plantcad/genecad_revision_models

GeneCAD revision models

Checkpoints of the GeneCAD gene-structure models trained for the revision of the GeneCAD manuscript (Bioinformatics, BIOINF-2026-2172). They are provided so that every result in the revised manuscript can be reproduced.

Training-data conditions

All models share the same eight training species, the same 8,192-bp windows, the same code and the same schedule. They differ only in which annotated genes are used as supervision (excluded genes are masked, not used as negatives):

Condition	Training genes
U	All genes the data pipeline can represent (no filtering)
B	U after basic structural QC
M	B after MMLR selection (positive set: PeptideAtlas-supported A. thaliana and O. sativa isoforms; no maize data)
R	B with genes removed at random to match the retention of M (one random draw per seed)

Each condition was trained with seeds 11, 22 and 33.

Common settings

• Code: plantcad/genecad commit 0701f7f0d3e077fe22f81c2ea7f416e2351519cb
• Backbone: emarro/pcad2-200M-cnet-baseline, revision 5574ff5d
• 1,753 optimizer steps, effective batch 384 (2 GPUs x 4 x 48 accumulation), learning rate 2e-4 with cosine decay, 8 head-encoder layers, token embedding 256, window 8,192 bp
• Hardware: 2 x NVIDIA GH200 (TACC Vista)
• last.ckpt is the step-1,753 checkpoint. training_evidence.json records the run (seed, steps, batch accounting, gradient checks, timing).

Files

<condition>/seed<NN>/last.ckpt
<condition>/seed<NN>/training_evidence.json
SHA256SUMS

Use

genecad predict -i genome.fa -o out/ -s MySpecies -m plant \
  --model-checkpoint path/to/M/seed11/last.ckpt