Here’s what’s on our plate today:

• 🧪 How AI is speeding up drug discovery—and what’s still slowing it down.

• 📰 Android audio upgrade, Apple’s AI search, and an AI stylus.

• 🎶 Weekend toolkit: AI song prompts, Replit agents, and model testing tools.

• 📊 Should AI be allowed to invent entirely new drugs?

Let’s dive in. No floaties needed…

The Laboratory

How AI is reshaping drug discovery

Modern medicine is a hallmark of interdisciplinary collaboration, working together to eradicate diseases and prolong human life. Breakthroughs in modern medicine are the result of a process that began with chance discoveries, like Penicillin in 1928 by Alexander Fleming. Rational design, such as beta-blockers that target enzymes or receptors, was followed by computational methods, where computational chemistry and bioinformatics helped drive discovery.

Now, this process is taking another approach. One that relies on deep learning to outperform even the most powerful computational models. Recently, Reuters reported that drug developers are increasingly turning to AI for the discovery and safe testing of new drugs. This is in line with the Food and Drug Administration’s (FDA) push to reduce animal testing. The report is part of a broader push across the world to delve into the use of AI for researching new drugs.

The traditional approach for drug research

The drug research industry is notoriously difficult to navigate. The traditional approach to discovering new drugs has been to first identify disease-related mechanisms and processes, then use computational models to evaluate the ability of chemicals or biological compounds to interact with the disease-related target. When a chemical or biological compound is found to have a positive effect on fighting a disease, it is sent for preclinical research, where they are tested on animals to ensure it is safe. It is only through positive animal testing that a drug is sent for human trials.

Once a drug is successful in human trials, it has to be approved by regulators before it can be released to the public. This process typically takes a decade or more and costs well over a billion dollars, depending on how you count and whether you include the cost of failed programs. This is where AI comes into the picture. AI models are used to speed up drug discovery by helping scientists decide where to aim, what to build, and how to make it safe.

The role of AI in drug research

The ability of AI models to sift through massive amounts of data and find patterns that may be missed by humans makes them an ideal fit for use in drug research. Between 2012-2015, a drug company called Merck sponsored an online challenge where participants had to predict how different chemical compounds might behave. It was during this challenge that deep learning, for the first time, outperformed older techniques. Soon after, researchers showed that these methods could also be used to predict whether a compound might be toxic or whether it might work against certain disease targets.

Then in 2015, a startup named Atomwise introduced AtomNet, a computer model that could look at the 3D shapes of proteins and potential drug molecules, and predict how well they would fit together.

By the 2020s, companies were connecting different AI tools into full loops that can design a drug, make it, test it, and then use the results to improve the next design. These loops often combine information about protein structures, images of how cells react to treatments, and AI systems that can invent new molecules.

Though it does replace the need for real-world lab experiments, the system allows fast-tracking of the development of new drugs. Since the beginning of the use of AI, researchers have been able to identify new antibiotics, molecules, and even enabled the repurposing of drugs. While they are yet to clear clinical trials, it is a promising start.

AI achievements

In 2020, a deep-learning system that screened existing molecules, called Halicin, found that it worked against several dangerous bacteria. In 2023, Abaucin, another machine-learning system aimed specifically at a hard-to-treat bacterium called A. baumannii, was developed. It was effective in mice and lab studies, but has not yet moved into human trials.

AI models have also successfully discovered new molecules like EXS-21546 and A2A, a molecule designed to act on brain receptors. In early safety trials with healthy volunteers, it worked as expected and was safe, allowing studies to move forward in patients.

These are just a few examples of discoveries made possible by AI systems. DSP‑2342 and Baricitinib are also drugs developed by leveraging the power of machine learning. The bottom line: AI‑designed molecules have reached Phase 1/2, repurposing has delivered clinical benefit; however, they are yet to be approved for use in humans, a fact critics often cite.

Lack of real-world medicines, however, is not the only criticism.

The downside of relying on AI

AI models are not flawless and can produce inaccurate results, called ‘hallucinations’. Such flaws can send teams down blind alleys, wasting precious time and effort. Then there is the question of bias. AI/ML models are intrinsically data-driven, as they extract or adapt their parameters from training data. This makes them vulnerable to the integration of bias into models.

Another area of concern is that AI models work on the concept of a black box, which does not sit well with the concept of auditability. As such, for a drug to be approved by the FDA and EMA, the explicit criteria on transparency and risk management are difficult to adhere to.

The same AI tools that design medicines could also be misused to design harmful chemicals. This is not just a theory; scientists have shown that such systems can quickly generate tens of thousands of possible toxic compounds. This is why rules, safeguards, and limits on access are important. While AI can help narrow down the best drug options, it cannot replace the hard work done through the study of chemistry, biology, and clinical trials. Most drug ideas still fail in testing, no matter how they were designed.

Artificial intelligence is not a magic bullet for medicine, but it is already reshaping how scientists look for new treatments. From scanning vast datasets to designing entirely new molecules, AI has shown that it can speed up the earliest and most uncertain stages of drug discovery. In just over a decade, it has gone from outperforming older computer models in contests to generating antibiotics, brain-targeting compounds, and even repurposing existing drugs to fight Covid-19. These achievements highlight both the promise and the pace of progress.

However, in the real world, no drug can be created solely with the use of AI. The long, expensive process of testing safety and effectiveness in people remains a barrier that algorithms cannot overcome. So, for the time being, regardless of the hype, AI remains a powerful new tool in medicine’s toolbox, not a replacement for the toolbox itself.

TL;DR

• Drug discovery is slow and expensive — often taking over a decade and billions of dollars.

• AI is now helping scientists design, test, and refine drug candidates faster than ever.

• Several AI-discovered molecules have reached human trials, but none are fully approved yet.

• Risks remain: bias, lack of transparency, and potential for misuse in creating toxic compounds.

• AI is a powerful accelerator, not a magic cure — real-world testing is still essential.

Friday Poll

Headlines You Actually Need

• reMarkable Pro bets on AI handwriting: The new Paper Pro stylus tablet adds local AI handwriting recognition—no cloud required.

• Apple eyes AI search for Siri: Apple is reportedly developing its own AI-powered search engine to compete with OpenAI and Google.

• Android finally supports multi-headphone audio: Google is rolling out Bluetooth LE Audio support to let Android users share audio to multiple headphones at once.

Weekend To-Do

• Try Suno’s new song prompts: Suno just dropped new prompt-based features that make generating full tracks even easier. Try creating a custom song with just one line.

• Build an AI app with Replit’s Ghostwriter Agent: Use Replit’s new Ghostwriter App Agent to spin up basic AI tools with natural language. No front-end needed—just prompts and logic.

• Compare AI model answers side by side: ChatPlayground AI’s multi-model chat interface lets you ask one question and get responses from GPT-4, Claude, and Mistral. Great for prompt testing.

Rate This Edition