DLTRNM

DLTRNM — distributed large-scale transcriptional regulatory neural network model — is a genome-scale model of transcriptional regulation in Saccharomyces cerevisiae. Rather than training one monolithic network over all genes, it builds a separate small neural network for every target gene, whose inputs are exactly the transcription factors already documented as regulating that gene. Curated interactions from SGD and YEASTRACT define the wiring; the data then fills in the quantitative relationships. The result is a distributed model whose architecture is an explicit encoding of prior biological knowledge rather than something the network must discover.

The work comes from Xingcun Fan, Hongzhong Lu, Xuefeng Yan and colleagues at East China University of Science and Technology and Shanghai Jiao Tong University, posted as a preprint in February 2025 and published in Synthetic and Systems Biotechnology later that year. It addresses a gap that scYeast approaches from the single-cell side: few pretrained regulatory models exist for budding yeast, and the large transformer-based alternatives demand computational resources most yeast labs do not have.

Training follows a two-stage transfer learning recipe. Pretraining on a pan-transcriptomic compendium teaches each subnetwork the static relationship between a transcription factor and its target; fine-tuning on time-course induction data teaches it how that relationship plays out dynamically.

#Key Features

• Knowledge-constrained architecture: Each target gene gets its own subnetwork whose input layer is restricted to its documented regulators, so the model cannot invent regulatory edges that the literature does not support.
• Fully distributed and independently runnable: Subnetworks are parallel and non-interacting, meaning a single gene can be trained, updated or queried without touching the rest of the genome-scale model.
• In silico perturbation: Tuning a transcription factor's expression propagates through the relevant subnetworks to predict target-gene response, and setting it to zero simulates a knockout across all genes at once.
• Key regulator identification: SHAP analysis on each subnetwork ranks transcription factors by contribution, which collapses a dense regulatory network into a legible core structure when only the top regulators are retained.
• Low computational demand: Small per-gene networks make the whole model tractable on modest hardware, in contrast to fine-tuning a large pretrained transcriptomic model.

#Technical Details

Each subnetwork is a small artificial neural network taking the expression of a gene's annotated transcription factors as input and that gene's expression as output. Pretraining used a compendium of 969 high-quality transcriptomes derived from RNA sequencing of 1,032 natural S. cerevisiae isolates, producing 6,085 pretrained subnetworks, of which 73.9% reached a summed variance below 100. Fine-tuning used time-course transcriptomics from over 200 transcription factor induction experiments, yielding 5,859 fine-tuned subnetworks covering 5,855 shared genes plus four newly introduced factors; when a new regulator appears for an existing target, hidden-layer weights are frozen and the remainder retrained. After fine-tuning, 98.02% of subnetworks reached a training loss below 0.2 and 86.87% did so on held-out data, with mean test loss 0.144. Under a three-level evaluation scheme, 82.84% of time-series samples were classified as excellent predictions and 13.75% as failures; 84.3% of subnetworks exceeded 80% accuracy. Directional validation against 1,963 SGD-annotated positive and negative regulatory relationships — data withheld from training — gave 53.2% correct directional calls, and knockout simulation across 225 transcription factors matched published expression changes in 51.95% of cases. SHAP coefficient-of-variation analysis found informative regulator rankings in 87.27% of subnetworks. The repository ships every checkpoint individually: 6,085 pretrained and 5,859 fine-tuned per-gene .pth files.

#Applications

DLTRNM is a hypothesis-generation tool for yeast systems biology and metabolic engineering. A strain designer asking which transcription factor to overexpress or delete to shift flux through ethanol or pyruvate metabolism can simulate candidate perturbations across the genome before building anything. The SHAP-derived key-regulator lists compress the full regulatory network into core subnetworks that are easier to reason about and to test, and the paper applies this to refine previously published regulatory subnetworks. Because subnetworks are independent, a lab interested in a handful of genes can download only the relevant checkpoints and run them locally.

#Impact

The model demonstrates that hard-coding curated regulatory topology into network architecture buys both interpretability and tractability: gradients are attributable to named transcription factors, and the compute cost stays within reach of an ordinary lab. Distributing every per-gene checkpoint makes the work genuinely reusable rather than merely reproducible. The limitations are stated plainly in the paper. Directional accuracy on held-out SGD relationships and on knockout data sits near 52–53%, so predictions are best treated as a ranked shortlist rather than a settled answer; subnetworks with too little fine-tuning data relative to their input dimension are unstable in both directions, overfitting as readily as underfitting. The model is also specific to S. cerevisiae by construction, since its topology comes from yeast-specific curation, and the repository ships no license file, which limits reuse in downstream work.