Porecast prediction results
RNA → Electrophysiology
Porecast predicts 7 electrophysiological features from a cell's 354-gene ion-channel expression profile. Validated on 161,145 Patch-seq sweeps from 1,039 donors.
What the traces mean.
Each feature is predicted from ion-channel RNA expression alone — no patch-clamp measurement needed.
Generated vs recorded, by cell type.
Each card shows a real Patch-seq recorded trace (black) alongside traces generated by Porecast from that cell's ion-channel RNA (blue, with uncertainty bands). Mouse visual cortex, 12 representative types.
Seven features, one expression vector.
Each bar shows Pearson r between predicted and measured values, on held-out donors.
n = 161,145 sweeps · 1,039 donors · 5-fold nested cross-validation with donor blocking.
RNA beats every shortcut baseline.
Probabilistic accuracy (CRPS)
| Model | CRPS | Δ |
|---|---|---|
| LightGBM (354 genes) best | 7.73 | — |
| probabilistic_feature_template | 7.73 | +0.0 |
| Ridge (354 genes) | 10.56 | +2.8 |
| Current step only | 11.79 | +4.1 |
| ttype_mean | 13.14 | +5.4 |
| deterministic_template | 13.36 | +5.6 |
| Brain-area mean | 13.65 | +5.9 |
| rna_nn | 13.97 | +6.2 |
For binary spike-counting classification (firing vs non-firing), RNA expression alone achieves AUC 0.94.
Lower CRPS = better. Donor-blocked bootstrap (95% CI excludes zero).
Predictions validated on mouse visual-cortex Patch-seq data only (Mouse visual cortex Patch-seq (DANDI 000020)). Body-wide extrapolation to human tissues is not claimed. The expression atlas itself covers 354 genes across 738 human cell types.
How the predictions work.
Ion-channel expression (354 genes) feeds a ridge regression on 354-gene ion-channel expression. 5-fold nested cross-validation, donor-blocked bootstrap. Donor blocking prevents leakage between train and test folds.
The generative waveform model (conditional flow matching) produces the blue traces shown above — plausible voltage traces from RNA. It is a proof-of-concept companion to the validated scalar predictions.