Spatial Dropout

Spatial dropout creates an excess of zeros beyond what standard Poisson or Negative Binomial models predict. This arises from two distinct sources: biological zeros (the gene is truly not expressed) and technical zeros (the transcript is present but not captured). The capture efficiency in spatial assays typically ranges from 10-40%, meaning 60-90% of molecules go undetected. This zero-inflation violates assumptions of standard differential expression tools and inflates false negatives in downstream analyses.

Multiple physical and chemical factors contribute to spatial dropout:

• Permeabilization inefficiency: Incomplete tissue digestion prevents mRNA release from cells to the capture surface
• Diffusion constraints: Transcripts may diffuse laterally before binding to spatially barcoded oligonucleotides
• Reverse transcription bias: GC-rich or highly structured RNA sequences resist efficient reverse transcription
• Optical crowding: In imaging-based methods like MERFISH or ISS, overlapping fluorescent signals cause missed detections
• Probe hybridization failure: Target sequence polymorphisms or secondary structures prevent probe binding

To account for spatial dropout, specialized models extend standard count distributions:

• Zero-Inflated Negative Binomial (ZINB): Adds a Bernoulli dropout component to the NB distribution, modeling the probability that a zero comes from a failed capture event
• Hurdle models: Separate the process into a binary presence/absence component and a positive count component, explicitly modeling the detection threshold
• Bayesian hierarchical models: Place priors on per-gene and per-spot capture rates, allowing uncertainty quantification around dropout probabilities
• Deep generative models: Variational autoencoders like scVI learn latent representations that disentangle biological signal from technical dropout noise

Several computational methods aim to recover the true expression landscape:

• Spatial smoothing: Borrowing information from neighboring spots via Gaussian processes or graph-based diffusion to fill in dropout zeros
• Markov Random Fields: Model spatial dependencies in the dropout process itself, recognizing that adjacent spots often share similar capture efficiencies
• Multi-modal integration: Using paired histology images or co-registered scRNA-seq data to inform which zeros are likely technical artifacts
• Gene-gene correlation transfer: Leveraging co-expression patterns learned from high-quality reference datasets to predict missing values in spatial data
• Deep learning imputation: Methods like Sprod and Tangram use neural networks trained on matched spatial and dissociated data to de-noise spatial measurements

Uncorrected spatial dropout systematically biases biological conclusions:

• Spatially variable gene detection: Dropout inflates false negatives, causing genuinely spatially patterned genes to appear uniformly expressed
• Cell-type deconvolution: Missing transcripts distort cell-type proportion estimates, particularly for low-abundance cell types
• Ligand-receptor analysis: Dropout in either the ligand or receptor gene can break true signaling relationships, reducing sensitivity of cell-cell communication inference
• Trajectory inference: Gaps in gene expression along spatial gradients disrupt pseudotime ordering algorithms
• Differential expression: Power is severely reduced for lowly expressed genes, which are precisely those most affected by dropout

Despite the shared name, spatial dropout in transcriptomics is fundamentally different from the dropout regularization used in deep learning. In neural networks, dropout is an intentional training strategy that randomly deactivates neurons to prevent overfitting. In spatial transcriptomics, dropout is an unwanted measurement artifact that obscures true biological signal. The confusion arises because both produce 'missing' values, but the underlying mechanisms and goals are opposite: one is a deliberate model design choice, the other is a physical limitation of the assay that must be modeled and corrected.

A two-component mixture model that is the standard for modeling spatial dropout. It combines a negative binomial (NB) distribution for true expression counts with a point mass at zero for technical failures. The model estimates the probability that a zero comes from the dropout process versus true biological absence, making it essential for differential expression testing in sparse spatial data.

A probabilistic graphical model used to smooth spatial gene expression by encoding the assumption that neighboring cells or spots share similar states. In the context of dropout, an MRF prior is often layered on top of a ZINB model to inform the dropout probability based on the expression state of adjacent locations, effectively borrowing statistical strength from the tissue architecture.

Computational techniques that recover missing transcript counts caused by dropout by leveraging information from similar cells or genes. Common approaches include:

• Matrix factorization: Decomposing the expression matrix into latent factors.
• k-Nearest Neighbors (kNN): Averaging expression from similar spatial neighbors.
• Deep generative models: Using variational autoencoders to learn the underlying data distribution and reconstruct missing values.

The probability of successfully capturing a transcript that is physically present at a location. This is the inverse of the dropout rate. Sensitivity varies significantly across different spatial technologies (e.g., imaging-based vs. sequencing-based) and even across individual genes within the same experiment, often correlating with a gene's mean expression level and local tissue context.

A statistical measure of the degree to which a variable's values at nearby locations are more similar than expected by random chance. In spatial transcriptomics, Moran's I is frequently used to identify spatially variable genes (SVGs). High dropout rates can artificially deflate autocorrelation statistics, requiring robust models that can distinguish true spatial patterning from technical noise.

A computational process that estimates the proportions of different cell types within a spatial transcriptomics spot by separating the mixed gene expression signal. Dropout events complicate deconvolution by creating sparse input vectors that can mislead standard algorithms. Specialized deconvolution tools incorporate dropout-aware likelihood models to improve cell-type proportion estimates.