AI in Drug Discovery: Chasing Dreams, Facing Realities

After completing my PhD, ages ago by now, I remember standing at a crossroads. My research had taken me deep into the world of cisplatin and its binding to DNA, and now, a new chapter was starting. Around me, former colleagues and friends were venturing into computer-aided drug design, applying various computational tools to streamline the drug discovery process. I was torn. A part of me even toyed with the notion of a dramatic field switch, to computational astrochemistry… but in the end I decided to stay and delve even deeper into the investigation of the interactions of potential drugs with their biological targets. There was so much room for creativity in using DFT, docking, and molecular dynamics to model those beautiful systems. This was a time when machine learning still felt like a plot point in a science fiction novel. And today? Well, we’re knee-deep in a different narrative: AI is the panacea for all our problems, right? But before we get too carried away, let’s tap the brakes on the hype machine and ask: what’s the real deal with AI in drug discovery? That, my reader, is the question we are addressing today.

AI in Drug Discovery: Chasing Dreams, Facing Realities | From Atoms To Words | Arturo Robertazzi

AI in Drug Discovery: Why Do We Need It?

12-15 years. That’s the average time it takes to transform a spark of insight in a lab into a brand-new drug on pharmacy shelves. 12-15 years and an avalanche of dollars. Roughly 2.5 billion, give or take. 

But why’s drug discovery such a challenge?

First off, you’ve got to identify the biological target involved in a disease. This could be a fragment of a nucleic acid or a protein-based entity, like an enzyme or a receptor. Just identifying that target is an epic quest on its own.

But hang onto your pipettes, because that’s just the beginning.

Once you have your target, you need to design a drug that can grapple with the disease. And that’s when the molecular mayhem really kicks off.

As we’ve previously explored on From Atoms To Words, screening the chemical space to find a specific molecule is like trying to find a grain of sand of a very specific blue in the Sahara desert. Quite literally.

Chemical space is vast. Insanely vast, featuring a staggering 1022 to 1060 possible combinations. Your potential wonder drug is hidden in that gigantic desert.

So, what’s the game plan?

AI in Drug Discovery: Chasing Dreams, Facing Realities | From Atoms To Words | Arturo Robertazzi
Copper-based anticancer agents binding their biological target, a DNA triplet. Adapted from Robertazzi 2009

You dive into the discovery stage with a legion of candidates and start whittling them down via techniques like high-throughput screening of compound libraries against these targets.

For most drugs, the spotlight’s on small molecules. These molecular marvels are then passed to the medicinal chemists, who tweak them into more effective forms, smoothing out any rough edges.

If that work pans out, you level up to preclinical trials.

This phase is all about running a gamut of tests tracking the drug candidate’s journey in vivo — its metabolism path, interactions, and, ultimately, its exit strategy during excretion. Safety and dosage are also part of the deal, and only the molecules that clear these tests move on to the next step — the clinical trials.

Sound like a Herculean task? Well, it is. This discovery and preclinical stage takes about six years.

Aiming to trim timelines, cut costs, and boost success rates, we’ve armed our medicinal chemists with an arsenal of computational techniques. Think computer-aided drug design, where we mix in a slew of computational methods—molecular docking, quantitative structure-activity relationship, DFT, molecular dynamics, and more.

These computational tools have been helping and enhancing, sure, but the road to fame for a potential drug is still long and winding.

Now, with a new buzz sweeping through the digital ether, there’s been some serious chatter about AI potentially turning those grueling preclinical stages into a cakewalk.

But is AI the real deal in drug discovery? And if it is, how exactly is it going to disrupt the drug discovery process?

Further reading: AI’s potential to accelerate drug discovery needs a reality check (2023)

AI in Drug Discovery: Chasing Dreams, Facing Realities | From Atoms To Words | Arturo Robertazzi
Antibacterial sulfonamide drug locked inside its biological target (PDB: 1AZM).

More on From Atoms To Words:
Large Language Models for Chemistry: Is the Beginning of a New Era?
60 Years in the Making: AlphaFold’s Historical Breakthrough in Protein Structure Prediction
Is Machine Learning Going to Replace Computational Chemists?

The Dream: What could AI do for Drug Discovery?

AI has been a force of transformation. Think of advancing speech and image recognition technologies. It also brought us the revolutionary AlphaFold, changing the game entirely in protein structure prediction. Don’t just take my word for it; see for yourself.

With such feats under its belt, it’s not a stretch to envision AI streamlining every stage of the drug discovery pipeline, honing the efficiency and efficacy of the process from start to finish.

Let’s take a closer look. Shall we?

▸ Biological Target identification

Our quest for the perfect drug begins with zeroing in on the right biological target, one that has a causal link to the disease. By hitting this target, we hope to mitigate the disease, while steering clear of a barrage of side effects.

When discussing biological targets, we’re often—but not always—referring to proteins. So, what’s our goal? To inhibit their action, alter their structure, or maybe block a specific region or surface. In essence: pinpoint that one protein that’s causally linked to our disease and smack it with a drug.

Here, AI’s unique strength in sifting through massive data volumes and spotting patterns becomes invaluable. With machine learning, you could:

  • Identify the target: Machine learning algorithms can scrutinize gene expression profiles and protein interaction networks, finding potential targets critical to disease-related mechanisms.
  • Determine causal relationships: Graph Neural Networks and tree-based methods can uncover causal relationships between genes and diseases.
  • Characterize and classify biological targets: Decision tree meta-classifiers can predict morbidity-associated genes, presenting opportunities for novel drug targets. Machine learning approaches can also classify proteins as either potential drug targets or non-targets.
  • Analyze the scientific literature: Large Language Models can give their share of help by parsing through the extensive array of scientific literature. They can extract key target-disease associations, enriching databases for target identification.
  • Infer drug-target interactions: Large language models can also identify potential drug-target interactions by comparing molecular descriptors with those of known compounds.
▸ Fast screening of compounds

This is where, in my opinion, AI truly comes into its element, virtual screening. Particularly through the creation of predictive models that detect compounds likely to bind a biological target.

AI models are typically trained on rich datasets that include known protein-ligand complexes, structural details, and molecular descriptors, such as solubility, partition coefficient, degree of ionization. Once you identify a potential candidate, AI can even support its synthesis, planning efficient routes for chemical reactions.

Now, remember computer-aided design? Machine learning can be employed to amplify the efficacy of certain computational methods. I am thinking of machine-learning-enhanced molecular dynamics and docking, for instance.

▸ Beyond the Preclinical Phase

When you’ve pinpointed the target and designed a potential drug, the next big question is, how will the biological system respond?

This is the critical moment where similarity and feature-based machine learning methods strut onto the stage. These tools aim to forecast the response of drug-target interactions, estimating specifics like binding affinity and the free energy of binding.

But why stop at predicting drug responses? Imagine using AI to meticulously choose patients for preclinical trials, identifying relevant human-disease biomarkers, sidestepping potential toxic side effects, or navigating through a maze of clinical variables to handpick the perfect patient group.

And what if AI could even predict the results of clinical trials, drastically reducing risks for participants? It’s a tantalizing prospect, though maybe we’re tiptoeing into a (dystopian?) sci-fi story here.

So, let’s get back to reality. While all this is in principle within reach of what AI may accomplish, it begs the questions: where do we stand now? What have we really achieved so far with AI in drug discovery?

Further reading: AI in drug discovery and its clinical relevance (2023)

AI in Drug Discovery: Chasing Dreams, Facing Realities | From Atoms To Words | Arturo Robertazzi
Molecular dynamics simulation showing the release of a inhibitor from a metalloenzyme. Animation Credit

More on From Atoms To Words:
The Lifesaving Hunch: How Rosenberg’s Unexpected Discovery of Cisplatin Changed Medicine
When Will RNA Structure Prediction Get Its AlphaFold Breakthrough?
Chemical Space to Material Discovery: Simulations and Machine Learning Leading the Way

The Reality Check: What’s Holding up AI in Drug Discovery?

It’s undeniable: AI-driven drug discovery startups are stealing the limelight, fast-tracking molecules into trials. But what is the real-world impact? What’s the status of AI in drug discovery?

To answer these questions, let’s go through a recent report published in Nature, by Meier and colleagues. They put AI-driven drug discovery under the microscope, analyzing 24 AI startups against the backdrop of the top 20 pharma heavyweights. Their findings are quite eye-opening.

Unlike the big players, AI companies tend to go for the tried-and-true therapeutic targets. Enzymes like kinases are their darlings, making up over 60% of their focus. Why? Well, it’s partly because they want to play it safe. Going after established targets reduces risks and lets them focus on the fine-tuning of their machine learning models.

According to Meier, these AI companies are already flaunting about 160 discovery programs and preclinical candidates (with roughly 15 assets braving the clinical trials). This is indeed pretty impressive, especially if you consider that the combined output of AI discovery and preclinical trials represents about 50% of the portfolio of the top 20 big pharma companies.

But here’s the million-dollar question: will AI-derived drug candidates shorten the process, reduce costs, and, most importantly, be more successful in the clinical trials?

Only time will tell, yet we can’t ignore some recent hiccups. Take, for instance, the case of Exscientia, which made headlines when it had to halt its Phase I/II trial for their highly-anticipated AI-generated cancer drug. Then there’s BenevolentAI and Recursion Pharmaceuticals, each nursing their bruises from similar setbacks.

The crux of these stumbling blocks, in my opinion? The intrinsic complexity of patient biology, notoriously tough to replicate in computer-aided design (in silico), lab settings (in vitro), or even in living models (in vivo).

Biology’s intricacy is incredibly context-dependent, as varies from person to person. It’s kind of like trying to solve a jigsaw puzzle, with a billion pieces, all coming from different boxes.

Quite a challenge, right? So, what’s our move?

We need models smart enough to get the quirks of every individual’s biology, while also catching the common threads that run through all that data. Because, at the end of the day, finding those golden similarities can make or break decisions in drug discovery.

Yup, drug discovery has never been easy for humans. It’s still pretty hard for AI.

Further reading: AI in small-molecule drug discovery: a coming wave? (2022)

More on From Atoms To Words:
7 Noncovalent Interactions in Proteins: The Hidden Architects of Structures and Functions
Multiscale Simulations of DNA: From Quantum Effects To Mesoscopic Processes
Can Quantum Chemistry Simulations Help Trace the Origin of Life?

A Final Personal Touch

It’s probably everybody’s expectation that AI will grow to play a pivotal role throughout the drug discovery pipeline, leading AI-derived drug candidates to excel in clinical trials and ensuring the process becomes faster, more cost-effective, and ultimately, more successful.

Yet biology isn’t an open book. It’s rather a cryptic tome that AI must learn to read from cover to cover.

Besides the scientific complexity of drug discovery, there’s also a philosophical roadblock to contend with: what exactly does it mean for a drug to be AI-discovered? As AI plays an increasingly crucial role in drug development, with machine learning models becoming integral tools in the hands of humans, where do we draw the line for its contributions? And how does this ambiguity affect our ability to truly assess the effectiveness of AI in drug discovery?

The thing is, we’re just scratching the surface, and the landscape of AI in drug discovery is riddled with questions. But of this, I’m certain: if AI and machine learning models continue to evolve and prove their mettle in real-world clinical settings, they will reshape pharmaceutical research.

Maybe not today, but who’s to say not tomorrow?

AI is set to play vital roles in our lives, making critical decisions on our behalf. Should we brace for existential threats? Not a fan of the word “existential,” but caution is definitely our ally.

For example, pondering over AI in drug discovery, especially in clinical stages, I find myself caught between optimism and a touch of unease. I see a gigantic ethical elephant in the room: the role of AI in patient selection for the trials, to name one. Is a future where AI holds the reins in such significant decisions something we desire?

It’s quite a lengthy conversation, I admit it. What I know is that such an idea gives me the same chills as seeing Arnold Schwarzenegger on a Harley.

If you enjoyed this dive into AI in drug discovery, I’d love to hear your thoughts. Agree, disagree, or have a totally wild theory of your own? Let’s connect! Subscribe to my LinkedIn newsletter and let’s keep the conversation rolling.


From Atoms To Words

Today’s Story Filed Under: