Update README.md

README.md

@@ -1,3 +1,60 @@
----
-license: mit
----
+---
+license: mit
+---
+
+## Overview
+
+- 3.4M 3D small molecule structures and thier associated HOMO-LUMO energy gap and SMILES string. 
+- Molecules range from 2 to 53 atoms (including hydrogens) 
+- 22 different elements appear in throughout this dataset 
+
+Practical Relevance: Density Functional Theory (DFT) is a powerful and widely-used quantum physics calculation that can accurately predict various molecular properties such as the shape of molecules, reactivity, responses by electromagnetic fields. However, DFT is time-consuming and takes up to several hours per small molecule. Using fast and accurate ML models to approximate DFT enables diverse downstream applications, such as property prediction for organic photovaltaic devices and structure-based virtual screening for drug discovery.
+
+Overview: PCQM4Mv2 is a quantum chemistry dataset originally curated under the PubChemQC project. Based on the PubChemQC, we define a meaningful ML task of predicting DFT-calculated HOMO-LUMO energy gap of molecules given their 2D molecular graphs. The HOMO-LUMO gap is one of the most practically-relevant quantum chemical properties of molecules since it is related to reactivity, photoexcitation, and charge transport. Moreover, predicting the quantum chemical property only from 2D molecular graphs without their 3D equilibrium structures is also practically favorable. This is because obtaining 3D equilibrium structures requires DFT-based geometry optimization, which is expensive on its own.
+
+Prediction task and evaluation metric: The task is graph regression: predicting the HOMO-LUMO energy gap in electronvolt (eV) given 2D molecular graphs. Mean Absolute Error (MAE) is used as evaluation metric.
+
+more details on this dataset and benchmarking can be found here
+https://ogb.stanford.edu/docs/lsc/pcqm4mv2/
+
+### Preprocessing 
+56 structures in the original training set have been removed after rdkit failed to load and parse them. The original indicies of these structures are listed in 'invalid_structures.csv'
+
+in both the test-dev and test-challange splits 3 SMILES strings were removed after rdkit failed to parse them. The original indicies are listed in 'invalid_test-dev.csv' and 'invalid_test-challenge.csv'.
+
+### Data Splits
+Original datasplit: 90/2/4/4 
+
+- 'train' : 3378550 molecules, 3D structures, SMILES and labels
+- 'valid' : 73545 molecules, SMILES and labels only
+- 'test-dev' : 147037 molecules, labels only
+- 'test-challange' : 147432 molecules, labels only
+
+### Node Labels
+None
+
+### Molecule Labels
+- 'SMILES' : SMILES string 
+
+- 'gap' : the HOMO-LUMO energy gap in electronvolt (eV)
+
+### Structure Distribution
+number of points (atoms including hydrogens)
+
+SEE GITHUB PAGE FOR FIGURES
+
+![Structure Training Set](./train_histogram.png)
+
+![Structure Validation Set](./valid_histogram.png)
+
+![Structure Validation Set](./test-dev_histogram.png)
+
+![Structure Validation Set](./test-challenge_histogram.png)
+
+
+## Sources 
+Data origin:
+https://ogb.stanford.edu/docs/lsc/pcqm4mv2/
+
+Original publication
+https://pubs.acs.org/doi/full/10.1021/acs.jcim.7b00083