What Is Molecular Machine Learning?

The intersection of machine learning with chemistry has been going on for decades. In recent years, with the advent of sophisticated deep learning methods, machine learning in molecular studies has garnered interest from the scientists’ community. Molecular machine learning has seen tremendous growth in recent years and has also seen an increase in predictions about molecular properties.

By applying machine learning, researchers have now changed the way creativity is being considered, with artificial intelligence can now able to create original images, music and text. With such an advancement in other industries, researchers have now applied machine learning in molecular studies which are popularly known as molecular machine learning.

What is Molecular Machine Learning?

All machine learning approach needs data to learn, and this data is crucial for further studies in the required field, and in this case, the data is necessary on molecules. This molecular data is tough to find, and therefore, it is not available in a large quantity and is very expensive to acquire. But, before this data is fed to machine learning algorithms, it is required to find a way to represent these molecule properties in the form of data, so that the machine can understand and learn. 

Usually, when it comes to molecules, it is known as a bunch of atoms connected through chemical bonds, and machines aren’t capable of understanding this type of data. So, for better understanding, molecular data is being represented in the form of line notation and this system of representing the molecular data in a line notation is called SMILES — Simplified Molecular-Input Line-Entry System. The original SMILES was initiated in the 1980s as an advantageous system to describe the chemical structure of a molecule using ASCII String.

For example, Nicotine, C10H14N2, is represented as CN1CCC[C@H]1C2CCCNC2.

One doesn’t have to worry when it comes to remembering these long and complicated notations as during the time of converting the molecule structure into models, and there are many molecule editors which can automatically convert SMILES string into 2D or 3D models of molecules. Also, it can be easily be converted in Python using the cheminformatics and machine learning toolkits like RDKit.

The problem with SMILES is that the sequence of strings is still not readable by some of the machine learning models as the molecules can come with a different number of atoms and have different sequence length. ML models usually require inputs where the molecule sequence length must be the same. So, to solve this problem, the SMILES strings have been converted into a vector of numbers which can be passed through machine learning algorithms — this method is called molecular featurization.

Once a molecule is featurized, the machine then needs to learn from it. There are many algorithms for learning functions on molecules — simple fully connected networks or a complicated one like graph convolutional technique. So, post such functions, the machine can analyse and hopefully meet the demand for new molecular designs.

Applications

The molecular machine learning has been one of the trending approaches in recent years. It possibly offers a new way of creating molecules. With this as a reference, let’s take a look at the three most critical applications of the molecular machine learning.

Accelerates Development Time

In today’s world, material and medicinal failures are a result of slow research and development, which can be improved by machine learning. The average time for a molecule to go from a piece of knowledge to an application is average ten years and not only time but a lot of money is also required into the creation of a drug or a molecule.

According to Scientifist, the avg time for a drug to complete going from a discovery stage to the development stage is 12 years.

Discovering Materials

Not only does molecular machine learning accelerate the development process, but also helps in identifying new drugs. About 30 years ago, it was believed impossible to predict the structure of even the simplest crystalline solid just from their chemical composition — and till date remains challenging. But in recent years, generation algorithms such as random sampling, simulated annealing, metadynamics, minima hopping and evolutionary algorithms as well as the progress in energy evaluation methods have helped in application of predicting classical crystal structure to a broader range of molecules of solid forms. However, these methods are still highly expensive when it comes to computation and spending energy.

Predicting Material Properties

The machine learning methods have helped in predicting the property of a large number of elements. This prediction of properties comes into play when specific attributes of materials help in better performance of applications. For example, machine learning can predict material properties like superconductivity, bulkiness, shear, and band gaps.

Outlook

Machine learning has shown strong potential to compete with conventional molecular computation. Molecular machine learning methods have the potential to revolutionise the research on material properties. However, it has hit certain limitations like lack of standard evaluation platform for the proposed algorithms, lack of expertise for obtaining precise and accurate results, and lack of data, which creates problems in transforming of molecules into a suitable form that machines can read. With so much potential and vast scope for research and discoveries, it will be interesting to see where molecular machine learning ends up in future.