of free parameters and accuracy.

The deep learning network takes batches of one-dimensional (1D) univariate time-series data as input, which are converted into a tensor of shape (batch size, data length, 1). The data is passed to three convolutional blocks. Each block consists of a 1D convolutional layer with a leaky Rectified Linear Unit (leaky ReLU) activation function followed by an average pooling layer. To avoid overfitting, each convolutional block uses batch normalization (Ioffe and Szegedy, 2015). Batch normalization reduced training time by about two times and improved the accuracy of the resulting model. We added a fourth convolutional block that includes a convolutional layer, batch normalization and a leaky ReLU activation function, but no average pooling layer. The output of the final convolutional layer passes through a bidirectional recurrent layer of LSTM units. The final layers consist of a fully connected dense unit layer and a dropout layer (Srivastava et al., 2014), followed by a single unit with sigmoid activation. The output of the neural network is a scalar between [0, 1]. The layers and parameters used, including output shape and number of trainable parameters, are summarized in Table 1. The total number of trainable parameters was 190,913.

The network was trained with Tensorflow 2.12 and Python 3.10 with CUDA 11.4. Datasets were scaled between zero and one, and split into training and validation data (0.75/0.25). The model was compiled using the Adam optimizer (Kingma and Ba, 2014) with AMSGrad (Reddi et al., 2019). We trained the classifier using a learning rate η = 2E–5 and batch size of 128 on the training data. Training was run for maximum 100 epochs with early stopping to avoid overfitting. Validation data was used to evaluate training performance. Early stopping was applied when the validation loss did not improve for eight consecutive epochs. Typically, training lasted for 20–40 epochs. We used binary cross-entropy loss and binary accuracy as measures of performance during training. The best performing model was selected from a training run, and a receiver-operating characteristic (ROC) was calculated. We used accuracy and area under the ROC curve to evaluate training performance. To accelerate training, we used a GPU; training time for the neural network was ∼8 min on a workstation with NVIDIA Tesla P100 16 GB GPU (Intel Xeon 2.2 GHz CPU, 16 GB RAM).

To analyze the discriminative data segments, we calculated per-sample saliency maps with SmoothGrad using the Python package tf-keras-viz (Simonyan et al., 2013). Saliency maps were smoothed with a 10-point running average filter for display purposes. To illustrate the transformation by the model, we performed dimensionality reduction of the training data by Uniform Manifold Approximation and Projection (UMAP), using either the raw dataset or the input to the final layer of the deep learning model.

The trained miniML classifier takes sections of data with predefined length as input. To detect events in arbitrarily long recordings, a sliding window procedure is used. Time-series data from voltage-clamp or current-clamp recordings is segmented using a sliding window with stride. The resulting 2D-tensor is min-max scaled and then used as model input for inference. To overcome the potential limitation of long computation times, we used a sliding window with stride procedure. Using a stride >1 significantly reduces the inference time of the model (Figure 2—figure supplement 1), especially for data with high sampling rates or long recording times. This approach results in a prediction trace with data being spaced at samplinginterval∗stride. With stride=20, a 120 s recording at 50 kHz sampling rate can be analyzed in ∼15 s on a laptop computer (Apple M1Pro, 16 GB RAM, Figure 2—figure supplement 1). To maintain temporal precision, the prediction trace is resampled to the sampling frequency of the raw data.

Events in the input data trace result in distinct peaks in the prediction trace. We applied a maximum filter to enhance post-processing of the data. Thus, by applying a threshold to the prediction trace, synaptic event positions can be detected. Our analyses indicate that the absolute threshold value is not important in the [0.05, 0.95] range (Figure 3—figure supplement 1). Data following a threshold crossing in the prediction trace are cut from the raw data and aligned by steepest rise. To find the point of steepest rise, a peak-search is performed in the first derivative of the short data segment. If multiple peaks are detected, any peak that has a prominence ≥0.25 relative to the largest detected prominence is treated as an additional event that is in close proximity or overlapping. The resulting 2D-array with aligned events can be further analyzed to obtain descriptive statistics on, e.g., amplitude, charge, or rise and decay kinetics of individual events.

We compared the deep learning-based miniML with the following previously described detection methods: template-matching (Clements and Bekkers, 1997), deconvolution (Pernía-Andrade et al., 2012), a finite threshold-based approach (Kudoh and Taguchi, 2002), the commercial MiniAnalysis software (version 6.0.7, Synaptosoft), SimplyFire (Mori et al., 2024), and a Bayesian inference procedure (Merel et al., 2016). Detection methods were implemented in Python 3.9, except for the Bayesian method, which was run using Matlab R2022a, and MiniAnalysis running as stand-alone software on Windows.

We used generated standardized data to benchmark the detection performance of different methods. Event-free recordings (see section Electrophysiological recordings for details) were superimposed with simulated events having a biexponential waveform:I(t)=(1−e(−tτrise))∗e(−tτdecay)

where I(t) is the current as a function of time, and τrise and τdecay are the rise and decay time constants, respectively. Simulated event amplitudes were drawn from a log-normal distribution with variance 0.4. Mean amplitude was varied to generate diverse signal-to-noise ratios (SNR; mean event amplitude / standard deviation of the noise). Decay time constants were drawn from a normal distribution with mean 1.0 ms and variance 0.25 ms (Delvendahl et al., 2019). Generated events were randomly placed in the event-free recording with an average frequency of 0.7 Hz and a minimum spacing of 3 ms. Generated traces provided ground-truth data for the evaluation of the different methods.

To quantify detection performance of different methods over a range of signal-to-noise ratios, we calculated the number of true positives (TP), false positives (FP) and false negatives (FN). From these, the following metrics were calculated:Precision=TPTP+FPRecall=TPTP+FNF1=2∗Precision∗RecallPrecision+Recall

For all metrics, higher values indicate a better model performance. We also evaluated detection performance under non-optimal conditions (i.e., events that did not precisely match the event template or the training data). To do this, we varied the kinetics of simulated events by either increasing (mean = 4.5 ms) or decreasing (mean = 0.5 ms) the average decay time constant (Figure 3—figure supplement 1).

For all benchmarking conditions, threshold settings were: −4 for template-matching, 5∗SD of the detection trace for deconvolution, and −4 pA for the finite-threshold method. Both template-based methods used a template waveform corresponding to the average of the simulated events. For the Bayesian detection approach, we used the code provided by the authors (Merel et al., 2016) with the following adjustments to the default hyperparameters: minimum amplitude = 3.5, noise ϕ=[0.90;−0.52], rate = 0.5. We chose a cutoff of 6∗SD of the event time posteriors as threshold. For SimplyFire, we used the following hyperparameters: direction = −1; kernel = 300; stride = 100; minimum amplitude = 4 pA. MiniAnalysis (Synaptosoft) was run in the automatic detection mode with default settings for EPSCs and amplitude and area thresholds of −4 pA and 4, respectively. We used a minimum peak height of 0.5 and minimum peak width of 10 strides for miniML. To compare different event detection methods in real-world data Figure 4, we used the following detection hyperparameters: We applied miniML with a window size of 600 and minimum peak width of 5 strides. For the Golgi cell data, window size was set to 900 to account for the slower event kinetics. Detection thresholds for template matching and deconvolution thresholds were set at −3.5 and 6, respectively. We adjusted the template for each preparation based on the average waveform extracted with a generic template. MiniAnalysis was run with area and amplitude thresholds of 5; amplitude threshold was increased to −20 pA and −8 pA for Calyx of Held and Golgi cell data, respectively.

To make the miniML method applicable to data with different characteristics, such as different event kinetics, noise characteristics, or recording mode, we used a transfer learning (TL) approach (Pratt et al., 1991). We froze the convolutional layers of the fully trained MF–GC miniML model, resulting in a new model with only the LSTM and dense layers being trainable. Thus, the convolutional layers act as pre-trained feature detectors and much fewer training samples are required. Hyperparameters and training conditions were the same as for full-training (see Training and evaluation) with the following exceptions: learning rate η = 2E–8, patience = 15, batch size = 32, dropout rate = 0.5. The training data were resampled to 600 data points to match the input shape of the original model.

To compare TL with full training, random subsets of the training data (mEPSPs in cerebellar GCs, spontaneous EPSCs in zebrafish, or mEPSCs at the Drosophila neuromuscular junction) with increasing size were generated and used to train models using fivefold cross-validation. We always used the same size of the validation dataset. After comparing TL with full training (Figure 5, Figure 5—figure supplement 2), we trained separate models to analyze the different datasets using subsets of the available training data.

Computing times were quantified using the performance counter function of the built-in Python package time and are given as wall times. Statistical comparisons were made using permutation t-tests with 5000 reshuffles (Ho et al., 2019). Effect sizes are reported as Cohen’s d or mean difference, with 95% confidence intervals obtained by bootstrapping (5000 samples; the confidence interval is bias-corrected and accelerated) (Ho et al., 2019).

Raw data were filtered with a Hann window for event analysis after the detection. The number of points of the filter window was adjusted based on the sampling rate (default, 20 samples). Event amplitudes were quantified as difference between detected event peaks and a short baseline window before event onset. Decay times refer to half decay times (time until signal reaches 50% of the amplitude), and rise times were quantified as 10–90% rise times (time between 10% and 90% of the amplitude). These event statistics were calculated for each individual event and averaged per cell.

All code was implemented in Python version 3.9 or 3.10 (RRID:SCR_008394) with the following libraries: TensorFlow (RRID:SCR_016345), SciPy (RRID:SCR_008058), NumPy (RRID:SCR_008633), Matplotlib (RRID:SCR_008624), Pandas (RRID:SCR_018214), h5py (RRID:SCR_024812), scikit-learn (RRID:SCR_002577), pyABF, dabest (RRID:SCR_022340).