Telen-NOCAP

The perturbation half of the atlas behind this model measures knockdowns of 68 genes. Asking which of the 19,424 protein-coding genes carries autism risk therefore means predicting responses to perturbations nobody has run — and, less obviously, choosing the cell type in which to ask, because the same knockdown means different things in a radial glial cell and in a cortical neuron. Telen-NOCAP is a transformer over single-cell transcriptomes of human brain organoids, fine-tuned on telencephalic neurons, that predicts a genome-wide transcriptional response to the in silico knockdown of any protein-coding gene.

The model was developed at Korea University with collaborators at Seoul National University, the University of Toronto and the Hospital for Sick Children, and released as a bioRxiv preprint in August 2026. It sits on NOCAP (Neural Organoid Cell Atlas with Perturbation), a 3,643,130-cell integration of 501 organoid samples from 40 published scRNA-seq studies, split into a neurotypical Core Atlas of 1,418,675 cells used for pretraining and an Extended Atlas of 2,224,455 cells — including 143 perturbed or chemically treated samples — used for fine-tuning.

Telen-NOCAP adopts the architecture of scGPT but not its weights. The transformer is pretrained from scratch on the Core Atlas, so its gene vocabulary and representations come entirely from organoid tissue. That distinction is the paper's central experiment rather than an implementation detail: matched variants built on the public scGPT brain-tissue and whole-human checkpoints form the comparison arm.

#Key Features

• Organoid-native pretraining: The backbone is pretrained on 27,245 genes across 1,418,675 neurotypical organoid cells, giving it representations of a developing in vitro neural system rather than adult post-mortem tissue.
• Cellular resolution as a design variable: Three fine-tuned models — Pan-NOCAP (all cell types), Neuro-NOCAP (all neurons) and Telen-NOCAP (telencephalic neurons) — isolate the effect of context. Autism risk genes organize most coherently in Telen-NOCAP's perturbation space.
• Genome-wide in silico knockdown: A single fixed checkpoint, fine-tuned on 60 perturbation conditions, predicts responses across all 19,424 protein-coding targets, reported over the same 19,424 genes.
• Selection on a four-domain benchmark: Models were ranked on prediction accuracy, embedding quality, differential-expression recovery and perturbation specificity, plus Moran's I coherence of risk genes — a design that penalizes predictions that correlate well but collapse into an uninformative embedding.
• Released checkpoints and code: Pretrained and fine-tuned weights ship through Git LFS alongside the pretraining, fine-tuning, prediction and benchmarking scripts under an Apache-2.0 license.

#Technical Details

The backbone is a four-layer transformer with four attention heads, an embedding size of 64, a hidden dimension of 64 and a maximum sequence length of 1,536, with expression discretized into 51 continuous bins. Pretraining ran 10 epochs on A100 hardware using both masked-value-completion and generative objectives at masking ratios of 0.25, 0.50 and 0.75. Fine-tuning used 512,740 telencephalic neurons and 60 perturbations, 15 epochs at learning rate 1e-4, with conditions split 75/25 into training and held-out test sets; ten seeds were trained and the checkpoint with the highest correlation on differentially expressed genes was kept.

Across the 17-model benchmark, Telen-NOCAP was the most balanced rather than the most accurate: Pearson ΔDE 0.901, RMSE 0.081, centroid accuracy 0.772, kNN distance 0.616 and Moran's I 0.0305 (FDR < 0.05). Telen-scFoundation, built on scFoundation, scored the highest correlation of any telencephalic model at 0.985 but collapsed to a kNN distance of 0.0086 and centroid accuracy of 0.341, and the scGPT whole-human variant reached 0.963 correlation with weaker perturbation specificity. Baselines built on GEARS, Geneformer, GeneCompass and CellFM span correlations from 0.727 upward. Two robustness variants accompany the release: Telen-NOCAP-Lite, refit after excluding two dominant studies, and Telen-Mouse, trained on a mouse forebrain atlas.

#Applications

The model is built for gene prioritization in neurodevelopmental genetics. Clustering the predicted responses of all 19,424 genes yielded two programs — a mid-fetal synaptic neuronal cluster and an early radial glial ubiquitin-signaling cluster — tested against damaging-variant enrichment and clinical phenotypes in 89,916 family-based samples. Logistic-regression prioritization over those profiles nominated 343 candidate genes, 167 inside the two programs, with recurrent support for NBEA and KLHDC10. The same profiles give organoid labs a way to rank candidate knockdowns before committing bench time.

#Impact

Telen-NOCAP argues that a perturbation model's training substrate and cellular context are load-bearing, not incidental: the paper's own benchmark shows a model pretrained on far more cells of adult brain tissue predicting expression more accurately while representing perturbations less usefully. That is a caution against ranking perturbation models by correlation, which is how most of them are compared. The work is a preprint awaiting peer review under CC-BY-NC-ND 4.0, and its authors describe the predictions as functional priors requiring prospective validation. The integrated NOCAP atlas is deposited on Zenodo under restricted access, so reproducing pretraining requires an access request even though the checkpoints and code are open.