Unlocking Porphyrin Power: Predicting Drug Activity with Machine Learning

Hey there! Let’s talk about something pretty fascinating: porphyrins. You might not know the name, but these cool molecules are like nature’s building blocks for things like chlorophyll in plants and hemoglobin in our blood. But guess what? They’re also showing massive potential in medicine, especially for fighting tough diseases like cancer through something called Photodynamic Therapy (PDT).

The idea behind PDT is neat: you get these special molecules, like porphyrins, to hang out in tumor cells. Then, you zap them with light, and they produce little reactive powerhouses that kill the cancer cells. Pretty clever, right? The catch is, finding the *perfect* porphyrin for the job – one that goes where it’s supposed to, works efficiently, and doesn’t cause too much trouble elsewhere – is a huge challenge. Traditionally, this involves tons of lab experiments, which can take forever and cost a fortune.

That’s where we thought, “Hold on, can’t we get smart about this?” And by ‘smart’, I mean bringing in the big guns: machine learning. Imagine being able to look at a molecule’s structure and predict how good it will be at inhibiting growth (that’s what IC50, or its negative log, pIC50, tells us) *before* you even make it in the lab. That’s exactly the kind of shortcut we were aiming for!

Diving into the Data Pool

So, our first step was to gather all the info we could. We hit up a fantastic resource called ChEMBL, which is basically a massive database of chemical compounds and their biological activities. We pulled down data on porphyrin derivatives, looking for their structures and how well they inhibited growth (those precious IC50 values).

Now, real-world data is rarely perfect. It’s like trying to bake a cake with a recipe written on a napkin in the rain – you need to clean it up! We filtered out entries that weren’t measured in a consistent way (like sticking to nM for IC50), removed duplicates (because who needs the same info twice?), and tossed out any entries missing crucial structural details. After all that tidying, we ended up with a solid dataset of 317 unique porphyrin compounds. Sounds like a good number to get our machine learning models chewing on!

Translating Molecules into Numbers

Computers aren’t great at looking at pretty molecule pictures and knowing what they do. They need numbers! So, we used some cheminformatics magic (specifically, a tool called RDKit) to turn the structural information of each porphyrin into hundreds of numerical values called molecular descriptors. Think of these as giving the computer a detailed report card on each molecule, covering things like its size, how many squiggly bits it has, how it interacts with water or oil, and so on.

We calculated all sorts of descriptors – 2D ones based on the flat drawing, 3D ones that consider its shape in space, and even topological ones based on how the atoms are connected. We also paid special attention to descriptors related to Lipinski’s Rule of Five, which is a handy guideline for figuring out if a molecule is likely to be orally bioavailable (basically, can you take it as a pill and will your body absorb it?).

Among all these numbers, we wanted to see which ones seemed to matter most for bioactivity. We looked at how well each descriptor’s value correlated with the pIC50. Interestingly, two descriptors stood out with the highest positive correlation (though still not super strong, around 30-34%): qed (Quantitative Estimation of Drug-likeness) and fr_Al_COO (Fraction of Aliphatic Carboxylic Acids). Qed is a score that tries to capture how ‘drug-like’ a molecule is based on several properties, while fr_Al_COO looks at a specific chemical group. This told us that both general drug-likeness and the presence of certain chemical features might play a role.

Sorting the Chemical Crowd

With our molecules translated into numbers, we could start looking for patterns. We used techniques like hierarchical clustering to group the porphyrins based on how similar their numerical descriptors were. This is like sorting a pile of LEGO bricks by color and shape – you start seeing groups emerge. Our analysis showed that the porphyrins in our dataset fell into about nine distinct groups, but they weren’t super tightly related, suggesting our collection had a good amount of structural diversity.

We also looked at their core structures, or scaffolds, using methods like Murcko scaffold analysis. This helps us see the basic skeleton of the molecules, stripping away the dangling side chains. Identifying the most common scaffolds is useful because it can point to structural backbones that are more likely to be biologically active. For instance, Tetraphenylporphyrin (a core with benzene rings) showed up a lot, likely because it’s easy to make and modify.

Are They ‘Drug-Like’ and Do They Work?

Remember Lipinski’s Rule of Five? We checked how many of our porphyrins followed these guidelines. Out of the 168 compounds we identified as biologically active (meaning they had a pIC50 above a certain threshold), only 31 also met Lipinski’s criteria. This is interesting because while Lipinski’s rule is a good predictor for *oral* drugs, many porphyrins are used in ways that don’t require swallowing a pill (like injecting them for PDT). So, a molecule can be super active even if it breaks some of Lipinski’s rules, which our data confirmed.

We also peeked at some data on tumor response – basically, how well certain porphyrins worked in studies looking at actual tumor reduction. We found a few superstars that achieved a 100% tumor response! Looking at their structures gives us clues about what makes them so effective. It seems like specific chemical groups and even the length of certain chains attached to the core structure can make a big difference.

Putting Machines to Work

Okay, time for the main event: machine learning! We wanted to build models that could predict two things:

• Predict the actual pIC50 value (a regression task).
• Classify whether a compound is ‘active’ or ‘inactive’ based on a threshold (a classification task).

Instead of using the hundreds of molecular descriptors directly, we opted for something called Morgan fingerprints. These are like unique digital barcodes for molecules, capturing structural features in a way that’s often more effective for machine learning. We generated these fingerprints for all our compounds.

Then, we split our dataset into a training set (what the models learn from) and a testing set (what we use to see how well they perform on data they haven’t seen before). To be super thorough, we used a cool tool called LazyPredict, which automatically trains and evaluates *dozens* of different machine learning models for you. It’s like having a whole team of data scientists trying out different approaches at once!

The Results Roll In

For the regression task (predicting the pIC50 value), LazyPredict tested 42 different models. The best performer was something called the Tweedie Regressor, which achieved an R-squared value of 0.63. Now, 0.63 isn’t a perfect score (1.0 would be perfect prediction), but it shows the model can explain a decent chunk of the variation in bioactivity based on the molecular structure. It tells us we’re onto something, but there’s definitely room to improve, maybe by trying more complex models or finding even better ways to represent the molecules.

For the classification task (predicting if a compound is active or inactive), the results were even better! Among the many models tested, Logistic Regression came out on top with an impressive 83% accuracy. This is fantastic! It means our model is pretty good at sorting the promising porphyrins from the less active ones just by looking at their fingerprints. What’s more, Logistic Regression is relatively simple and super fast, making it really practical for quickly screening lots of potential compounds.

We also used a technique called SHAP values to try and understand *why* the models made certain predictions. For the fingerprint-based models, this basically tells us which specific structural fragments within the porphyrin molecules were most important for predicting activity. It’s like getting hints about which parts of the molecule are the ‘active ingredients’.

What We Learned and Where We’re Going

So, what’s the takeaway from all this? Well, first off, our study successfully crunched a whole bunch of data on porphyrin derivatives. We highlighted some of the most studied ones, like TMPyP4 and Temoporfin, which are already known players in the PDT world. We saw that while general drug-likeness (like the qed score) is somewhat correlated with activity, it’s not the whole story. Specific structural features and scaffolds really matter.

Crucially, we showed that machine learning is totally feasible and quite effective for predicting whether a porphyrin derivative is likely to be biologically active. Our Logistic Regression model, with its 83% accuracy, is a great starting point for quickly identifying promising candidates for further investigation. This could potentially save a ton of time and resources in the drug discovery process.

Of course, this is just one step. Our models were validated internally (we split our own data), but the real test is how they perform on completely new data they’ve never seen before. Future work should definitely include testing on an independent dataset. Also, while 317 compounds is good, having even more data would help train potentially more powerful models, maybe even deep learning ones, to capture even more complex relationships between structure and activity. But hey, we’ve built a solid foundation!

Ultimately, this work underscores the exciting potential of combining cheminformatics and machine learning to accelerate the discovery and optimization of therapeutic agents like porphyrin derivatives. It’s all about using smart tools to find the best molecules for the job, helping us move faster towards new treatments.

Source: Springer