CRE.AI.TIVE

Raising the expression of a native plant gene is harder than knocking one out. Promoter edits usually reduce activity, because the cis-regulatory elements that drive transcription are unknown, redundant, and scattered across kilobases of noncoding sequence. Cataloguing those elements species by species, then designing a promoter around them, does not scale to the number of crops and traits breeders care about.

CRE.AI.TIVE is the pretrained sequence-to-coverage model that Phytoform Labs built to sidestep that catalogue. It takes 65 kbp of plant DNA and predicts transcriptomic and epigenomic coverage tracks at 32 bp resolution, learned jointly across 12 plant species so that regulatory signal from well-annotated model plants supports the poorly annotated crops. The model gives the platform of the same name its scoring oracle: an in silico mutagenesis algorithm evolves promoter sequences against the model's predictions, and a protoplast-based massively parallel reporter assay (MPRA) measures what the winners actually do. The work comes from Phytoform Labs in Harpenden, UK, and was posted to bioRxiv in December 2024.

Architecturally it follows the human regulatory-genomics lineage — Enformer and especially Borzoi, whose RNA-seq coverage objective it adopts — and transplants it into plants, where prior work had mostly used short promoter windows from one to four genomes.

#Key Features

• Species-specific output heads: A single convolutional-transformer trunk feeds a separate head per species and per assay category, so one pretrained backbone predicts coverage for 12 plants without splitting into 12 models.
• Long input window: Inputs are 65,536 bp centred on a transcription start site, wide enough to reach distal regulatory sequence that short promoter windows miss.
• Multi-assay targets: Training targets combine RNA-seq coverage with ChIP-seq, ATAC-seq, DNase-seq, FAIRE-seq and DAP-seq tracks where available, encoding chromatin context alongside transcription.
• Cultivar fine-tuning: A new output head trained on in-house RNA-seq adapts the frozen backbone to a cultivar the model never saw, reaching accuracy comparable to the pretrained reference cultivar.
• Directed in silico mutagenesis: A generational search accumulates substitutions, insertions and deletions in a promoter and scores each candidate with the model, optimizing expression in one tissue while holding others constant.

#Technical Details

Sequences are one-hot encoded and passed through five convolutional layers, each halving resolution, to reach 32 bp bins; a stack of multi-head attention units then models long-range interactions, the outer 25 % of the binned sequence is cropped, and output heads emit binned coverage per target condition. The model has 140 million parameters, was trained with a mean squared error objective on log2-transformed counts, and pretraining took two weeks on eight H100 GPUs. Genomes and annotations came from Ensembl Plants, RNA-seq from NCBI SRA guided by PlantExp metadata, and epigenomic data from ChIP-Hub; splits were drawn along gene trees so homologues stay together.

Across the 12 species, gene-level Pearson correlation between predicted and measured log2 counts averages 0.82, with 0.84 for Arabidopsis thaliana, 0.81 for Oryza sativa and 0.84 for Zea mays; the transpose Pearson, measured across conditions within a gene, averages 0.48. Fine-tuned on tomato cv. Ailsa Craig, the model reaches 0.80 across genes and 0.46 across tissues. Applied to the tomato gene SlbHLH96, 100 generations of mutagenesis over a 211 bp proximal promoter produced millions of variants, of which 2,000 were synthesized and assayed by MPRA in cotyledon protoplasts. Replicates correlated at Pearson 0.965; predicted versus measured activity correlated at 0.396. Of 1,700 promoters designed for increased activity, 163 exceeded two-fold over controls and 32 exceeded four-fold, and the three strongest gave 1.5–2× higher GFP brightness by flow cytometry. Motif analysis of the best variant found 26 overlapping binding sites for 13 transcription factors where the wild-type promoter had four.

#Applications

The intended users are crop improvement teams that want to raise expression of a native gene without a viral promoter or a transgene. The workflow — predict, evolve, screen by MPRA, confirm by reporter — produces a graded series of promoter variants whose edits can then be installed by genome editing, and in principle extends to any gene in the 12 species covered by pretraining. The tomato case study targeted SlbHLH96, previously linked to drought tolerance.

#Impact

CRE.AI.TIVE shows that the coverage-prediction recipe developed on human and mouse data transfers to plants, and that multispecies pretraining lets crops with an order of magnitude less functional genomic data ride on signal learned from Arabidopsis. The candid part of the result is the gap between assays: a 0.396 correlation between bulk RNA-seq predictions and protoplast MPRA measurements is modest, and the model is more accurate on later mutagenesis generations than earlier ones. Variants have been validated in protoplasts, not in whole plants, so the link to drought tolerance remains to be shown. The model itself is proprietary — no weights, code, or license have been released, and a patent covering the technology has been filed.