mycoCirc

Circular RNAs (circRNAs) are covalently closed transcripts produced by backsplicing, in which a downstream splice donor joins an upstream acceptor. They are largely invisible to standard experimental practice: circRNAs carry no poly(A) tail, so poly(A)-selected mRNA-seq discards them, and the total RNA-seq protocols that retain them are too costly for large-scale screening. Discovery therefore depends on computational prediction. The available predictors were trained exclusively on human or plant sequence, and they transfer poorly to fungi, whose genomes span an enormous architectural range — from intron-poor Candida species, where most genes are single-exon, to intron-rich filamentous fungi.

mycoCirc, developed by Xueyan Hu and colleagues in Ence Yang's group at Peking University Health Science Center and posted as a bioRxiv preprint in July 2026, is an end-to-end pan-fungal multi-modal pretrained model built for this gap. It predicts which genes in a fungal genome produce circRNAs, and which backsplice junction within a gene is most likely, taking only a genome FASTA and a GTF or GFF3 annotation as input. Before it, no computational framework existed for de novo fungal circRNA prediction, leaving most non-model fungi inaccessible.

The design is deliberately compact — 775,858 parameters across five modality encoders — and leans on pretraining plus per-lineage fine-tuning rather than scale.

#Key Features

• Annotation-only inference: Mode A runs on genome sequence and gene annotation with no expression data, the regime that matters for the thousands of fungal species with an assembly but no total RNA-seq. Mode B optionally adds host-gene expression.
• Bidirectional donor–acceptor cross-attention: The JunctionEncoder embeds every exon 5' donor and 3' acceptor site in a gene and computes a pairwise compatibility matrix between the two arrays, turning gene-level classification into a ranked list of candidate backsplice junctions.
• Explicit species conditioning: A SpeciesEmbedding combines an 8-dimensional phylogenetic PCA with a learnable per-species vector, related to the conditioning used in the Species-Aware DNA Language Model.
• Lineage-specific checkpoints: Three fine-tuned models are released — Ascomycete yeast (Candida), Basidiomycete yeast (Cryptococcus), and filamentous fungi — so users select by phylogenetic proximity.
• mycoCircAtlas: A companion database of 319,860 high-confidence gene–circRNA predictions across 768 fungal species from Ensembl Fungi Release 113, with primer-design support.

#Technical Details

The architecture fuses five encoders. A GenomicContextEncoder applies three dilated 1D convolutions (kernel 7, dilations 1/2/4) and a BiGRU over a ±5 kb window discretized into 200 bins of 50 bp, yielding a 128-dimensional context vector. A GTFEncoder maps 17 gene-structure features — exon count, exon and intron length statistics, CDS length, GC content, biotype — through an MLP to 128 dimensions. The JunctionEncoder, the largest module at 245,984 parameters, combines a k-mer BiGRU path with cross-attention over donor and acceptor arrays (up to 50 exons per gene, ±150 bp flanks) and a convolutional path over one-hot flanks. A SpeciesEmbedding and an ExpressionEncoder complete the set; the FusionModule compresses the 640-dimensional concatenation to 128 dimensions before a gene head and a junction head.

Pretraining used 22 fungal strains and 16,483 positive gene–circRNA associations in two stages: 50 epochs of gene-level binary classification with the JunctionEncoder frozen, then 100 epochs optimizing gene and junction objectives jointly. Each lineage variant was then fine-tuned with 5-fold leave-one-strain-out cross-validation. On held-out test species under Mode A, mycoCirc reached AUROC 0.6985 on C. auris, 0.6902 on C. neoformans, and 0.6976 on F. venenatum, against 0.51–0.57 for JEDI and 0.49–0.53 for CircPCBL. Cross-species evaluation on four independent fungal datasets gave AUROC 0.63–0.72. Ablations show the GTF modality dominates (removal costs 0.06–0.14 AUROC), while models trained from scratch collapse to roughly 0.50, confirming that pretraining carries the transferable signal.

#Applications

The model targets researchers working on fungi that have a reference assembly but no circRNA sequencing data, which is nearly all of them. Typical use is to rank candidate circRNA host genes genome-wide, select the top-scoring junctions, and design divergent primers for RT-PCR validation, collapsing an otherwise open-ended experimental search. That matters for human pathogens such as Candida auris and Cryptococcus neoformans, where circRNAs are an unexplored layer of regulation, and for industrial filamentous fungi used in enzyme and metabolite production.

#Impact

mycoCirc establishes the first de novo prediction framework for a transcript class fungal genomics has had essentially no computational access to, and does so with a small, MIT-licensed model whose weights ship in the repository. AUROC near 0.70 makes it a screening prior rather than a definitive call, and the authors document why: within-group cross-validation exceeds 0.99 while held-out species sit near 0.69, indicating the model captures lineage-specific rather than universal fungal circRNA grammar. The expression modality helps within a group but does not transfer across species, and junction-level ranking has not been extensively validated against experimental backsplice data. The work is a preprint and has not yet been peer reviewed. Even so, it converts fungal circRNA discovery from a prohibitively expensive sequencing problem into a tractable prioritization one.