Graph neural networks (GNNs) are models that update each node by pooling features from its neighbors and then passing the result through a small dense network. They are now used to recover which cell types make up a bulk tissue sample (deconvolution) and to repair missing values in single-cell data. These biological graphs are getting large. Reference atlases such as the 3CA collection hold millions of cells. Specialized GNN chips have cut the memory traffic that dominates GNN runtime by storing node features at very low precision, and the most effective policy gives more bits to well-connected "hub" nodes and as few as 2 bits to sparsely connected ones. On a cell-similarity graph, that policy has a predictable failure. The rare cells a biologist most wants to see, such as stem-like tumor cells or unusual macrophage subsets, sit on the sparse fringe of the graph, so the hardware would compress them hardest. Average accuracy would barely change while the rare populations vanish. Compression would bring back the very masking that single-cell methods were built to undo.
This note proposes to measure that effect and then fix it. The fix keeps existing accelerator hardware unchanged and changes only the rule that assigns bit-widths to nodes. The rule is computed without labels, from how distinctive each node is relative to its neighborhood. The key question is whether such a rule beats both the degree-based rule and its obvious inversion at equal memory budgets, on rare-population metrics rather than averages.
Core Claim
Two claims follow, and they can be tested separately. The first is a measurement claim. When GNNs that analyze cell graphs are quantized to the 2–4 bit range, the resulting error falls disproportionately on small cell populations. Aggregate metrics such as mean concordance across cell types or overall classification accuracy hide this. Degree-aware mixed precision makes the problem worse, because rare cells are low-in-degree "anti-hubs" in k-nearest-neighbor similarity graphs. The second is a method claim. A per-node precision rule driven by neighborhood distinctiveness, rather than degree, preserves rare-population accuracy at the same average bit budget and the same hardware cost.
The tension comes directly from prior work. MEGA observed that a node's in-degree tracks the dynamic range of its aggregated features. It learns to give hubs around 8 bits and low-degree nodes around 2. It reports 38.3× speedup and 47.6× energy savings over HyGCN with up to 18.6× compression relative to FP32. All of this was evaluated on citation and social graphs, where low-degree nodes are seldom what the user cares about. In deconvolution, the user cares about exactly those nodes. Work on rare therapy-resistant and stem-like subpopulations (tamoxifen-resistance stem-like cells; conserved stemness profile across cancers; HCC macrophage subsets) shows that the clinically actionable signal is often a small fraction of cells. Results from LLM quantization motivate the specific mechanism we expect. CherryQ found that under 1% of parameters carry most of the quantization sensitivity, and that protecting them avoids the 2-bit collapse. Our hypothesis is the node-level analogue: a small, identifiable set of nodes carries most of the biologically meaningful sensitivity, and degree is the wrong way to find them.
Research Program
The study has three parts.
Stratified damage curves. First, we characterize where the error lands. We take the GNN deconvolution model GraphDEC, which builds sample-similarity graphs from autoencoder latents and applies a neighborhood-aware GNN. We also take the cell–gene bipartite GraphSAGE and the small GCN from the single-cell GNN imputation benchmark. We quantize each model's weights and post-aggregation activations at 8, 4, 3 and 2 bits under uniform, per-node, and degree-aware schemes. For each setting we report error separately for every cell population, ordered by population size, and for nodes binned by in-degree and by k-occurrence (how many other cells list this cell among their nearest neighbors).
To separate rarity from cell identity, we run controlled spike-in experiments. A chosen cell type is downsampled in the reference to a range of frequencies, and we track how the bit-width at which that type's accuracy collapses shifts as the type becomes rarer.
A label-free precision rule. Second, we build the allocation rule. For node $v$ with hidden feature $h_v$ and neighbor set $N(v)$, we define a distinctiveness score
$$s_v = \frac{\lVert h_v - \bar h_{N(v)} \rVert}{\lVert h_v \rVert + \epsilon},$$
where $\bar h_{N(v)}$ is the mean neighbor feature and $\epsilon$ is a small constant. A large $s_v$ means aggregation will pull the node strongly toward its neighbors. That is the situation in which rounding noise can push a rare cell's representation into the majority type. We combine $s_v$ with k-occurrence into a small lookup table indexed by quantized buckets of both quantities. Bit-widths per bucket are fit once on a calibration split under an average-bits budget, with an objective of minimizing worst-population error.
The table has the same shape as the degree-indexed lookup that TopGQ uses to assign scales to unseen nodes. That makes the rule usable in inductive settings, where new bulk samples arrive at test time. It also lets us reuse TopGQ's absorption of per-node scales into the adjacency matrix, which keeps aggregation in integer arithmetic. In GraphDEC the edges are re-weighted by estimated proportions, but the graph is fixed at inference, so the absorbed scales are recomputed once per inference graph rather than per step. We will measure that recomputation cost.
Hardware cost. Third, we check that the new rule costs nothing at the hardware level. We reimplement MEGA's dataflow in a cycle-level simulator: grouped packages for each bit-width, bit-serial units for the dense step, and reordering of features to make memory access sequential. We then swap only the allocation policy. The packaging hardware is unchanged, so any change in cycles, DRAM traffic, or energy comes from the bit-width distribution and from package fragmentation. We run the simulator on graphs built from 3CA (2.7 million cells from 1,631 samples, as described in POPPY), since this scale is what justifies sub-4-bit storage in the first place.
Baselines
The naive reference points are the full-precision model and uniform INT8. They set the accuracy ceiling and represent what a practitioner would do by default. On the GraphDEC benchmarks, the full-precision ceiling is GraphDEC's reported average concordance correlation coefficient (CCC) of 80.7% on proteomics deconvolution and 93% on cross-species sets (GraphDEC). For cell classification on Paul15 and PBMC3K, the reported reference is 59.44% accuracy for a one-layer GCN with group normalization, against 59.4% for an SVM (GNN imputation benchmark). We will rerun both ourselves to obtain per-population numbers. The file contains no per-population figures, and no reported result was found on Lacuna for any quantized GNN on biological deconvolution or single-cell data.
The strongest existing competitor is MEGA's degree-aware mixed precision. It is the only published method that allocates bits per node and is co-designed with hardware. We reimplement its learned allocation and give it the same average bit budget, the same calibration data, and the same tuning budget, measured in hyperparameter trials and GPU-hours, as our rule. On standard graphs, our reimplementation must come close to MEGA's reported behavior before we trust it on cell graphs. The second strong competitor is TopGQ's per-node post-training quantization, which reports 93.05% accuracy for 4-bit GCN on Reddit while some baselines fell to near zero. We run it at 4, 3 and 2 bits with its lookup indexed by degree.
Two ablations isolate our key ingredient. The inverse-degree rule gives more bits to low-degree nodes. The oracle rule protects nodes by their true rare-population labels and sets the upper bound for any label-free rule. A random-protection control keeps the same fraction of nodes at 8 bits. Beating the baselines means lower worst-population error, measured as the CCC of the smallest population and the recall of the rarest class, than MEGA and TopGQ at equal average bits. The gap must hold across seeds. Average accuracy must stay within one point of INT8, and simulated memory traffic must be no worse than MEGA's.
Why not just…?
Why not just use uniform INT8 on a GPU? For many current deconvolution cohorts this is probably enough, and Tango shows INT8 GNN training keeps over 99% of FP32 accuracy. Our damage curves will report the bit-width at which rare populations start to degrade. If that threshold is at or below 4 bits for every model and dataset, the method claim loses its motivation and the contribution reduces to the measurement study. The case for going lower rests on atlas-scale graphs, where INT8 gives only a 4× reduction in feature memory. The 3CA simulation tests whether that gap actually matters.
Why not just invert the degree rule? This is the most serious threat. In k-nearest-neighbor graphs, low in-degree already marks fringe cells. Low in-degree also marks doublets and noisy cells, however, so inverse-degree may spend bits on junk. Distinctiveness is meant to separate a coherent rare cluster, which has low k-occurrence but is consistent with its few neighbors, from isolated noise. The inverse-degree ablation settles the question directly. If it matches our rule, we will report that, and the practical recommendation becomes the simpler rule.
Why not just reweight rare classes during quantization-aware training? Class-balanced losses need reference labels and act on the weights, not on how each node is stored. At inference on new bulk or spatial samples, nothing tells the hardware which nodes are rare. We will still run class-balanced quantization-aware training combined with MEGA's policy as an extra baseline, because it may close part of the gap.
Why not just apply CherryQ or QuIP? Both protect or rotate dense weight matrices. Our hypothesis places the damage in post-aggregation node activations, where hub-driven outliers have already been documented (TopGQ). We will run CherryQ-style weight protection at 2 bits. If it alone recovers rare-population accuracy, then our activation-side diagnosis is wrong.
Proposed Experiments
The datasets are the GraphDEC benchmarks: synthetic proteomics, cross-species, the Mouse Brain, Pancreas and PBMC transcriptomics sets, and spatial tonsil and PDAC. We add Paul15 and PBMC3K for classification, and spike-in variants of each with controlled rarity. Cora and Reddit serve as checks that the new rule does not hurt standard graphs relative to MEGA.
Accuracy metrics are per-population CCC and recall, worst-population scores, and averages. Hardware metrics are the average bits per node, the fraction of nodes per package, and simulated cycles, DRAM bytes and energy.
There are three decisive checks. The first is whether degree-aware 2–3 bit quantization degrades the rarest populations at least twice as much as common ones while average metrics stay flat; if not, the masking hypothesis fails. The second is whether, at matched bits, the distinctiveness rule closes most of the gap between MEGA and the label oracle on rare populations. The third is whether it does so better than inverse-degree, with simulated traffic within a few percent of MEGA's.
References
- MEGA: A Memory-Efficient GNN Accelerator Exploiting Degree-Aware Mixed-Precision Quantization
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers
- Global immune characterization of HBV/HCV-related hepatocellular carcinoma identifies macrophage and T-cell subsets associated with disease progression
- Cherry on Top: Parameter Heterogeneity and Quantization in Large Language Models
- Deciphering Cell Type Abundance in Proteomics Data Through Graph Neural Networks
- Benchmarking Graph Neural Network-based Imputation Methods on Single-Cell Transcriptomics Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Tango: rethinking quantization for graph neural network training on GPUs