In the introduction the authors lay out how tumor cells evade immune recognition by expressing the ligand PD-L1 to bind PD-1 protein on T-cells, leading to reduced immune responses against the tumor. Blocking the PD-1/PD-L1 interaction would help the immune system attach the tumor cells. Most of the immune checkpoint inhibitors are monoclonal antibodies and the authors in this study are trying to find small molecules that mimic the mAB activity. Early work identified three key regions on PD-L1 where small molecules can bind, thereby inhibiting its interaction with PD-1. The authors review Bristol Myers Squibb's work in this area (BMS-200 and BMS-202) and other small molecules that are undergoing clinical trials at the moment. The authors provide an extensive review of other virtual screening follow up studies done on this target and highlight that the small molecules block the protein-ligand activity by inducing dimerization of PD-L1. The authors introduce the benefits of drug repurposing and present their approach of picking molecules from CHEMBL and ZINC databases and performing docking and testing compounds in an immunoassay.

Few clarifications needed on the following:
- "A total of 29,197 molecules from the ChEMBL and PubChem databases, along with recent literature from the Web of Science, were utilized to build these models....
...
This search was conducted in October 2022, utilizing the following search keywords: PD-L1; PD-1; PD-1/PD-L1." - It was not clear how these 29K molecules were picked from ChemBL, can you please provide some more details on any filters applied. The molecules from ZINC are annotated with their IDs but the associated CHEMBL IDs do not exist for the CHEMBL data in the SI.
- "For duplicates with different activity values, the respective bioassays were consulted to align the “active label” provided by the assays with the defined goal." - I do not understand this statement, can you please elaborate on the process here.
- In Table S4, Compound 195
"Cmpd #, SMILES, Experimental activity, Random Forest (RF), Support vector machine (SVM), Deep Learning Multilayer Perceptron Networks (dMLP)"
"195, SMILES, 0 (Inactive), 0.110, 0.001, 0.000"
May be I am misinterpreting, can you explain why the classifier is assigning numbers (or probablities?) instead of categorical labels, such as "Active" or "Inactive" here, what do these predicted activity numbers mean in this context?
- Table 8 is difficult to interpret with the keys, the first column is slightly human readable. Also, what is the relative difference between the ranked key attributes, how much is rank_1 attribute important than rank_20th?
- " PDI-1 inhibitor and sonidegib showed a 50% of inhibition at 5.523 μM and >10^8 μM, respectively." - I think there is an extremely grave typo here regarding Sonidegib's IC50. Same in Figure 7 legend.

• Minor typos here:
• "a corroboration technique, manly based on".
• Two "Table 8"s in the proof.
• I think the journal automatically appended all attachments and the proof is 4000+ pages because of the SI excel data, may be the journal team can provide some guidance on that.