The paper, "Cell morphology and gene expression: tracking changes and complementarity across time and cell lines," by Islam et al., introduces a label-free machine learning framework for cell classification for real-time sorting with bright-field microscopy images. The idea is innovative and supported with the implementation by a FGPA. However, there are some concerns regarding whether it is really practical for biological applications. It is recommended that the paper should be significantly revised before publication. Detailed comments are listed below:

1. When the users want to use the sorter to sort new cell types or sub-types for a new application, is it needed to re-train the model and do transfer learning? If yes, it will be very complicated and impractical.

2. Related to the previous question, for a new application, how to do gating with a label-free setup?

3. The accuracy seems high, but probably still much lower than fluorescence based methods. Can the authors provide some comparison? For certainly biology, the contamination of a few percent is not acceptable for proving or disproving a hypothesis.

4. The authors demonstrate the separation between B and T cells. Given the difference between B and T cells is huge, the accuracy is not perfect. The situation might be worse when sorting sub-populations with a cell type. It is recommended that the authors can demonstrate the feasibility.

5. The team compares the performance with some complicated ResNet50/ResNet18 models. However, according to prior literature (e.g., PMID: 36251981), it is possible that a very simple model can intrinsically perform the task with high accuracy. It is not convincing that the transfer learning is really useful and necessary.

6. Typo of "Figure ??(a))" in section 2.1. Please fix it.

The reviewers addressed the comments from the reviewers.