Biotech & Health Research 2026
gCas9-OTNet
Off-target prediction for CRISPR gene therapy in Alzheimer's disease
The problem
CRISPR gene therapy for Alzheimer's disease needs to edit APP, BACE1, and PSEN1/2 with high specificity, but the standard off-target scoring tools (MIT score, CFD score) were built for wild-type SpCas9 and don't model the binding dynamics of Geminin-tagged Cas9 (gCas9), a variant with improved specificity for neuronal targets.
The approach
Built gCas9-OTNet, a hybrid CNN-Transformer that encodes each sgRNA/off-target DNA pair as a 20-channel-by-20-position matrix (one-hot sequences plus 12 mismatch-substitution channels) alongside 14 hand-engineered scalar features and optional chromatin-accessibility context, then trained it with focal loss against real GUIDE-seq and CIRCLE-seq off-target data.
gCas9-OTNet is a deep learning pipeline for predicting off-target cleavage of Geminin-tagged Cas9 (gCas9) in CRISPR gene therapy aimed at Alzheimer’s disease targets: APP (amyloid precursor protein), BACE1 (beta-secretase 1), and PSEN1/2 (presenilin).
why existing scores fall short
MIT score and CFD score are the standard off-target predictors, but both were fit to wild-type SpCas9 behavior. gCas9 is a Geminin-tagged variant engineered for improved specificity in neuronal targets - a different binding profile that off-the-shelf heuristics don’t capture, which matters a great deal when the therapy is editing genes directly implicated in amyloid pathology.
architecture
Each sgRNA/off-target DNA pair is encoded as a 20×20 matrix (4 channels for sgRNA one-hot, 4 for DNA one-hot, 12 for mismatch-substitution type) plus a 14-dimensional scalar feature vector (mismatch count and position, GC content, transition/transversion counts, PAM-proximal fraction) and an optional 6-dimensional chromatin-accessibility vector from ATAC-seq data.
Multi-Scale CNN (kernels 1,3,5) -> 64ch
-> Sinusoidal Positional Encoding
-> Transformer Encoder (2 layers, 4 heads, GELU)
-> Learned-Query Attention Pooling -> 64-dim embedding
+ Scalar Feature MLP (14 -> 32)
+ Chromatin Context MLP (6 -> 16)
-> Classifier Head (112 -> 64 -> 1 logit)Training uses focal loss (α=0.75, γ=2.0) to handle the severe class imbalance inherent to off-target data - the overwhelming majority of candidate sites never actually cleave - with AdamW and cosine-annealed warm restarts, early-stopped on validation AUPRC.
validation strategy
Because off-target claims feed into a therapeutic safety decision, the pipeline validates four different ways rather than trusting one train/test split: within-assay holdout on GUIDE-seq, cross-assay transfer from GUIDE-seq (in cellulo) to CIRCLE-seq (in vitro), 6-fold leave-one-sgRNA-out cross-validation, and bootstrapped 95% confidence intervals over 1000 iterations.
results
| Model | AUROC | AUPRC |
|---|---|---|
| gCas9-OTNet | 0.876 | 0.714 |
| CFD Score (Doench 2016) | 0.819 | 0.613 |
| MIT Score (Hsu 2013) | 0.803 | 0.582 |
| Mismatch Count (naive) | 0.724 | 0.489 |
gCas9-OTNet holds up under leave-one-sgRNA-out CV (0.873 ± 0.056 AUROC) but, as expected, loses substantial ground on cross-assay transfer (0.698 AUROC, GUIDE-seq → CIRCLE-seq) - a real limitation the pipeline reports rather than hides.
interpretability
Three complementary methods explain predictions rather than just scoring them: transformer attention-weight visualization across sequence positions, gradient saliency maps decomposed by channel group (sgRNA identity, DNA identity, mismatch type), and SHAP (TreeExplainer / LinearExplainer) on the XGBoost and logistic-regression baselines for direct feature attribution.
why it matters
A CRISPR safety score that can’t explain itself, or that silently fails to generalize across assay types, is a liability in a therapeutic pipeline. gCas9-OTNet is built to beat the standard heuristics on the numbers while surfacing exactly where its own confidence should be discounted.
What I took away
- A multi-scale CNN (kernels 1/3/5) followed by a small Transformer encoder with attention pooling beats both classical heuristics and a plain CNN baseline on AUROC and AUPRC, but the model's cross-assay transfer (train on GUIDE-seq, test on CIRCLE-seq) drops sharply, from 0.876 to 0.698 AUROC - in cellulo and in vitro assays measure meaningfully different things, and a model trained on one doesn't automatically generalize to the other.
- Leave-one-sgRNA-out cross-validation is a much harder and more honest test than a random split for this task, since it forces the model to generalize to guide sequences it has never seen rather than memorizing sgRNA-specific patterns.
- Interpretability (attention weights, gradient saliency by channel group, SHAP on the tree baselines) matters as much as raw accuracy for a therapeutic safety model - a black-box score isn't good enough when the downstream decision is which genome edits are safe to make in neurons.
0 comments
Loading…