The rise of artificial intelligence (AI) over the past year has raised a lot of concerns, like the risk of fake news and job losses. People are worried about deepfakes messing with public opinion, biased algorithms, and more.
But amid all this worry, there’s a bright side where these technologies can actually speed up scientific progress and help tackle big issues. A great example is in healthcare, especially pharma and life sciences, where AI is starting to change the process of drug discovery, which is usually really expensive, slow, and full of setbacks.
Traditional drug discovery pipeline
Drug development, traditionally, is a complex and time-consuming journey. It kicks off when scientists figure out what they want to target, usually an organic compound, all twisted and shaped in complex ways. They use computer programs to see a digital model of this protein spinning around on the screen, also known as in silico modelling. They search for spots where a new compound might attach, kind of like a ship finding a place to dock. Building this compound happens one atom at a time.
Once the compound is made, scientists test it on living cells in controlled settings. However, the results can be disappointing. Many cells die, and it’s often unclear why. Biological systems are complicated, which makes these experimental drugs pretty unpredictable. Scientists then have to come up with new versions, tweaking and refining them over many years. Even after a long research time, some of the most promising compounds can turn out to have some issues, such as carcinogenicity, adverse effects, drug-drug synergies with new medications, long-term resistance patterns, or immunological complications, and as a result get scrapped.
If luck is on their side, researchers might find a compound that works well in a preclinical discovery and only then move with it to the clinical trials.
Such a conventional development process can be characterised by three sobering statistics:
- Timeline: Over a decade from initial discovery to market approval.
- Cost: More than $2 billion to bring a single drug to market.
- Attrition rate: Only about 1 in 10 drug candidates successfully pass through clinical trials and receive regulatory approval.
The high-risk, trial-and-error approach makes drug development one of the most resource-intensive endeavours in the industry. However, AI, especially things like large language models and generative AI, have the potential to redefine that process, promising, maybe not revolutionary changes overnight, but subtle, meaningful improvements across the whole workflow.
The path to AI-powered drug discovery
Back in 1981, Fortune magazine's cover declared that computers were set to change drug discovery for the better. The article talked about how scientists were using computer visualisation to figure out which molecules to test. While computers did change things up, the biggest hurdles in drug development didn't just go away with a simple click of a mouse. Scientists began exploring whether neural networks, the latest form of machine learning, could uncover patterns in patient data, helping to explain why some folks responded to treatments while others didn't. As time went on, it became clear that AI could do much more than just spot patterns in medical records.
However, the breaking point happened in 2020 when AlphaFold, an AI system developed by DeepMind, proved it could predict how proteins would fold into their final shapes. This AI essentially solved a problem that had baffled scientists for years: figuring out the three-dimensional shapes of proteins from their amino acid sequences, leveraging templates and evolutionarily related sequences. The significance of this breakthrough was recognised in 2024 when the Royal Swedish Academy of Sciences awarded it the Nobel Prize in Chemistry.
So, how does AI drug discovery actually work?
In this process, each compound in the pharmaceutical formulation, including the drug molecule, excipients, and components in suspensions and colloidal systems, is seen as a data point with its own set of traits: its chemical structure and structure-activity relationship (SAR), its pharmacokinetics and pharmacodynamics, and its ADMET (absorption, distribution, metabolism, excretion, and toxicity). These compounds are studied through high-throughput screening of massive datasets filled with numbers that describe them in different ways.
For instance, each candidate drug is evaluated based on how well it binds to its target, how the body handles it, its safety, stability, and potential effectiveness. When these models are trained properly, molecules with similar therapeutic potential or chemical features tend to be closer together in the computational space.
How AI transforms each stage of drug development
Target identification
Target identification is a crucial step in the fields of chemical genetics and drug discovery. This process involves determining which gene or protein is responsible for a disease. AI processes large and complex datasets that integrate various layers of biological information simultaneously.
In simpler terms, AI analyses all this information together, creating a comprehensive and interconnected understanding of how some chemical or protein structure should be changed (or have another property) to be more efficient for some disease. This holistic perspective is something a human researcher would struggle to compile manually from such vast amounts of data.
Industry examples
For instance, let's consider AstraZeneca, though it's worth noting this particular example about a narrower field closer to genetics, more aligned with AlphaFold than classical drug discovery. They have announced plans to analyse up to 2 million genomes by 2026. Their goal is to identify subtle genetic variations that may be linked to disease progression and patient responses to treatment. Processing this vast amount of genomic data is only possible because AI can effectively manage such a large scale of information.
Virtual screening
AI models can be used to predict how different compounds will interact with specific proteins. Such an approach can help quickly analyse a large number of potential molecules, from thousands to millions, in a much shorter time than traditional experimental methods would take.
Industry examples
Tools such as AtomNet and Schrödinger's Drug Discovery Suite use deep learning to predict how well small molecules will bind to protein targets, significantly reducing the time it takes to find promising drug candidates.
Molecular generation
This is where AI moves beyond just screening existing libraries of chemicals. AI shifts to de novo design and molecular docking, actually generating and evaluating totally new molecules. This is where models like LLM-based MoEs with reasoning (or scientific multi-agent AI), or other types of generative models come into play. In addition to conducting high-throughput screening of existing molecules, these AI agents/models can actually generate completely novel molecular structures.
The AI can optimise these novel structures as it designs them. It optimises them for specific properties simultaneously, such as their potency, predicted safety and ADMET properties. So, it's designing and optimising for multiple factors all at once, right at the start. Advanced techniques, like Generative Adversarial Networks (GANs) and reinforcement learning (RL) algorithms, are particularly effective for this type of de novo drug design, as they can generate and evaluate molecules with specific desired properties through iterative learning processes.
Industry examples
Insilico Medicine developed a novel drug candidate for a lung disease called idiopathic pulmonary fibrosis, using an artificial intelligence platform. They went from target to a candidate ready for preclinical testing in just 18 months. Exscientia designed a really potent inhibitor for a target called PKC-Theta, using an AI generative design platform, in 11 months. These timelines were previously unimaginable, considering the fact that traditionally preclinical process take anywhere from 3-6 years.
But it's important to note that those timelines are generally for the discovery and optimisation phase, getting from the target idea to a candidate molecule that looks promising enough to start the formal clinical process.
Synthesis
AI's impact extends beyond computational design into physical production through the "lab-in-the-loop" approach. Data from real-world experiments, lab tests, and clinical data feeds back to train the AI models. The improved models then make better predictions, maybe about the compound properties, maybe about how to synthesise it. And those predictions get tested in the lab again, generating new data, which feeds back again to retrain the models. It's this continuous loop of prediction, verification, and refinement.
Industry examples
IBM has an AI tool for chemical reactions estimation, called RXN for Chemistry, that uses deep learning to predict the best chemical steps to actually synthesise. AI can suggest routes chemists might not have thought of, saving months of planning.
Clinical trials
Even in the later stages involving actual patients, AI can bring significant efficiencies:
- It can help optimise the design of the clinical trial itself, the protocol.
- It can analyse data to predict which patients are most likely to respond to the drug or who might be at higher risk for side effects. This helps in patient stratification.
- AI can also help predict potential trial outcomes based on early data.
Challenges preventing wider adoption of AI in drug discovery
Despite the technology's promise, the likelihood of AI-discovered molecules completing all clinical phases successfully predicted to improve from 5–10% to about 9–18%, no AI-designed drugs have reached the market yet. While companies like Recursion and Insilico have advanced candidates through phase II clinical trials, demonstrating safety in patients, a key question remains: if AI is so beneficial, why isn't it used to develop every drug?
- Key main challenges and roadblocks:
-
- Data quality. AI models, especially deep learning, need vast amounts of high-quality, diverse data to learn effectively. But biological data can be messy, inconsistent, and incomplete. Sometimes it's locked away in proprietary silos. If that's the case or the data is biased, the model's predictions won't be accurate or reliable in the real world. Poor data leads to poor models and inaccurate predictions.
- Black box problem. We don't always know how the AI reached its conclusion. It can be very difficult, sometimes impossible, to trace back exactly why the model predicted, say, that a specific molecule would be effective or why it flagged a safety concern. In medicine, where lives are on the line, that lack of interpretability, that lack of transparency, understandably makes people nervous about reliability and accountability. And it's a major hurdle for regulatory agencies; they need to understand why a decision was made before approving a drug.
- Regulatory adaptation. The technology is moving incredibly fast, even faster than the regulatory frameworks can keep up. We need new frameworks to handle things like data privacy links, especially with sensitive health data, establish clear guidelines for how to validate AI models used in drug development, how to ensure data integrity, and how to handle the ethical considerations. We saw the EU formally adopt its AI Act in August 2024, which is a big step towards setting standards. But it's still an ongoing process.
- Skills shortage. The biotechnology sector increasingly requires specialists who can work at the intersection of life sciences and AI. However, there is a shortage of professionals who are skilled in both areas. This lack of talent makes it difficult to effectively use AI solutions in existing processes and limits chances for major breakthroughs in biotechnology. Organisations might need to create training programs or partner with universities to develop this specialised skill set.
- Financial barriers. Adopting an AI solution demands considerable capital expenditure. Companies need to spend on important infrastructure, like powerful computers and storage, along with appropriate software platforms. On top of that, training employees to use AI and machine learning systems adds to these costs. This financial pressure can be too much for smaller biotechnology companies and new businesses, making it hard for them to benefit from AI technology.
- Data privacy. Protecting personal information is one of the key ethical issues when using AI in healthcare. Collecting and analysing confidential medical information raises concerns about patient consent and the risk of data misuse. Healthcare organisations need to set up strong security measures to keep patient data safe and follow laws like the US Health Insurance Portability and Accountability Act (HIPAA). It's important to be clear and honest with the public about how data is collected, why it is used, and how it is shared.
Final thoughts
The use of AI in the drug discovery process is changing the game when it comes to tackling one of medicine's toughest challenges. The idea that algorithms can create new drugs in just a few months sounds incredible, and companies like Insilico Medicine and Exscientia are showing that we might be at a real turning point.
Still, we need to keep our expectations in check; this isn't a magic fix for all diseases, it's more like a powerful set of tools that can help make a slow process a lot better.
Related Insights
