Tumor cells can export long non-coding RNAs inside small membrane-wrapped packets called exosomes. When a neighboring cell absorbs them, these RNAs act as "sponges": they soak up the cell's microRNAs, which are short regulatory RNAs that normally silence growth genes, and so they release those genes from silencing. The statistical test for sponging is conditional mutual information. It asks whether the link between a microRNA and its target gene changes depending on the level of a candidate sponge. Running that test over every microRNA, target, and sponge triplet is too slow to do exhaustively. Running it with a permutation null, at single-cell scale, is out of reach. This note proposes a change in representation that makes it cheap. Discretize each gene's expression into a few levels per cell, with "zero / dropout" as its own level. Store each level as a bitmask over cells. Every count the estimator needs then becomes a bitwise AND followed by a population count (the number of set bits). Wide vector registers on ordinary CPUs execute exactly these two instructions at full width.
The engineering idea is simple. The open question it raises is scientific: does coarse, fixed discretization recover the sponge interactions that slower adaptive estimators find? If it does, two analyses become possible that are not feasible today. The first is exhaustive permutation-tested screening. The second is a spatial version of the test that asks whether a sponge in neighboring cells explains a cell's target derepression beyond the sponge in the cell itself. That second test is the signature of intercellular, exosome-borne sponging rather than ordinary within-cell sponging.
Core Claim
The claim is that discrete conditional mutual information over $L=4$ levels (2 bits) can be computed directly on bit-sliced expression data. Bit-sliced means one bitmask per (gene, level) across $n$ cells. This runs one to two orders of magnitude faster than tree-partition estimators. It should agree with them closely enough on ceRNA calls, meaning calls about competing endogenous RNAs that share microRNAs, that the cheaper estimator can replace them for screening. The speed also leaves room to add permutation nulls and a spatial neighbor-conditioning term that no current tool computes.
The estimator is
$$ \widehat{I}(X;Y\mid Z)=\frac{1}{n}\sum_{x,y,z} c_{xyz}\log\frac{c_{xyz}\,c_{z}}{c_{xz}\,c_{yz}},\qquad c_{xyz}=\mathrm{popcount}\big(B^X_x \wedge B^Y_y \wedge B^Z_z\big), $$
where $B^G_\ell$ is the bitmask of cells in which gene $G$ falls at level $\ell$, $c$ are cell counts, and the lower-order counts are marginals of $c_{xyz}$.
The bitmask layout is the important design choice. It reverses the usual packing choice in low-precision kernels. DeepGEMM packs 2-bit codes side by side and looks up their products in register-sized tables. It reports that unpacking those codes takes "approximately 80% of the execution time in the LUT-Conv stage". Bitplanes need no unpacking at all, because the logical AND is the computation. For the same reason, the dense 5-values-per-byte ternary encoding of xTern is the wrong fit here. It saves memory, but it breaks the AND/popcount property, so that paper acts as a constraint on the design rather than a method to reuse. The idea of mapping continuous values to discrete bins and then doing table work comes from MADDNESS. It hashes subvectors with learned trees at over 100 GB/s per thread. Here the bins are per-gene quantiles rather than learned, and the "table" is a contingency table rather than a set of precomputed products.
Research Program
The first step is to build the kernel. Screening has a natural loop order: fix a microRNA $X$ and a candidate modulator $Z$, then sweep all targets $Y$. Under that order, the $L^2=16$ masks $B^X_x\wedge B^Z_z$ are computed once and reused across every target. Each target then costs $L^3=64$ AND-plus-popcount operations per 512-cell register word. The $c\log c$ terms are read from a precomputed table indexed by count. Permutation nulls get the same treatment. A set of permuted copies of $X$ is packed once and reused across all $(Y,Z)$. This amortization is what turns permutation testing from prohibitive into routine. The kernel will target AVX-512 with hardware popcount and ARM NEON, and will be checked against an exact scalar implementation.
The discretization takes one fixed level for exact zeros. The remaining $L-1$ levels are quantiles of the nonzero values. This mirrors the zero-expression handling that JAMI found necessary to avoid false ceRNA edges caused by dropouts. Bias of the plug-in estimator grows with table size relative to $n$. That bias follows from the standard Miller–Madow correction, which scales as (number of nonzero cells)/$2n$. It sets the practical choice of $L$, and we will test $L=4$ and $L=8$.
Second, a spatial extension. For spatial transcriptomics, each cell's modulator variable is computed in two ways. One is its own sponge level. The other, $Z^{\text{nbr}}_i$, is a $1/d$-weighted sum of the sponge in surrounding spots, excluding spot $i$ itself. This uses the reciprocal-distance decay that stMLnet found to fit spatial signaling better than constant or radial-basis kernels. The test statistic is $I(Z^{\text{nbr}};Y\mid Z^{\text{self}})$, restricted to targets $Y$ of microRNAs that the sponge is known to bind. It asks whether the neighbors' sponge predicts target derepression beyond the cell's own sponge. Standard spatial capture does not measure microRNAs, so this test deliberately avoids needing them. This is a known limitation of the field, and the test is designed around it rather than around a proxy that cannot be validated.
The biology motivating the test is concrete. Exosomal MALAT1 showed that metastatic colorectal cells export MALAT1, which sponges miR-26a/b and releases FUT4. CD44+ leading-edge cells showed that aggressive colorectal cells cluster at the invasive boundary and remodel nearby fibroblasts. The prediction is that neighbor-MALAT1 signal against FUT4 should be strongest at that boundary.
The closest computational work leaves this question untouched. miRTalk scores exosome-microRNA communication on dissociated cells as direct repression, with no sponge layer and no spatial coordinates. stMLnet, COMMOT, and CellChat model only protein ligand–receptor signaling. MulNet includes microRNA–target edges, but treats them as static, non-spatial edges.
Baselines
The strongest competitor is JAMI (JAMI). It computes conditional mutual information with adaptive tree partitioning and multithreading, and is the current fast reference for ceRNA network inference. The file reports its speed only relative to an older tool, an "approximately 70-fold" speedup over CUPID on TCGA breast cancer. It gives no absolute runtimes and no accuracy figures, so on accuracy there is no reported result on Lacuna for JAMI ceRNA call accuracy. JAMI and our kernel will run on the same TCGA breast cancer expression and miRNA data and the same candidate triplets, on the same machine and core count, with each tool's own default permutation or p-value settings.
Beating JAMI means two things together. Wall-clock time must be at least 10× lower at matched core count. Call sets must overlap at Jaccard ≥ 0.8 at matched FDR, and recovery of curated sponge interactions from the literature, including the MALAT1/miR-26/FUT4 axis, must be no lower.
The second baseline is the same discrete estimator implemented as dense histogram counting with standard array code on a GPU. This isolates whether the bitplane layout matters, or whether discretization alone delivers the speed.
The naive reference is the simplest possible screen. For each triplet it uses partial Pearson correlation of microRNA and target given the modulator, with a BLAS implementation. Every method must at least beat this reference in recovering curated interactions.
Two ablations isolate the key ingredients. The first removes the zero level, using plain quantiles of all values. The second removes the amortized permutation reuse. A third comparison, $L=4$ versus $L=8$ versus JAMI's adaptive partitioning, measures what fixed discretization costs in accuracy.
Why not just…?
Why not just run JAMI with more cores? JAMI already scales through threads. Multicore speedup leaves the per-triplet cost unchanged, and with permutation nulls that cost is multiplied by the number of permutations. Bitplanes reduce the per-triplet cost itself and share permutation work across triplets. The head-to-head timing at equal core count, run with and without permutations, settles this directly.
Why not just put the histograms on a GPU? This is a real threat. Once the data are discretized, a well-written GPU histogram may match a CPU bitplane kernel. If it does, the main contribution moves to the discretization fidelity result and the spatial test, which work on either hardware, and the hardware claim is dropped. The GPU baseline above is included specifically to make this call. The strongest version of the CPU argument is memory: 4 bitplanes per gene take 4 bits per cell. That lets whole-atlas datasets fit in a workstation's cache hierarchy, much as packing let Agile-Quant fit LLaMA-7B into about 2.53GB on edge CPUs.
Why not just use partial correlation? It is fast and needs no new code. However, sponging is a titration effect: the target is released only once the sponge exceeds the microRNA, which is a threshold behavior that linear measures handle poorly. Linear measures are also distorted by dropout zeros, which is the motivation JAMI gives for its zero handling. Still, this is an empirical claim. If partial correlation recovers curated interactions as well as either information-based method on TCGA, the information-theoretic machinery is not justified, and we will report that.
Proposed Experiments
The fidelity experiment uses TCGA breast cancer, which has matched miRNA-seq, and compares bitplane calls against JAMI calls as described in the Baselines section.
The scaling experiment uses dissociated tumor cohorts from miRTalk, such as GSE138709 cholangiocarcinoma. Gene expression is subsampled from $10^3$ to $10^5$ cells while the triplet set is held fixed, measuring throughput in triplets per second with 1,000 permutations.
The spatial test uses the CD44+ colorectal Visium data and the Visium datasets used by stMLnet. Its decisive check has two parts. The neighbor-sponge statistic for MALAT1 against miR-26 targets should be significant and concentrated at the annotated leading edge. Under spatial shuffling of spot coordinates, which preserves expression but destroys neighborhoods, the statistic should collapse.
A negative control uses sponges with no curated exosomal evidence. Their neighbor statistics should not exceed those of randomly chosen genes with matched expression levels.
If coarse discretization loses JAMI's calls, or if spatial shuffling fails to kill the neighbor signal, the proposal fails on its own terms.
References
- DeepGEMM: Accelerated Ultra Low-Precision Inference on CPU Architectures using Lookup Tables
- xTern: Energy-Efficient Ternary Neural Network Inference on RISC-V-Based Edge Systems
- Multiplying Matrices Without Multiplying
- JAMI: fast computation of conditional mutual information for ceRNA network analysis
- Modeling and inference of spatial intercellular communications and multilayer signaling regulations using stMLnet
- Exosomal MALAT1 sponges miR-26a/26b to promote the invasion and metastasis of colorectal cancer via FUT4 enhanced fucosylation and PI3K/Akt pathway
- CD44+ cells enhance pro-tumor stroma in the spatial landscape of colorectal cancer leading edge
- Extracellular vesicle-derived miRNA-mediated cell–cell communication inference for single-cell transcriptomic data with miRTalk
- Screening cell–cell communication in spatial transcriptomics via collective optimal transport
- Inference and analysis of cell-cell communication using CellChat
- MulNet: a scalable framework for reconstructing intra- and intercellular signaling networks from bulk and single-cell RNA-seq data
- Agile-Quant: Activation-Guided Quantization for Faster Inference of LLMs on the Edge