diff --git a/README.md b/README.md
index 0903961..717b4c8 100644
--- a/README.md
+++ b/README.md
@@ -1,5 +1,5 @@
 # Molecule-RNN
-Molecule-RNN is a recurrent neural network built with Pytorch to generate molecules for drug discovery. It is trained with the [Zinc](https://zinc.docking.org/) dataset.
+Molecule-RNN is a recurrent neural network built with Pytorch to generate molecules for drug discovery. It is trained with the [Zinc](https://zinc.docking.org/) dataset. The [SELFIES](https://github.com/aspuru-guzik-group/selfies) is used as the representation of molecules. The SMILES files are converted to SELFIES during training on-the-fly.
 
 ## Training
 1. Dowdload the SMILES files of molecules from [Zinc](https://zinc.docking.org/). Select "SMILES(*.smi)" and "Flat" opitions when downloading.
@@ -14,7 +14,15 @@ python train.py
 The training loss:
 
 ## Sampling
-We can generate molecules by sampling the model according to the output distribution.
+We can generate molecules by sampling the model according to the output distribution. 
 ```
 python sample.py
-```
\ No newline at end of file
+```
+The sampled output is in the format of SELFIES:
+```
+```
+Note that the SELFIES is an [automata](https://en.wikipedia.org/wiki/Automata_theory), and it terminates when there is no chemical bonds to build. So the converted SMILES could be shorter than the SELFIES:
+```
+```
+
+The advantage of SELFIES is that the output is always a valid molecule: