Find a open source sample dataset to be used for development and testing.

Select a molecule -> (1) Identify all MMPs based on the most common transformations. Compare the predicted difference and the actual difference accross the MMPs
(2) Identify MMPs based on the molecule to "neutral" fragment transformation. Predict the target property for the identified molecules. Use the difference as an estimate of how much that fragment contributes to the prediction

Building on H4, investigate the following:
1. How many molecules fall within X distance (in feature space) of the selected molecule?
Maybe X is a function of number of features? Or a constant.
Explanation: this could be used as a measure of how well a specific area of chem space is represented, at least based on the selected feature set.
(To discuss: what is chem-space? Is it a function of molecules and the features selected? E.g. even if we have all the molecules but we only count the number of carbons as a feature, we're not really capturing a wide chem space)

2. What is the variance of the target values for molecules within the radius (defined above)?
Explanation: high variance within similar molecule could indicate: activity cliff-ish effect if localised, features not capturing the underlying science enough to differentiate between molecules.

which features to be decided. will need to look at MP prediction papers and what's freely and easily available.
Try grouping the features into categories based on what they intend to capture, so specific features could be "turned off" to show the functionality of the library.

Will probably split this into multiple issues later.

Select a MMP transformation-> compare the predicted vs actual difference across the MMPs with the selected transformation

Rather everything into one place and evaluate the state of the project. See which components work / which dont. Have an internal discussion regarding improvements / new ideas.

set-up a ML pipeline to train a set of ML models based on the sets of features calculated.
Details to be confirmed later.

The Stream of Consciousness

• Wouldn't it be useful to know what part of the molecule contributes to certain parts of the molecule?

Possible solution ?

• Could we run descriptors over each fragment to figure what their contribution would be? Granted this won't work for everything.

Select feature -> (1) get example molecules from parts of distribution (2) most common fragments in specific part of distribution
select fragment -> how widely distributed are molecules with the specific fragment? (i.e. if for a descriptor, molecules with the given subject are only found in one part of the distribution, it suggests that that desriptor is related to that fragment)

Select fragment -> distribution of predicted error of molecules with the fragment vs general population

Select a feature, select a fragment -> compare the distribution of the feature between molecules with fragment vs general population

Problem

It's always annoying as a scientist when you are not sure exactly what kind of chemical space your data covers. Typically people will say to you:

But how do you know you're covering the correct chemical space?

Idea

A set of simple descriptors should be used to describe the possible chemical space and the space your data occupies.

Current suggestions :
Molecular Weight, Fragment Types

Not sure how to best present this information?
% Complete? I.e. of the range of fragments available your dataset covers x% as a whole and the least/most counts are as follows....

splitting #5 into two tickets.

Come up with categorisation of discriptors.
Come up with a clever way to store these categorisations.

set of feature and performance explainability hypothesis to test

https://github.com/Trusted-AI/AIX360

See how easy it is to use
what the required inputs are
judge if it's worth integrating with it

For a given set of descriptors and a selected molecule, what are the most similar molecules? (based on distance in selected feature space).
Something similar but fragment based? e.g. average descriptor value for mols with the fragment and compare to averages of mols with other fragments.

get rdkit, try mmpdb, see how fragment data is stored / retrieved.
see if this is going to be suitable

Present the key findings somewhere to get an outside perspective on usefulness as well as ideas for future PoCs and improvements

select feature -> get distribution within the dataset.
select molecule -> show where in the distribution the molecule falls.