BoYueGRN

Ask a conventional gene regulatory network (GRN) tool about a new single-cell dataset and it starts from nothing: it re-optimizes its parameters against that expression matrix, then returns a ranked list of gene pairs that move together, with no statement about which gene regulates which. Refitting is slow and unstable when genes vastly outnumber cells, and an undirected edge cannot say whether a transcription factor drives a target or merely tracks it.

BoYueGRN moves the fitting offline. Trained once, entirely on 10,000 synthetic datasets generated from structural causal models, it answers any new dataset with a single forward pass emitting both an edge probability and a regulatory direction. The synthetic corpus is what makes direction recoverable: preferential-attachment topology reproduces the scale-free hub structure of real networks while guaranteeing acyclicity, and additive-noise structural equations open a deliberate magnitude gap between injected and observation noise — the asymmetry that separates cause from effect. No real expression profile is seen during training, so every benchmark is out of distribution by construction.

The framework was developed by Jingyue Wu and Ying-Qiang Shen at Sichuan University — West China Hospital and the State Key Laboratory of Oral Diseases at West China Hospital of Stomatology — and posted to bioRxiv in August 2026. It shares the amortized, train-on-priors shape of PerturbPFN while targeting network structure rather than perturbation response, and is positioned against per-dataset learners such as GRNformer and GeneCompass.

#Key Features

• Directed edges as a first-class output: A dedicated direction head scores ordered pairs through separate source and target projections, so the score for i→j is decoupled from j→i by construction rather than oriented from motif or ChIP-seq priors.
• Dual direction experts routed by target type: Separate specialists are trained on TF-to-TF and TF-to-non-TF ground-truth edges, and inference routes each candidate edge by the type of its target gene.
• Whole-transcriptome coverage from a fixed-size model: The network accepts 200 genes at a time; TF-centric sliding windows with asymmetric max fusion tile it across the genome, reaching 99.4–100% gene coverage and averaging out single-window noise.
• Two complementary dependence channels: Each window is summarized by a Ledoit–Wolf shrunk precision matrix for linear conditional dependence and a distance correlation matrix for nonlinear dependence, not by raw expression values.
• DEG-Expanded gene selection: Building the input gene set from differentially expressed genes rather than highly variable genes enlarged the NAFLD input twentyfold with no loss of GO terms.

#Technical Details

Three eight-layer graph Transformer encoders share one architecture — 200 gene nodes, d_model 512, eight attention heads, and edge-biased attention that adds a projection of the pairwise statistics to the standard query-key product. One encoder serves edge existence and is trained from scratch for 20,000 steps with Adam and positive-weighted binary cross-entropy; the two direction encoders are warm-started from its weights and fine-tuned independently, sharing no parameters at the end. The edge model totals about 25.4 million parameters and the asymmetric direction head adds roughly 1.05 million. Zero-shot on the six BEELINE benchmarks, single-window AUROC leads seven baselines on five of six datasets (mean 0.679), and genome-wide window fusion raises the mean to 0.760 with an early precision ratio above 4 everywhere — including mDC, a 383-cell outlier that recovers from 0.472 to 0.791. On two genome-wide CRISPRi Perturb-seq screens, directional accuracy reaches 0.821 (HCT116) and 0.810 (K562), rising to 0.859 and 0.947 on edges retained above probability 0.2, with shuffled controls returning to chance. Dual-expert routing reaches 0.910 directional accuracy against 0.905 for the acyclicity-constrained DAGMA, without any optimization at inference.

#Applications

Because inference needs no retraining or case-specific tuning, the same checkpoint applies to any disease with scRNA-seq data. The authors reconstruct cell-type- and stage-resolved regulatory dynamics across five settings spanning more than 270,000 cells — NAFLD/MASLD progression, hepatocellular carcinoma, periodontitis, proliferative verrucous leukoplakia, and Alzheimer's disease — recovering known pathological programs and nominating testable regulators that surface only at cell-type resolution. Groups working with disease atlases can treat directed network reconstruction as a routine annotation step rather than a per-study modeling project.

#Impact

BoYueGRN's contribution is a paradigm claim as much as a benchmark one: that directed GRN inference can be amortized into a reusable checkpoint, opening the way to atlas-scale regulatory mapping. Code is Apache-2.0 on GitHub, and the checkpoints — a four-seed edge ensemble, five direction specialists, and seven ablation models, about 1.6 GB in total — are Apache-2.0 on HuggingFace. The work is a preprint and has not been peer reviewed, and because training is wholly synthetic, accuracy on real data rests on how well the simulated causal distribution spans real regulation. Performance also depends on the tiling: at the default 200-gene window, directional accuracy on the Perturb-seq screens falls below chance, and low cell-to-gene ratios need a rank-deficiency guard on correlation-based target preselection, as the authors found on the 758-cell hESC benchmark.