Update content/02.main-text.md

Co-authored-by: Casey Greene <[email protected]>

content/02.main-text.md

@@ -253,7 +253,7 @@ For the 5-layer neural networks, the generalization results were similar to the
 Using public cancer genomics and transcriptomics data from TCGA and CCLE, we studied generalization of mutation status classifiers for a wide variety of cancer driver genes.
 We designed experiments to evaluate generalization across biological contexts by holding out cancer types in TCGA, and to evaluate generalization across datasets by training models on TCGA and evaluating them on CCLE, and vice-versa.
 We found that, in general, smaller or more parsimonious models do not tend to generalize more effectively across cancer types or across datasets, and in the absence of prior knowledge about a prediction problem, simply choosing the model that performs the best on a holdout dataset is at least as effective for selecting models that generalize.
-Given that similar "smallest good" heuristics are used broadly across genomics studies (see, e.g. [@doi:10.1007/s00405-021-06717-5; @doi:10.1089/dna.2020.6193; @doi:10.1186/s12859-021-04503-y]), we think this is a useful finding that will have implications on current practices.
+Given that similar "smallest good" heuristics are used broadly across genomics studies (see, e.g. [@doi:10.1007/s00405-021-06717-5; @doi:10.1089/dna.2020.6193; @doi:10.1186/s12859-021-04503-y]), we expect these results to have implications on current practices.
 
 Our results were similar in both linear models (LASSO logistic regression) and non-linear deep neural networks when using hidden layer size as the regularization parameter of interest.
 In our non-linear model experiments, we did not observe better generalization across datasets for fully connected neural networks with fewer hidden layer nodes, and our preliminary results indicated a similar trend for dropout and weight decay.