Molecular Datasets

This repository contains a curated collection of datasets linking molecules to their activity.

Motivations

Many methodologies have been developed to tackle the data inefficiency of traditional machine learning. Most of them use the meta-learning paradigm, where the algorithm is trained on a collections of tasks to extract prior knowledge, which is then used at test time to better generalise on a novel problem.

Although this setting has shown impressive results, the testbeds are for the most part disconnected real-world applications.

Hence, we present two dataset collections that map molecules to scalar values. These can be of value for the community because :

1. This is an actual, real-world problem with tremendous potential. Solving molecular scoring will lead the way to in silico drug optimisation, which has the potential to change healthcare as we know it.

2. Molecules are best represented with graphs, but cheap vectorisation methods exist. As such, molecular datasets can be used to try out an array of different methods.

Collections

Each collection can be seen as a dataset matrix, where the columns are the tasks and the rows are the examples. Note that said matrices are not necessarily dense.

The collections have three main properties, which we report hereafter :

1. The number of individual tasks
2. The number of molecules
3. The density of the collection matrix

TL;DR

Collection	Category	# tasks	# molecules	# examples	density
binding	measured	5 717	573 035	1 088 K	0.03%
antibacterial	measured	3 255	26 993	90 K	0.04%

Description

We provide two measured collections :

• binding is a collection that measures the binding force between a given protein and molecules. In this context, the task is equivalent to the protein.

• antibacterial, wherein we report the measured antibacterial activivty of a set of molecules.

Note that in both cases the collection matrix is sparse : not all (molecule, task) tuples have been tested.