scMaize

Maize is both a staple crop and one of the oldest genetic model organisms, and dozens of single-cell RNA-seq studies have profiled its roots, leaves, reproductive tissues and developing seeds. Those datasets remain partitioned by study, platform, genotype and processing pipeline, so the community has had no unified coordinate system in which cells from different experiments can be compared — and no maize-specific foundation model, because algorithms trained on isolated datasets learn study-specific structure rather than reusable expression programs.

scMaize, from Jun Yan's group at China Agricultural University, addresses both gaps at once. It pairs scMaizeAtlas, an integrated reference of 385,675 quality-controlled cells drawn from 20 public projects and 66 samples across seven tissue systems, with two Transformer encoders pretrained on that atlas. scMaizeExp is an expression-only baseline; scMaizeGO adds Gene Ontology functional embeddings as an explicit inductive bias, and is the variant whose parameter count is recorded here.

The design argument is that species-focused pretraining at modest scale can beat broad cross-taxa representations within a target organism. Where scGPT, Geneformer and scFoundation were trained on tens of millions of human cells, and cross-species efforts such as TranscriptFormer and GeneCompass span many organisms, scMaize commits its capacity entirely to maize.

#Key Features

• Integrated maize reference: 681,032 collected cells were filtered to 385,675 and batch-corrected with scVI, then annotated through a hierarchy of 34 coarse and 46 fine-grained cell types built from 3,074 non-redundant curated markers.
• Gene Ontology as an inductive bias: scMaizeGO injects IDF-weighted SVD embeddings of GO annotations through a dedicated input branch, improving rank-order expression prediction and aligning attention with gene co-expression structure.
• Cross-species representation transfer: embeddings generalize from maize to rice root tip and Arabidopsis leaf datasets via one-to-one orthologs, spanning the monocot–dicot divide without retraining the backbone.
• Lightweight few-shot adaptation: a DualStream head of roughly 300K trainable parameters fuses raw expression with the frozen cell embedding, reaching high accuracy from a small labeled fraction.
• Perturbation-aware gene discovery: passing mean expression through the pretrained projection layer scores condition-responsive genes in the model's learned space, without the model ever seeing treatment labels.
• Zero-code web platform: scmaize.com serves interactive atlas browsing, downloadable weights, and four hosted pipelines for annotation, embedding extraction, imputation and gene similarity.

#Technical Details

Both models are 6-layer Transformer encoders with hidden dimension 384, 4 attention heads, feed-forward dimension 1,536 and dropout 0.1; scMaizeExp has 16.7M parameters and scMaizeGO 16.8M. Each cell is represented as 2,048 gene tokens sampled from a 15,000-gene vocabulary — 13,000 highly variable genes plus 2,000 chosen to maximize GO coverage — with separate branches encoding gene identity, continuous expression, batch label (removable at inference) and, for scMaizeGO, the GO embedding. A learnable CLS token yields the 384-dimensional cell representation. Pretraining used masked gene modeling with 15% masking and an MSE loss weighting non-zero values fivefold, AdamW at 2e-4, effective batch size 256, and 80 epochs on a single A100 40GB with peak memory under 15 GB.

Test-set Pearson correlation on masked expression was 0.7825 for scMaizeExp and 0.7841 for scMaizeGO — essentially tied — but the GO prior separates the two downstream. On held-out high-confidence cells, scMaizeGO reached 86.0% cell-type and 97.1% tissue classification accuracy, against 72.9%/79.7% for Harmony-corrected HVG embeddings and 46.8%/51.7% for a maize-fine-tuned scPlantLLM roughly seven times larger. Zero-shot k-NN macro F1 on external data was moderate (0.292 maize root, 0.398 rice, 0.553 Arabidopsis), rising to 0.746, 0.894 and 0.926 after few-shot tuning on 15% of labels.

#Applications

The atlas gives maize researchers a common reference for marker discovery, cell-type annotation and cross-study comparison, while the pretrained encoders supply cell embeddings, expression imputation and gene-similarity analysis for new experiments. Because transfer to rice and Arabidopsis works from ortholog mapping alone, groups working on other grasses can label small datasets without pretraining anything. The perturbation analysis — which recovered chitin response and salicylic acid signaling under Fusarium verticillioides, and glutathione and water-deprivation pathways under heat — points at stress-gene prioritization for breeding programs. The hosted apps make all of this reachable without writing code.

#Impact

scMaize is a template for species-specific plant single-cell modeling: a curated atlas, a structured functional prior, and a compact model that a single GPU can train. Its honest boundaries are stated in the paper — zero-shot classification is only moderate, transcription factors are underrepresented among attention hubs because attention tracks co-expression rather than regulation, and batch mixing is traded away for biological structure. The work is a preprint under a non-commercial license, and the code repository is new, though it is MIT-licensed and ships inference scripts that need only a checkpoint and an .h5ad file.