first pass at documentation

app.py

@@ -98,7 +98,6 @@ if __name__ == '__main__':
         st.session_state.off_target = None
 
     # title and documentation
-    st.title('TIGER Cas13 Efficacy Prediction')
     st.markdown(Path('tiger.md').read_text())
     st.divider()
 

tiger.md

@@ -1,3 +1,45 @@
+## TIGER Cas13 Efficacy Prediction
 
+Welcome to TIGER!
+This Hugging Face space is an online tool that accompanies our Nature Biotechnology article.
+TIGER's ability to make accurate on- and off-target predictions enables biologists to both design highly effective gRNAs and precisely modulate transcript expressing by engineering gRNA mismatches.
+If you utilize our model, please consider citing us:
 
-Wessels, H.-H., Stirn, A., Méndez-Mancilla, A., Kim, E. J., Hart, S. K., Knowles, D. A., & Sanjana, N. E. (2023). Prediction of on-target and off-target activity of CRISPR–Cas13d guide RNAs using deep learning. Nature Biotechnology. https://doi.org/10.1038/s41587-023-01830-8
+> Wessels, H.-H., Stirn, A., Méndez-Mancilla, A., Kim, E. J., Hart, S. K., Knowles, D. A., & Sanjana, N. E. (2023). Prediction of on-target and off-target activity of CRISPR–Cas13d guide RNAs using deep learning. Nature Biotechnology. https://doi.org/10.1038/s41587-023-01830-8
+
+[//]: # (In our article, TIGER predicts log2 fold-change (LFC) from target sequence, guide sequence, and additional scalar features.)
+
+[//]: # (Prior to training, we normalize our survival screen's LFC values on a per-gene basis to discourage TIGER from learning which target transcripts come from more essential genes.)
+
+[//]: # (As such, TIGER outputs a normalized LFC estimate.)
+
+This tool differs from our manuscript in two ways.
+First, this version of TIGER predicts using just target and guide sequence, which will marginally reduce performance (fig 3c).
+Second, we map TIGER's outputs to the unit interval to make estimates more interpretable: a one corresponds to high gRNA activity and a zero denotes no activity.
+This transformation, maps estimates with no detectable Cas13 activity to (0,0.025] and the most active 2.5% of estimates to [0.975,1)
+This transformation is monotonically decreasing and therefore preserves Spearman, AUROC, and AUPRC performance.
+We label these transformed LFC estimates as `Guide Score` in our prediction tables.
+
+
+### Using TIGER
+
+We support two methods for transcript entry:
+- Manual entry of a single transcript
+- Uploading a FASTA file that can contain one or many transcripts provided each has a unique ID
+
+We currently offer three run modes:
+- We report all on-target gRNAs for each provided transcript. This mode does not support off-target identification due to current computational constraints.
+- We report the top ten most active, on-target gRNAs for each provided transcript. This mode allows for the optional identification of off-target effects.
+- We report the top ten most active on-target gRNAs for each provided transcript and their titration candidates (all possible single mismatches). This mode also does not support off-target identification due to current computational constraints.
+
+We use version 19 of gencode (protein-coding and lncRNA) to identify off-target candidates.
+
+### Feature Roadmap
+
+- Off-target scanning speed improvements
+- Off-target scanning for titration mode
+- Allow user to select more than the top ten guides per transcript
+- Incorporate non-scalar features (target accessibility, hybridization energies, etc...)
+
+To report bugs or to request additional features, please click the "Community" button in the top right corner of this screen and start a new discussion.
+Alternatively, please email [Andrew Stirn](mailto:andrew.stirn@cs.columbia.edu).