Jonathan W. Pillow
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Jonathan W. Pillow has a total of 346 co-authors. Here are the top co-authors by number of shared publications:
- Adam S. Charles
Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD
10 shared publications - Carlos D. Brody
Princeton Neuroscience Institute, Princeton University, Princeton, NJ, 08544, USA
7 shared publications - Zoe C. Ashwood
Princeton Neuroscience Institute, Princeton University, Princeton, USA
7 shared publications - Jacob Lachenmyer Yates
University of California, Berkeley
6 shared publications - Alexander Christopher Huk
Departments of Ophthalmology and Psychiatry & Behavioral Sciences, Fuster Laboratory for Cognitive Neuroscience, UCLA, Los Angeles, 90095, CA, USA
5 shared publications - Eero P. Simoncelli
Flatiron Institute, Simons Foundation
5 shared publications
Latest Publications
- Open Access
- 4 November 2026
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 22 (11) , e1014894
Bayesian efficient coding
The efficient coding hypothesis, which proposes that neurons are optimized to maximize information about the environment, has provided a guiding theoretical framework for sensory and systems neuroscience. More recently, a theory known as the Bayesian Brain hypothesis has focused on the brain’s ability to integrate sensory and prior sources of information in order to perform Bayesian inference. Although pieces of a connection between these two hypotheses have appeared in prior work, a general formulation that treats the optimality criterion as an arbitrary functional of the posterior distribution – and thereby admits both information-based and other objectives within a single framework – has remained largely implicit. Here we make this formulation explicit, developing a Bayesian theory of efficient coding that defines Bayesian efficient codes in terms of four basic ingredients: (1) a stimulus prior distribution; (2) an encoding model; (3) a capacity constraint, specifying a neural resource limit; and (4) a loss functional, quantifying the desirability or undesirability of various posterior distributions. Classic efficient codes arise as the special case in which the loss functional is the posterior entropy, leading to a code that maximizes mutual information, but alternate loss functionals give solutions that differ dramatically from information-maximizing codes. Within this framework we introduce covtropy , a novel family of losses defined on posterior distributions and parameterized by a single exponent, and use it to show that decorrelation of sensory inputs – optimal under classic efficient codes in low-noise settings – can be disadvantageous for objectives that penalize large errors. We then reanalyze Laughlin’s seminal data on contrast coding in the blowfly large monopolar cell, and find that the measured response nonlinearity is better explained by minimizing L p reconstruction error with p = 1/2 than by information maximization, overturning a forty-year-old interpretation. Bayesian efficient coding thus enlarges the family of codes that are optimal under different objectives and provides a more general framework for understanding the design principles of sensory systems.
- 4 November 2026
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 22 (11) , e1014894
Open AccessBayesian efficient coding
The efficient coding hypothesis, which proposes that neurons are optimized to maximize information about the environment, has provided a guiding theoretical framework for sensory and systems neuroscience. More recently, a theory known as the Bayesian Brain hypothesis has focused on the brain’s ability to integrate sensory and prior sources of information in order to perform Bayesian inference. Although pieces of a connection between these two hypotheses have appeared in prior work, a general formulation that treats the optimality criterion as an arbitrary functional of the posterior distribution – and thereby admits both information-based and other objectives within a single framework – has remained largely implicit. Here we make this formulation explicit, developing a Bayesian theory of efficient coding that defines Bayesian efficient codes in terms of four basic ingredients: (1) a stimulus prior distribution; (2) an encoding model; (3) a capacity constraint, specifying a neural resource limit; and (4) a loss functional, quantifying the desirability or undesirability of various posterior distributions. Classic efficient codes arise as the special case in which the loss functional is the posterior entropy, leading to a code that maximizes mutual information, but alternate loss functionals give solutions that differ dramatically from information-maximizing codes. Within this framework we introduce covtropy , a novel family of losses defined on posterior distributions and parameterized by a single exponent, and use it to show that decorrelation of sensory inputs – optimal under classic efficient codes in low-noise settings – can be disadvantageous for objectives that penalize large errors. We then reanalyze Laughlin’s seminal data on contrast coding in the blowfly large monopolar cell, and find that the measured response nonlinearity is better explained by minimizing L p reconstruction error with p = 1/2 than by information maximization, overturning a forty-year-old interpretation. Bayesian efficient coding thus enlarges the family of codes that are optimal under different objectives and provides a more general framework for understanding the design principles of sensory systems.
- Open Access
- 4 November 2026
- Published by Springer Nature in Nature
- p. 1-12
Aggression experience and observation drive shared behavioural and neural changes
- 4 November 2026
- Published by Springer Nature in Nature
- p. 1-12
Open AccessAggression experience and observation drive shared behavioural and neural changes
- Open Access
- 6 October 2026
- Published by eLife Sciences Publications, Ltd in eLife preprints
- Vol. 15
Improved inference of latent neural states from calcium imaging data
Calcium imaging (CI) is a standard method for recording neural population activity, as it enables simultaneous recording of hundreds-to-thousands of individual somatic signals. Accordingly, CI recordings are prime candidates for population-level latent variable analyses, for example, using models such as Gaussian Process Factor Analysis (GPFA), hidden Markov models (HMMs), and latent dynamical systems. However, these models have been primarily developed and fine-tuned for electrophysiological measurements of spiking activity. To adapt these models for use with the calcium signals recorded with CI, per-neuron fluorescence time-traces are typically either de-convolved to approximate spiking events or analyzed directly under Gaussian observation assumptions. The former approach, while enabling the direct application of latent variable methods developed for spiking data, suffers from the imprecise nature of spike estimation from CI. Moreover, isolated spikes can be undetectable in the fluorescence signal, creating additional uncertainty. A more direct model linking observed fluorescence to latent variables would account for these sources of uncertainty. Here, we develop accurate and tractable models for characterizing the latent structure of neural population activity from CI data. We propose to augment HMM, GPFA, and dynamical systems models with a CI observation model that consists of latent Poisson spiking and autoregressive calcium dynamics. Importantly, this model is both more flexible and directly compatible with standard methods for fitting latent models of neural dynamics. We demonstrate that using this more accurate CI observation model improves latent variable inference and model fitting on both CI observations generated using state-of-the-art biophysical simulations and imaging data recorded in an experimental setting. We expect the developed methods to be widely applicable to many different analyses of population CI data.
- 6 October 2026
- Published by eLife Sciences Publications, Ltd in eLife preprints
- Vol. 15
Open AccessImproved inference of latent neural states from calcium imaging data
Calcium imaging (CI) is a standard method for recording neural population activity, as it enables simultaneous recording of hundreds-to-thousands of individual somatic signals. Accordingly, CI recordings are prime candidates for population-level latent variable analyses, for example, using models such as Gaussian Process Factor Analysis (GPFA), hidden Markov models (HMMs), and latent dynamical systems. However, these models have been primarily developed and fine-tuned for electrophysiological measurements of spiking activity. To adapt these models for use with the calcium signals recorded with CI, per-neuron fluorescence time-traces are typically either de-convolved to approximate spiking events or analyzed directly under Gaussian observation assumptions. The former approach, while enabling the direct application of latent variable methods developed for spiking data, suffers from the imprecise nature of spike estimation from CI. Moreover, isolated spikes can be undetectable in the fluorescence signal, creating additional uncertainty. A more direct model linking observed fluorescence to latent variables would account for these sources of uncertainty. Here, we develop accurate and tractable models for characterizing the latent structure of neural population activity from CI data. We propose to augment HMM, GPFA, and dynamical systems models with a CI observation model that consists of latent Poisson spiking and autoregressive calcium dynamics. Importantly, this model is both more flexible and directly compatible with standard methods for fitting latent models of neural dynamics. We demonstrate that using this more accurate CI observation model improves latent variable inference and model fitting on both CI observations generated using state-of-the-art biophysical simulations and imaging data recorded in an experimental setting. We expect the developed methods to be widely applicable to many different analyses of population CI data.
- Open Access
PIANO: Probabilistic Inference Autoencoder Networks for multi-Omics enables robust generative modeling of gene expression and scales single-cell integration to 100 million cells
SUMMARY: Single-cell RNA technologies enable the routine acquisition of transcriptomic atlases. However, these molecular profiles are influenced by overlapping sources of variation. Since these covariates confound comparisons, data integration is the first step in most analyses. Three challenges remain: correcting strong batch effects, scaling to millions of cells, and modeling how covariates influence gene expression. To address these challenges, we developed PIANO: Probabilistic Inference Autoencoder Networks for multi-Omics, a deep learning framework whose central feature is a generative model of gene expression data. Additionally, PIANO achieves robust integrations and trains 10x faster than previous methods. PIANO accurately integrates single-cell data across species and across single-cell and spatial transcriptomics modalities. As practical applications, PIANO models spatially-resolved gene expression during Alzheimer’s disease progression in human brains and integrates over 100 million cancer cells to model drug perturbations. In summary, PIANO’s integration and generative modeling capabilities will empower novel insights for countless future studies.
Open AccessPIANO: Probabilistic Inference Autoencoder Networks for multi-Omics enables robust generative modeling of gene expression and scales single-cell integration to 100 million cells
SUMMARY: Single-cell RNA technologies enable the routine acquisition of transcriptomic atlases. However, these molecular profiles are influenced by overlapping sources of variation. Since these covariates confound comparisons, data integration is the first step in most analyses. Three challenges remain: correcting strong batch effects, scaling to millions of cells, and modeling how covariates influence gene expression. To address these challenges, we developed PIANO: Probabilistic Inference Autoencoder Networks for multi-Omics, a deep learning framework whose central feature is a generative model of gene expression data. Additionally, PIANO achieves robust integrations and trains 10x faster than previous methods. PIANO accurately integrates single-cell data across species and across single-cell and spatial transcriptomics modalities. As practical applications, PIANO models spatially-resolved gene expression during Alzheimer’s disease progression in human brains and integrates over 100 million cancer cells to model drug perturbations. In summary, PIANO’s integration and generative modeling capabilities will empower novel insights for countless future studies.
- Open Access
- 4 August 2026
- Published by eLife Sciences Publications, Ltd in eLife preprints
- Vol. 15
Continuous partitioning of neuronal variability
Neurons exhibit substantial trial-to-trial variability in response to repeated stimuli, posing a major challenge for understanding the information content of neural spike trains. In the visual cortex, responses show greater-than-Poisson variability, whose origins and structure remain unclear. To address this puzzle, we introduce a continuous, doubly stochastic model of spike train variability that partitions neural responses into a smooth stimulus-driven component and a time-varying stochastic gain process. We applied this model to spike trains from four visual areas (LGN, V1, V2, and MT) and found that the gain process is well described by an exponentiated power law, with increasing amplitude and slower decay at higher levels of the visual hierarchy. The model also provides analytical expressions for the Fano factor of binned spike counts as a function of timescale, linking observed variability to underlying modulatory dynamics. Together, these results establish a principled framework for characterizing neural variability across cortical processing stages.
- 4 August 2026
- Published by eLife Sciences Publications, Ltd in eLife preprints
- Vol. 15
Open AccessContinuous partitioning of neuronal variability
Neurons exhibit substantial trial-to-trial variability in response to repeated stimuli, posing a major challenge for understanding the information content of neural spike trains. In the visual cortex, responses show greater-than-Poisson variability, whose origins and structure remain unclear. To address this puzzle, we introduce a continuous, doubly stochastic model of spike train variability that partitions neural responses into a smooth stimulus-driven component and a time-varying stochastic gain process. We applied this model to spike trains from four visual areas (LGN, V1, V2, and MT) and found that the gain process is well described by an exponentiated power law, with increasing amplitude and slower decay at higher levels of the visual hierarchy. The model also provides analytical expressions for the Fano factor of binned spike counts as a function of timescale, linking observed variability to underlying modulatory dynamics. Together, these results establish a principled framework for characterizing neural variability across cortical processing stages.
- Open Access
- 17 June 2026
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 123 (25)
State-switching navigation strategies in Caenorhabditis elegans are beneficial for chemotaxis
Animals employ different strategies for relating sensory input and behavioral output to navigate sensory environments, but what strategy to use, when to switch and why remain unclear. Caenorhabditis elegans navigate by combining “steering” (small heading changes) with “turn” (large reorientations). It is unknown whether transitions between these elements are driven solely by sensory input or also by persistent internal states. It is also unclear how worms sometimes appear to exit turns such that they are already oriented toward a goal, despite their presumed lack of spatial awareness during turns. We address these questions with measurements of sensory-guided navigation and a statistical model of state-dependent control. Worm navigation is well described by a sensory-driven, two-state-switching model whose states persist for seconds and produce distinct sensorimotor mixtures: one state is steer-enriched, the other turn-enriched. This hierarchical temporal organization challenges the view that gradient-climbing strategies are static and purely stimulus-locked. Instead, sensory input causally modulates transitions between persistent states, creating the appearance of “directed turns” when exiting the turn-enriched state. Measurements using genetically perturbed animals and modeling with data-constrained reinforcement-learning both show that state switching enhances gradient-climbing performance. Together, measurement, perturbation, and modeling reveal that state switching is functionally beneficial, organizing behavior across time—a principle that may generalize across species and contexts.
- 17 June 2026
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 123 (25)
Open AccessState-switching navigation strategies in Caenorhabditis elegans are beneficial for chemotaxis
Animals employ different strategies for relating sensory input and behavioral output to navigate sensory environments, but what strategy to use, when to switch and why remain unclear. Caenorhabditis elegans navigate by combining “steering” (small heading changes) with “turn” (large reorientations). It is unknown whether transitions between these elements are driven solely by sensory input or also by persistent internal states. It is also unclear how worms sometimes appear to exit turns such that they are already oriented toward a goal, despite their presumed lack of spatial awareness during turns. We address these questions with measurements of sensory-guided navigation and a statistical model of state-dependent control. Worm navigation is well described by a sensory-driven, two-state-switching model whose states persist for seconds and produce distinct sensorimotor mixtures: one state is steer-enriched, the other turn-enriched. This hierarchical temporal organization challenges the view that gradient-climbing strategies are static and purely stimulus-locked. Instead, sensory input causally modulates transitions between persistent states, creating the appearance of “directed turns” when exiting the turn-enriched state. Measurements using genetically perturbed animals and modeling with data-constrained reinforcement-learning both show that state switching enhances gradient-climbing performance. Together, measurement, perturbation, and modeling reveal that state switching is functionally beneficial, organizing behavior across time—a principle that may generalize across species and contexts.
- Top Cited Papers
- 25 February 2026
- Published by Springer Nature in Nature
- Vol. 652 (8111) , 947-954
Compact deep neural network models of the visual cortex
A powerful approach to understand the computations carried out by the visual cortex is to build models that predict neural responses to any arbitrary image. Deep neural networks (DNNs) have emerged as the leading predictive models1,2, yet their underlying computations remain buried beneath millions of parameters. Here we challenge the need for models at this scale by seeking predictive and parsimonious DNN models of the primate visual cortex. We first built a highly predictive DNN model of neural responses in macaque visual area V4 by alternating data collection and model training in adaptive closed-loop experiments. We then compressed this large, black-box DNN model, which comprised 60 million parameters, to identify compact models with 5,000 times fewer parameters yet comparable accuracy. This dramatic compression enabled us to investigate the inner workings of the compact models. We discovered a salient computational motif: compact models share similar filters in early processing, but individual models then specialize their feature selectivity by ‘consolidating’ this shared high-dimensional representation in distinct ways. We examined this consolidation step in a dot-detecting model neuron, revealing a computational mechanism that leads to a testable circuit hypothesis for dot-selective V4 neurons. Beyond V4, we found strong model compression for macaque visual areas V1 and IT (inferior temporal cortex), revealing a general computational principle of the visual cortex. Overall, our work challenges the notion that large DNNs are necessary to predict individual neurons and establishes a modelling framework that balances prediction and parsimony.
- 25 February 2026
- Published by Springer Nature in Nature
- Vol. 652 (8111) , 947-954
Top Cited PapersCompact deep neural network models of the visual cortex
A powerful approach to understand the computations carried out by the visual cortex is to build models that predict neural responses to any arbitrary image. Deep neural networks (DNNs) have emerged as the leading predictive models1,2, yet their underlying computations remain buried beneath millions of parameters. Here we challenge the need for models at this scale by seeking predictive and parsimonious DNN models of the primate visual cortex. We first built a highly predictive DNN model of neural responses in macaque visual area V4 by alternating data collection and model training in adaptive closed-loop experiments. We then compressed this large, black-box DNN model, which comprised 60 million parameters, to identify compact models with 5,000 times fewer parameters yet comparable accuracy. This dramatic compression enabled us to investigate the inner workings of the compact models. We discovered a salient computational motif: compact models share similar filters in early processing, but individual models then specialize their feature selectivity by ‘consolidating’ this shared high-dimensional representation in distinct ways. We examined this consolidation step in a dot-detecting model neuron, revealing a computational mechanism that leads to a testable circuit hypothesis for dot-selective V4 neurons. Beyond V4, we found strong model compression for macaque visual areas V1 and IT (inferior temporal cortex), revealing a general computational principle of the visual cortex. Overall, our work challenges the notion that large DNNs are necessary to predict individual neurons and establishes a modelling framework that balances prediction and parsimony.
- Open Access
- 22 January 2026
- Published by eLife Sciences Publications, Ltd in eLife preprints
Improved inference of latent neural states from calcium imaging data
Calcium imaging (CI) is a standard method for recording neural population activity, as it enables simultaneous recording of hundreds-to-thousands of individual somatic signals. Accordingly, CI recordings are prime candidates for population-level latent variable analyses, for example using models such as Gaussian Process Factor Analysis (GPFA), hidden Markov models (HMMs), and latent dynamical systems. However, these models have been primarily developed and fine-tuned for electrophysiological measurements of spiking activity. To adapt these models for use with the calcium signals recorded with CI, per-neuron fluorescence time-traces are typically either de-convolved to approximate spiking events or analyzed directly under Gaussian observation assumptions. The former approach, while enabling the direct application of latent variable methods developed for spiking data, suffers from the imprecise nature of spike estimation from CI. Moreover, isolated spikes can be undetectable in the fluorescence signal, creating additional uncertainty. A more direct model linking observed fluorescence to latent variables would account for these sources of uncertainty. Here, we develop accurate and tractable models for characterizing the latent structure of neural population activity from CI data. We propose to augment HMM, GPFA, and dynamical systems models with a CI observation model that consists of latent Poisson spiking and autoregressive calcium dynamics. Importantly, this model is both more flexible and directly compatible with standard methods for fitting latent models of neural dynamics. We demonstrate that using this more accurate CI observation model improves latent variable inference and model fitting on both CI observatons generated using state-of-the-art biophysical simulations as well as imaging data recorded in an experimental setting. We expect the developed methods to be widely applicable to many different analysis of population CI data.
- 22 January 2026
- Published by eLife Sciences Publications, Ltd in eLife preprints
Open AccessImproved inference of latent neural states from calcium imaging data
Calcium imaging (CI) is a standard method for recording neural population activity, as it enables simultaneous recording of hundreds-to-thousands of individual somatic signals. Accordingly, CI recordings are prime candidates for population-level latent variable analyses, for example using models such as Gaussian Process Factor Analysis (GPFA), hidden Markov models (HMMs), and latent dynamical systems. However, these models have been primarily developed and fine-tuned for electrophysiological measurements of spiking activity. To adapt these models for use with the calcium signals recorded with CI, per-neuron fluorescence time-traces are typically either de-convolved to approximate spiking events or analyzed directly under Gaussian observation assumptions. The former approach, while enabling the direct application of latent variable methods developed for spiking data, suffers from the imprecise nature of spike estimation from CI. Moreover, isolated spikes can be undetectable in the fluorescence signal, creating additional uncertainty. A more direct model linking observed fluorescence to latent variables would account for these sources of uncertainty. Here, we develop accurate and tractable models for characterizing the latent structure of neural population activity from CI data. We propose to augment HMM, GPFA, and dynamical systems models with a CI observation model that consists of latent Poisson spiking and autoregressive calcium dynamics. Importantly, this model is both more flexible and directly compatible with standard methods for fitting latent models of neural dynamics. We demonstrate that using this more accurate CI observation model improves latent variable inference and model fitting on both CI observatons generated using state-of-the-art biophysical simulations as well as imaging data recorded in an experimental setting. We expect the developed methods to be widely applicable to many different analysis of population CI data.
- Open Access
- 27 November 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1)
Identifying the factors governing internal state switches during nonstationary sensory decision-making
Traditional models of perceptual decision-making fail to capture dynamic strategy switching in non-stationary environments, and the factors governing these switches remain unknown. To address this gap, we developed an advanced internal state model with input-driven transitions and observations. Our approach employs a hidden Markov model (HMM) coupled with two sets of per-state generalized linear models (GLMs): a Bernoulli GLM for state- and stimulus-dependent choices, and a multinomial GLM for input-dependent transitions between states. We applied our model to a decision-making task in a non-stationary environment, analyzing hundreds of thousands of trials from a cohort of mice, and found that their behavior can be accurately described by a four-state model. This model identified two engaged states with low biases relative to the stimulus and two disengaged states with pronounced biases relative to the stimulus. Our analyses revealed that mice preferentially used left-bias strategies during left-bias stimulus blocks, and right-bias strategies during right-bias stimulus blocks, achieving high performance even in disengaged states by biasing choices toward the side with greater prior probability. Our model showed that past choices and past stimuli predicted transitions between left- and right-bias states, while past rewards predicted transitions between engaged and disengaged states. In particular, greater past reward predicted transition to disengaged states, suggesting that disengagement may be associated with satiety. Our approach uncovers links between animal behavior, input regressors, and state transitions, highlighting the complexity of adaptive strategies. This provides a foundation for future research in dynamic decision-making models.
- 27 November 2025
- Published by Springer Nature in Nature Communications
- Vol. 16 (1)
Open AccessIdentifying the factors governing internal state switches during nonstationary sensory decision-making
Traditional models of perceptual decision-making fail to capture dynamic strategy switching in non-stationary environments, and the factors governing these switches remain unknown. To address this gap, we developed an advanced internal state model with input-driven transitions and observations. Our approach employs a hidden Markov model (HMM) coupled with two sets of per-state generalized linear models (GLMs): a Bernoulli GLM for state- and stimulus-dependent choices, and a multinomial GLM for input-dependent transitions between states. We applied our model to a decision-making task in a non-stationary environment, analyzing hundreds of thousands of trials from a cohort of mice, and found that their behavior can be accurately described by a four-state model. This model identified two engaged states with low biases relative to the stimulus and two disengaged states with pronounced biases relative to the stimulus. Our analyses revealed that mice preferentially used left-bias strategies during left-bias stimulus blocks, and right-bias strategies during right-bias stimulus blocks, achieving high performance even in disengaged states by biasing choices toward the side with greater prior probability. Our model showed that past choices and past stimuli predicted transitions between left- and right-bias states, while past rewards predicted transitions between engaged and disengaged states. In particular, greater past reward predicted transition to disengaged states, suggesting that disengagement may be associated with satiety. Our approach uncovers links between animal behavior, input regressors, and state transitions, highlighting the complexity of adaptive strategies. This provides a foundation for future research in dynamic decision-making models.
- Open Access
Fast Optimization of Robust Transcriptomics Embeddings using Probabilistic Inference Autoencoder Networks for multi-Omics
SUMMARY: Advances in single-cell genomics technologies enable the routine acquisition of atlases with millions of cells. These datasets often include multiple covariates, such as donors, sequencing platforms, developmental timepoints, and species, which provide new opportunities for discovery and new challenges. To mitigate unwanted sources of variation, dataset integration is the starting point for most analyses. However, existing methods struggle with integrating large complex datasets. To address these limitations, we developed PIANO, a variational autoencoder framework that uses a negative binomial generalized linear model for stronger batch correction, and code compilation for ten times faster training than existing tools. We first demonstrate performant integration compared to commonly used methods on single-species datasets. We then show PIANO enables superior analyses of multiple atlases, solving challenging integration tasks across sequencing platforms, development, and species, while simultaneously preserving desired biological signals. Our contributions include a novel, high-performance integration method and recommendations for integration applications.
Open AccessFast Optimization of Robust Transcriptomics Embeddings using Probabilistic Inference Autoencoder Networks for multi-Omics
SUMMARY: Advances in single-cell genomics technologies enable the routine acquisition of atlases with millions of cells. These datasets often include multiple covariates, such as donors, sequencing platforms, developmental timepoints, and species, which provide new opportunities for discovery and new challenges. To mitigate unwanted sources of variation, dataset integration is the starting point for most analyses. However, existing methods struggle with integrating large complex datasets. To address these limitations, we developed PIANO, a variational autoencoder framework that uses a negative binomial generalized linear model for stronger batch correction, and code compilation for ten times faster training than existing tools. We first demonstrate performant integration compared to commonly used methods on single-species datasets. We then show PIANO enables superior analyses of multiple atlases, solving challenging integration tasks across sequencing platforms, development, and species, while simultaneously preserving desired biological signals. Our contributions include a novel, high-performance integration method and recommendations for integration applications.