Being a jack of all trades brings to mind someone who has extensive multidisciplinary expertise and is equipped with many tools in their toolbox to solve different problems. A jack of all trades is a great succinct description for computational chemists in drug discovery.
Recently I had a great conversation with Dr. Arjun Narayanan, a Senior Research Scientist at Vertex Pharmaceuticals and a jack of all trades as a computational chemist. In this blog post, I’ll describe what he does as a computational chemist, the problems he solves, and the new tools he’s looking forward to adding to his toolbox.
What is a computational chemist?
A computational chemist brings together expertise in chemistry, physics, statistics, and computer science to design and analyze molecules and their physical interactions. In pharma, they work to understand and predict the behavior of therapeutic molecules.
Arjun is a computational chemist in a project support role on various drug development projects within Vertex. Working hand in hand with medicinal chemists and biologists, he utilizes computational methods such as docking, molecular dynamics, pharmacophore modeling, and machine learning to inform the design decision of which chemistry should be performed next to modify a current set of compounds to find a better compound.
What exactly is a ‘better’ compound that you’re seeking as a computational chemist?
A better compound in a pharma sense means it has a high binding affinity, high selectivity for the correct binding site on the protein, good pharmacokinetics, low metabolism, good bioavailability (high solubility and permeability), and good safety measures. Basically, a potent molecule with as few side interactions as possible.
As a computational chemist you use models to predict these metrics of a compound. You are helping to develop an understanding of the relationship between a measured activity, like IC50, and its structure. For example, what is the predicted effect on the binding affinity if I switch this carboxylic acid with a sulfonamide? If I have crystal binding structure information, will it keep the same number of hydrogen bonding interactions as the carboxylic acid did? How will the torsion energy profile look for the conformations? These are the questions you’re looking to answer to have a full picture relationship between structure and activity to find better and better compounds.
What skills do you need to come into a beginning computational chemist position in pharma?
You of course need to have a strong background in chemistry and biophysics, importantly an understanding of how binding happens and where physical properties come from. Proficiency in Python is a must.
One of the most important skills that is often underdeveloped in incoming computational chemists is their comfort in communication with medicinal chemists. Arjun stresses that since you’re working directly with a medicinal chemist, your ability to ‘talk molecules’ with the chemist is vital. Understanding the phrases ‘protecting group strategies’ or ‘catalyst choice’ are specific examples. The clearer you can communicate your arguments for following a certain scaffold extension, the more likely your contribution will be synthesized in the lab. The reality is there are always going to be more experimental chemists than you, so you have to stand up for your findings.
Why is being a jack of all trades valuable?
Your job is to help the medicinal chemistry team move forward, it doesn’t matter which tools you use to do that, sometimes you need a mix of approaches. Therefore, you need at least a general familiarity with various computational methods such as molecular dynamics, density functional theory, and molecular mechanics.
You’re also just a scientist and can look at the data just as much as anyone else can. Sometimes your impact is asking the right questions and driving the project forward by providing a different perspective. Starting out, Arjun says that the tricky part is knowing when you should use a certain tool over another one, which takes years of experience to know and get right. New people commonly come in with asking how their set of tools can help the project, similar to being a hammer in search of a nail. You need to be open to the idea of learning new things and applying many different approaches to the project. You certainly learn as you go, developing your knowledge base.
Since I am interested in deep learning methods applied to chemistry, I always am interested in the current state of using these predictive tools in industry so I specifically asked the following question.
How are machine learning predictive tools used in a computational chemist’s toolkit?
Arjun typically sees machine learning tools being used for risk-assessment where you can identify compounds that have negative predicted properties thus giving you an insight if you need to change the structure of them to lower their risk further down in development. These tools can be anything from predicting physical properties such as solubility and lipophilicity or safety properties like toxicity and drug-drug interactions. He personally does not do any development of these predictive tools himself. At most pharma companies, like Vertex, Pfizer and Merck for example, they have a separate team that build and train machine learning based property prediction models.
In industry, there are two main approaches taken to design compounds and depending on the approach you take, the usefulness of your property prediction tools changes. One approach is singleton design, targeted designs to answer specific questions, made in a custom way and not a library enabled way. For a lead candidate compound with a difficult synthesis pathway, you need to answer specific questions about its predicted activity. Running through your tools, turns out that it has a high HERG activity or low permeability. Is there a way to modify my designs to get the same upsides but with less risk? For a compound that has these good properties that were hard to get in the first place, using these property prediction tools can be extremely helpful because you know you have this risk. It helps you think how can I modify the design to get the same upsides but with less risks? Is this level of risk worth the reward?
The other approach is the library design. You can run a larger set of compounds that you’re interested in through your models to start selecting subsets or clusters of compounds that have optimal levels of properties tested.
What is a new tool that would be useful in a computational chemist’s toolkit?
On the horizon, Arjun is excited about predicting the late-stage property of solubility of the compound from crystal lattice earlier on in development. Solubility is the driving force for absorption and therefore must be sufficiently high for levels of drug blood concentration to result in a therapeutic effect. Although, changing the solubility of a compound in crystal lattice packing can be difficult. Arjun has run into the problem where the compound packs beautifully into the crystal lattice where you have many hydrogen bonds and pi-pi stacking but posing a difficult problem on how to break the lattice. You don’t want to make many changes to the structure to affect potency, so you’re stuck. Predicting this problem earlier in development could save you this headache.
Predicting solubility is a difficult problem with many energy terms affecting it. You have free energy of conformation, interaction energy between each compound in a unit cell, and asymmetric unit cells. Arjun mentioned he’s interested in trying Open Eye’s Formulation Suite to predict crystal packing in formulation, so he can just add that as another tool to his toolkit.
Hopefully I could provide some perspective why being a jack of all trades is a good description for a computational chemist. The ability to perform free energy perturbation calculations to pharmacophore modeling and including machine learning predictive tools, all enables you to be an effective chemical problem solver.
If you’re interested in hearing other perspectives of what a career in computational chemistry looks like I would highly recommend reading this article. Happy tool building!
