DELIGHT-FM Epi-colon Age
Saved tissue-specific age-regression model for log2(PCW) prediction and verified age perturbations. Includes the fitted regression head and fine-tuned backbone.
Repository: chloehlee/delight-fm-epi-colon-age · License: Apache 2.0 · Selected foundation epoch: 7.0.
Start here
This repository includes age-prediction and gene/SCENIC+ age-perturbation notebooks. New-task classification fine-tuning is provided by the foundation companion. Human tissue-matched single-cell RNA expression is required.
Load this model
Use Python 3.10, the tested PyTorch/CUDA 11.8 stack and compiled Mamba extensions
described in the notebooks' Environment section and requirements.txt.
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained(
"chloehlee/delight-fm-epi-colon-age", trust_remote_code=True)
For preprocessing and the complete inference workflow, follow the linked notebook.
Review the custom Python before enabling trust_remote_code.
Inputs and saved outputs
Inputs comprise gene token IDs, float32 per-gene z-scored log-normalized expression
(clipped to ±5), and aligned HVG masks. Use the matching reference gene statistics
and tokenization files in resources/; resources/preprocessor_config.json records the
historical initial-ranking backend.
The saved single-output head predicts log2(PCW). Convert with 2**prediction without subtracting one. For edits, positive prediction deltas mean a higher predicted log2(PCW). Do not substitute a foundation backbone or refit the age head between perturbation jobs.
DELIGHT-FM model family
| Foundation | Task-specific companion |
|---|---|
| DELIGHT-FM | Saved broad-cell head included |
| DELIGHT-FM Epi-ileum | Age companion |
| DELIGHT-FM Epi-colon | Age companion |
| DELIGHT-FM Stromal | Age companion |
| DELIGHT-FM Immune | Age companion |
Selection and limits
Training subset: 85035 cells. Validation/test subset: 25155 cells. Reference gene features: 18253; context length: 2048. Pretraining combines masked gene modeling with auxiliary cell-cluster/broad-cell and learned age-scale supervision and adversarial batch/domain heads; all saved heads are preserved in the foundation export. Age companions use a separate fitted regression head.
Selection maximized the mean of cluster macro F1, location macro F1 and age Pearson,
excluding explicit single-class tasks. See metadata/selection.json for component values.
These are validation-selected results; the original test subset participated in model
selection. New downstream tasks should use donor-disjoint train/validation/test partitions
and evaluate generalization independently.
Perturbations measure model response, not experimentally established causal effects. Gene deletion removes the selected gene token. Regulon deletion jointly removes mapped target tokens and, when selected, the TF token. SCENIC+ sign suffixes identify reconstructed regulon programs; they are not used to invert deletion into overexpression.
Reproducibility and validation
Weights are unchanged from the final source checkpoint.
Perturbation tutorials reuse one saved model and one immutable baseline, verify
cell order, inputs, code/resources and weights, then compare predictions and embeddings.
Inference uses float32 with autocast/TF32 disabled. Deletion removes aligned tokens, expression values and HVG masks.
Model source metadata and selection records are in metadata/.
Credits and license
Published by chloehlee as part of DELIGHT-FM.
Release code and model weights: Apache 2.0. Reference assets and dependencies retain their original terms; see third-party attribution.
Full workflow and supporting data
Training, downstream evaluation and perturbation workflows: ChloeHJ/delight_fm_v1. Supporting training/expression inputs are staged for chloehlee/delight_fm_v1-data; this model package contains its own inference code, weights, preprocessing resources and tutorials.
Task-specific notebooks
Bundled resources/regulons.tsv contains epithelial SCENIC+ reconstructed regulons used for the colon epithelium age analysis. Columns include eRegulon_name, tf, target and compartments. The notebook uses this repository's saved log2(PCW) regression model to delete individual genes or jointly delete mapped targets of one reconstructed eRegulon, optionally including its TF. Distinct sign-specific eRegulon names are retained. Target mapping and token coverage are recorded for every program.
All edits reuse this saved age model and one verified baseline. Positive changes mean older predicted age and negative changes mean younger predicted age, in log2(PCW).