# Predicting Chemical Properties of Compounds from their Molecular Structure with Machine Learning
# Mentor Presentation on Tuesday

[Slides](https://drive.google.com/drive/folders/1XU96ZPkY2S1AwjmQteE8Aoace4GIt5bl?usp=sharing)

[Original biochemistry paper recording results](https://pubs.acs.org/doi/pdf/10.1021/jm00106a046)
# Dataset

## Loading the data
Link to google drive [folder](https://drive.google.com/file/d/1KNsSzFT1cEcve98BvB--J7eZCQOlCU3o/view?usp=sharing).

Alternative: [network repository](https://networkrepository.com/Mutag.php)

Importing the data in Python from pickle file:

import pickle

dataset = pickle.load(open('mutag_processed.pkl', 'rb'))

The dataset is a list of 188 dictionaries, where each dictionary describes the molecular structure of a compound in the dataset. The molecules have the following attributes:

* 'vert_type': a 1-d numpy array of integers from 0 to 6. The ith entry is the atom type of the ith vertex in the molecular structure graph
* 'edge_list': a list of tuples $((u,v),b)$, where $(u,v)$ is an edge in the molecular structure graph, and $b$ is the bond type of the edge, an integer from 0 to 3.
* 'class': a binary label {0,1} where 1 indicates that the molecule is a mutagen

# Discussion Log - 12th July 2022
1. Think of the property of a matrix that has nothing to do with the ordering of the matrix.
2. What the condition number mean? Conditioning measures how rapidly the output changed with tiny changes in input
3. Keep in mind of matrix, we will find it useful later.
4. Cross Validation
5. Think about Feynmann diagrams, it has a lot of similarities with this problem
6. Plot of number of edges vs number of vertices. Will this work? Possible feature
7. Belief propagation. Possible approach?

## TASKS - day 1
1. General properties of a graph (degree no./distribution etc.)
2. KL divergence
3. Spectral properties (eigenvalues, chateristic polynomial, trace)
4. Coding adjacency matrices

### Feature Extraction for Graphs
* number of edges/vertices
* number of connected components (cycles)
* node degree (histogram)
* eigenvector centrality
* clustering coefficient
* Hellinger distance (?)
* edge cutting (?)

## Slides
### Page 1

1. What is the problem?
* Mutagen: A compound is a mutagen if it causes mutations in cells. (Suggestion: put some emjois)
* We have data with molecules. How can we determine which molecules are mutagen or not?
* It is difficult to classify if a molecule is a mutagen or not, so before we use algorithms/ML, we have to do extract and exploit features.

2. Main Challenge - Feature Extractor
* Construct feature map
* Learn decision boundary in feature space

### Page 2
1. Feature candidates
* Eigenvalues differences and bounds of adj
* Characteristic polynomials
* Number of pentagons and hexagons
* Number of cut edge/cut vertex (connectivity)
* Maxzimun/Minimum degree
* (Closed) Euler trail
* Hamilton path/cycle