doench_et_al_paper_summary.md

change seq of sgRNA to change Cas-9; optimize on-target efficacy and minimize off-target potential

sgRNA sequence and experimental condition - important for high efficiency

1841 sgRNAs used for screening - resulted in Rule Set 1 score and used for the following libraries:

1. Avana - human
2. Asiago- mouse

6 sgRNAs per gene, are selected based on: (Supp Table 1- 3)

1. Rule Set 1 score
2. Specificity within protein-coding regions
3. Target site location within the gene

Libraries tested using resistance to vemurafenib (Zelboraf) in A375 melanoma cells which carry BRAF V600E mutation and are sensitive to MAPK pathway (Supp Fig 1 - 2)

More screening data in Supp Fig 3

sgRNAs were ranked by their log2 fold-change relative to their abundance in the plasmid DNA pool and averaged ranks from the two replicates - Supp Table 4

RIGER Algorithm screening for GeCKOv1, GeCKOv2 and Avana - Supp Table 5 + Supp Fig 4

RIGER only uses data from the top two perturbations so STARS was developed - generates false-discovery rates (FDR) - similar to MAGeCK algorithm

STARS observation - CUL3, MED12, NF1, NF2, TADA1, TADA2B - FDR < 1% - Supp Table 6

Annotation of PanCancer genes - loss may restore MAPK pathway or find other path for survival - Supp Table 7

Selumetinib screen - Supp Tables 8 and 9

Screening results - Supp Figs 5 and 6

Negative screens - Supp Fig 7 and Supp Tables 10 and 11

ROC-AUC of each sgRNA for the three libraries - Supp Fig 8 and Supp Table 12

Gene Set Enrichment Analysis - Supp Fig 8 and Supp Table 12

Avana + GeCKO -> STARS - Supp Table 13

Negative-selection screen performance of Avana in HT29 (colon cancer cell line) - Supp Table 10 and 11 and Supp Fig 10

More Avana library analysis - Supp Table 12

Screening of Avana to purine analog 6-thiguanine in A375 (HPRT1 was targeted most), HT29 and 293T cell lines (HPRT1 was enriched but NUDT5 targeting produced similar resistance) - Fig 2a, Supp Table 14 and Supp Fig 11

Asisago - BV2 cells challenged with interferon gamma and STARS - Table 1 and Supp Table 15

Rule Set 2 - Fig 3a and Supp Table 16 - 2549 sgRNAs targeting 8 genes make RES data set.

4000 sgRNAs targeting 17 genes - efficacy of each sgRNA versus its position in protein coding region - Supp Fig 12 and Fig 3c

Machine learning models already used:

1. Linear Regression
2. L1-regularized linear regression
3. L2-regularized linear regression
4. hybrid SVM plus logistic regression
5. Random Forest
6. Gradient-boosted regression tree
7. L1 logistic regression (classifier)
8. SVM classification

One-hot encoding done

20% versus 80% classification using SVM with logistic regression gave best results - Fig 4a and Supp Fig 13.

Linear vs Logistic regression model performance on FC data set - Supp Fig 14 - linear performed better

Factors considered: (Fig 4b) - L1 regression model

1. Previously single and dinucleotide position-specific nucleotides and GC count of sgRNA were used
2. Position-independent nucleotide counts
3. Location of sgRNA target site within gene
4. Biochemical and structural properties - specifically thermodynamic properties
5. Microhomology features - did not improve performance

For RES and combined data set - gradient boosted regression trees performed better

For FC data set, performance was comparable

Gradient boosted regression tree model proposed best - single-nucleotide and dinucleotide at each position of sgRNA, contributed most (58%) - Supp Fig 15 and Supp Table 17

Position-independent counts of single and dinucleotides - 16%; location of sgRNA within protein - 13% and melting temperatures of different regions - 11%

Model trained with combined data set performed best - Supp Fig 16

Gradient-boosted regression tress model with the augmented feature set trained on combined data set - Rule Set 2

CFD scoring - Supp Fig 19, Supp Table 19 and 20, Supp Fig 21 and 22

Optimized sgRNA libraries - Brunello and Brie - Supp Tab 21 and 22

CCtop algorithm implemented for scoring of Rule Set 2

Bowtie2 not suitable for large sequences

GUIDE-seq used for finding off-target sites