2. Obtain known mechanism of action annotations from databases like DrugBank [17] and PubChem [18], or from the
literature.
3. Extract the data for compounds with a known mechanism of
action into a separate dataframe.
4. Using a package like reshape2, reshape the known mechanism of action dataframe into a matrix, X_train, whose rows
are compounds and whose columns are strains, with the values
of the matrix being LFC.
5. Make a named vector of mechanism of action labels, Y_train,
for each compound, and reorder the vector to match the order
of the rows in X_train.
6. Use the glmnet function cv.glmnet with X_train and
Y_train to fit a Lasso model of the multinomial family with
k-fold cross-validation of the hyperparameter, λ, optimizing to
minimize mean squared error (MSE).
7. Reshape the entire dataframe into a matrix, X, whose rows are
compounds and whose columns are strains, with the values of
the matrix being LFC.
8. Use the model trained in step 6 to predict the mechanism of
action labels from X, using the R function predict with the
type parameter set to “class”.
Where no reference compounds are known, compounds
may be prioritized based on hypersensitivity of hypomorphs
corresponding to a desired target. For example, Gene Set
Enrichment Analysis (GSEA) [19] can be used with KEGG
annotations [20] of hypomorphic gene products to determine
biological pathways that are enriched among hypersensitive
strains. Alternatively, compounds can be prioritized based on
the separation between most and next most sensitive strains.
4 Notes
1. The library of interest can be screened in dose-response or at a
single concentration per compound. In our experience, the
dose-response (1) yields a richer dataset, since the individual
and collective hypomorph behaviors can vary across the dose
range, (2) maximizes the chance of collecting data at a concentration with maximal differential activity across the hypomorph
pool. However, dose-response format increases the costs and
therefore limits the number of compounds that can be
screened. Nevertheless, we have also applied singleconcentration format to increase the breadth of chemical
space that can be screened, while still being able to sensitively
detect compound activity and assign mechanisms of action.
Proteolytic Hypomorphs for Compound Mechanism of Action
339
literature.
3. Extract the data for compounds with a known mechanism of
action into a separate dataframe.
4. Using a package like reshape2, reshape the known mechanism of action dataframe into a matrix, X_train, whose rows
are compounds and whose columns are strains, with the values
of the matrix being LFC.
5. Make a named vector of mechanism of action labels, Y_train,
for each compound, and reorder the vector to match the order
of the rows in X_train.
6. Use the glmnet function cv.glmnet with X_train and
Y_train to fit a Lasso model of the multinomial family with
k-fold cross-validation of the hyperparameter, λ, optimizing to
minimize mean squared error (MSE).
7. Reshape the entire dataframe into a matrix, X, whose rows are
compounds and whose columns are strains, with the values of
the matrix being LFC.
8. Use the model trained in step 6 to predict the mechanism of
action labels from X, using the R function predict with the
type parameter set to “class”.
Where no reference compounds are known, compounds
may be prioritized based on hypersensitivity of hypomorphs
corresponding to a desired target. For example, Gene Set
Enrichment Analysis (GSEA) [19] can be used with KEGG
annotations [20] of hypomorphic gene products to determine
biological pathways that are enriched among hypersensitive
strains. Alternatively, compounds can be prioritized based on
the separation between most and next most sensitive strains.
4 Notes
1. The library of interest can be screened in dose-response or at a
single concentration per compound. In our experience, the
dose-response (1) yields a richer dataset, since the individual
and collective hypomorph behaviors can vary across the dose
range, (2) maximizes the chance of collecting data at a concentration with maximal differential activity across the hypomorph
pool. However, dose-response format increases the costs and
therefore limits the number of compounds that can be
screened. Nevertheless, we have also applied singleconcentration format to increase the breadth of chemical
space that can be screened, while still being able to sensitively
detect compound activity and assign mechanisms of action.
Proteolytic Hypomorphs for Compound Mechanism of Action
339