Are You A Kinase Looking For Love?
Are You A Kinase Looking For Love?
If yes, this is the right place for you! Our matchmaking agency leverages cutting-edge technology to make sure you find the one that will make your binding pocket melt. Whether you are looking for a lasting binding or a fast thunderbolt, we will find the right fit for you!
We know, dear kinases, how important you are for the proper functioning of living creatures. When you succumb to mutations, we succumb to disease. Of course, with such a busy schedule, it can be hard to find your soulmate. Even worse, many of our clients complain about ligands with commitment. One day their couple looks like a perfect match, then the ligand switches to a similar-looking kinase.
With us, you will have no more frustrations from nonspecific bindings. To find the perfect candidate, we meticulously go through each detail of structure and chemical properties of ligands.
Sounds interesting? Let’s find out how we do our data-driven magic.
What Is A Good Match?
How do you know you are happy with your partner? Does your ligand hug you with the highest possible affinity? Does it shield you from unwanted interactions? In other words, what binding criteria matter to you, our dear kinases?
Seems like everyone is talking about IC50 out there! It is the amount of ligand that makes you slow down to exactly a half of your activity. We understand why you prefer this metric – we too would love to find someone to quickly put us in a relaxed mood. There is another metric that shows promise – Ki. No wonder it is the case, as Ki directly measures the affinity in your couple, what a convenient metric! To make things easier, we will use pIC50 and pKi. This way, bigger values would mean better binding.
Our R&D department found that pIC50 and pKi are, in fact, very correlated. So, good news, we can safely rely on either metric to find your match!
Sorry for being a bit intrusive, but how spicy were your previous relationships? Do you like it burning hot, or are you more of a chill partner? We could definitely use the Temp (C) of your previous interactions to make good predictions. Also, how emphatic were your relationships? Was the ambiance rather sour, neutral, or basic? You guessed it, we will use the pH of your interactions as well.
Great, now that we know everything about your previous romantic endeavors, let’s look closer at what matters to you in terms of you partner’s chemical characteristics!
What Makes A Good Candidate?
The success in our matchmaking business depends on how accurately we gather information on our candidates. We want our clients to be able to decide whether it is a hit or a miss in a blink of an eye. That is why constructing meaningful profiles is so important. Let’s start with the obvious questions – what chemical properties do our clients want in their partners?
RDKit, a chemical detective agency and our dearest contractor, shared its insights with us.
Categorical Features: Functional Groups
Sometimes a smile or a glance can make you fall in love. As it turns out, in the chemical world functional groups are the ones that can produce some chemistry. Let’s see what kind of groups our ligands have.
Continuous Features
Okay, but what about features like weight and the amount of electrons available for some bond making? RDKit has some statistics for that too!
Of course, we asked RDKit detectives to investigate how different chemical features affect binding. As it turned out, the mutual information between the affinity metrics and the chemical properties is close to zero. Well, love is more complicated than some obvious features!
Dating Professionals
Does dating experience matter to you? Some of our ligands have their calendars full of dates. 45 of our candidates were involved in more than 25 dates!
A Match Made In Heaven
Embedding spaces are nice. There, we can see the ligands for what they really are. In an embedding space, inessential information is cut out. As a modern-day matrimony agency, we propose a variety of embedding spaces for you to choose from. If you do not care about the weight or the logP of your partner – which we totally support – take a look at our catalogues of candidates based on several embedding spaces.
Embedding Space From Rdkit
Well, let’s go all-in and look for every available piece of information regarding the chemistry of our candidates! Proud of our partnership with RDKit, an undercover chemical detective agency, we chemically profiled the candidates using all Descriptors provided by RDKit. Of course, nobody would want to go through such extensive profiling for the many candidates in our catalogues – kinases have better things to do – that’s why we propose a conveniently rendered summary obtained through dimensionality reduction techniques, so that our clients could quickly skim through the pages of available ligands and skip a heartbeat when seeing the one. The UMAP plots below show the embedding space of the ligands based on their RDKit descriptors and colored by their binding metrics in order to see if ligand groups can be clustered by their descriptors and if these clusters are related to the binding metrics.
UMAP of PCA-reduced Descriptors
As some ligands cluster together, we conclude that our clients do have their preferences when it comes to binding partners. Unfortunately, we still don’t understand them well enough to predict binding affinities. Maybe we can go further and uncover some deeply rooted preferences that even kinases do not know about?
Mol2Vec Embedding Space
Inspired by our insightful partnership with RDKit, we contacted our next contractor – Mol2vec. This agency does the dirty work of finding a meaningful representation of chemical properties and similarities between molecules for you. It considers the ligands for what they, in essence, are – atoms connected by bonds – and constructs vectors that capture everything you need to know. Straight to the point, so that our kinases do not lose their precious time on ligands that weren’t meant for them from the beginning. Once again, we care about the comfort of our customers, that’s way we propose a dimensionally reduced summary of our findings.
UMAP of PCA-reduced Mol2vec
This doesn’t look much better… We contacted Mol2vec for further explanation but they declined responsibility and accused us of providing ligands that were too structurally similar to be separated. Well, this partnership will not last any longer!
Morgan Fingerprints Embedding Space
Luckily, there is no shortage of embedding contractors and Morgan Fingerprints came to save the day. Let’s take a look at their insights.
UMAP of PCA-reduced Morgan fingerprints
Hmmm, almost the same picture. But do not worry, we will test everything to find the right embedding space for you!
Concatenated Embedding Space
We have gathered all the embeddings from our contractors and combined them into a single embedding space. This way, our clients can have a comprehensive overview of the ligands and quickly find the one that makes their binding pocket tingle.
UMAP of PCA-reduced concatenated embedding space
All our efforts to get that! No, we will not stop there. Let’s explore what machine learning has to offer!
Data-Driven Cupid
Here is the latest hit in matchmaking – predicting binding affinities using machine learning approaches. Of course, our R&D department already has some models to test out!
Multilayer Perceptron
Our economic tier model is a multilayer perceptron that takes the concatenated embeddings as input. Cheap, reliable, and fast. Training with all the new features, from shiny hyperparameter tuning to the latest techniques in data preprocessing, we are ready to make some predictions!
What a shame! Many of our clients were disappointed as the model was not able to predict the binding affinities with a satisfying accuracy. The R&D department insisted that machine learning would bring us billions, so in our last effort let’s try a more powerful model.
Graph Neural Network
For our truly special clients, we have a special model. Directly inspired by what is considered to be state of the art in terms of ML models in drug discovery, we trained our own graph neural network (GNN) to predict affinities. Each ligand was represented as a graph containing the information about its structure. We added some relevant RDKit Descriptors to the nodes and some bond information to the edges of the graphs. Then we trained a GNN that took in input a graph and was trained to output the pIC50 of the ligand. Hours of training and tuning later, we were ready to make some predictions. The model was still not able to predict the binding affinities with a satisfying accuracy.
To be honest, both our models had similar performances with a R² that most of our clients would consider a deal breaker. Proof that even the most advanced models can’t replace the human touch in matchmaking.
Closing Words
This was a rollercoaster! As it turns out, being a matchmaking agency is no joke. We found one of our disappointed clients crying while listening to You Should Be Stronger Than Me by Amy Winehouse on repeat. What can we say? The heart wants what it wants.
