Source: Hacker News — AI in drug discovery – what it is, where we stand and the path forward

Bringing a single new drug to market takes an average of ten to fifteen years. Of the thousands of compounds that enter the early research pipeline, only a handful ever reach patients. Every failure is a cost, and every cost is time — time that matters enormously to people waiting for treatments. AI has started to change that calculus, at least at the front end of the process.

A piece making the rounds on Hacker News this week — "AI in drug discovery: what it is, where we stand and the path forward" — offers a grounded, hype-free account of what AI is actually doing in pharmaceutical research right now. It's worth reading in full. This post distills the key points, adds some honest limits, and then draws a lesson that applies well beyond medicine.

The Protein Folding Breakthrough

Every biological process in the body runs on proteins. Proteins are built from chains of amino acids, but their function comes from their three-dimensional shape — how they fold. For more than five decades, figuring out how a protein folds from its sequence was one of biology's hardest unsolved problems. Determining a single protein's structure in a lab could take months or years of painstaking experimental work.

DeepMind's AlphaFold changed this. Using deep learning trained on databases of known protein structures, AlphaFold can predict the 3D shape of a protein from its amino acid sequence with remarkable accuracy. DeepMind released predictions for hundreds of thousands of proteins in an open database accessible to any researcher in the world. What used to require expensive equipment and months of effort now takes minutes on a computer.

Why does this matter for drugs? Most medicines work by binding to a specific protein — a receptor, an enzyme, a transport channel — and either activating or blocking it. If you know the shape of your target protein, you can design molecules that fit it precisely. Without knowing the shape, you're essentially searching blind through a vast chemical space. With the shape in hand, the search becomes dramatically more targeted.

Virtual Screening: Finding Needles in a Chemical Haystack

Even with a known protein structure, finding which molecule among billions of possible compounds will bind effectively is a formidable task. Traditional high-throughput screening — physically testing tens of thousands of compounds in the lab — is expensive and slow. You test a lot to find a little.

AI-powered virtual screening flips the ratio. Machine learning models evaluate hundreds of millions of molecular structures computationally, ranking them by predicted binding affinity and other properties before a single experiment is run. Only the top candidates go into the lab. This doesn't just save time — it changes the economics of early-stage drug research, making it feasible to explore parts of chemical space that would have been unreachable before.

Generative AI takes this further: rather than just selecting from existing compound libraries, models can propose entirely new molecular structures designed to meet specific criteria. The drug isn't found — it's designed.

What AI Cannot (Yet) Do

Here is where the honest account diverges from the hype. Everything described above accelerates the discovery phase — the search for promising candidates. What comes after discovery is clinical development: Phase 1, 2, and 3 trials in humans. This is where most drugs still fail, and it is where AI's impact is far more limited.

Biology in a living human is not fully simulatable. A compound that looks perfect in a protein binding simulation may behave differently in a cell, differently again in an animal, and differently still in a human body. Toxicity, off-target effects, drug metabolism, immune responses — these emerge from the full complexity of a living system and cannot be reliably predicted from structure alone, at least not yet.

Regulatory requirements — the safety and efficacy standards set by agencies like the FDA — require human trial data. The process of recruiting patients, running trials safely and ethically, collecting and adjudicating endpoints: none of this compresses easily. Clinical timelines are constrained by biology and ethics, not computational speed.

A headline claiming "AI will make drug discovery ten times faster" is usually talking about the discovery phase only. The full pipeline from initial research to an approved medicine remains measured in years, sometimes decades. The failure rate remains high. AI has widened the entrance to the pipeline significantly. The exit is still a long walk away.

The Lesson That Applies to Everyone

Here is the part of the drug discovery story that goes well beyond medicine.

AlphaFold predicts protein structures. But interpreting those predictions, connecting them to disease biology, designing the right experiments, and making judgment calls about which candidates to advance — that requires structural biologists, medicinal chemists, and clinicians with deep domain expertise. AI virtual screening narrows the field. The final selection is still made by experts. AI has not replaced the scientists. It has made the scientists dramatically more productive.

This pattern holds across fields:

• A lawyer using AI can research case law and draft contracts faster — but legal judgment still belongs to the lawyer.
• A marketer using AI can run copy tests and analyze audience data faster — but brand positioning and creative decisions still require human understanding of people.
• A software engineer using AI can write and debug code faster — but architecture decisions, trade-off analysis, and system design remain human work.
• A financial analyst using AI can process more data faster — but the interpretive frame — what matters, what doesn't, what questions to ask — is the analyst's.

The consistent pattern: AI amplifies expertise. It does not produce expertise from scratch. The tools are powerful precisely because they are wielded by people who already understand the domain. A language model given a protein sequence cannot tell you whether to advance a drug candidate. A structure biologist with AlphaFold can do in a day what took months before.

The competitive advantage in an AI-augmented world is not simply having access to AI tools — those are increasingly commoditized. The advantage is knowing where to apply them. Identifying the specific bottleneck in your own workflow where AI leverage is highest. Knowing which outputs to trust and which to check. Understanding when the AI is confidently wrong.

Drug discovery researchers found their bottleneck: protein structure prediction and massive compound screening. AlphaFold and virtual screening tools unlocked those bottlenecks. The question for everyone else is the same: which bottleneck in your field is ripe for this kind of acceleration?

AI has widened the entrance to the drug discovery pipeline considerably. The exit — a safe, effective treatment in a patient's hands — is still a long walk. The walk is still taken by people. In drug development as in every other field, AI is a force multiplier for expertise, not a substitute for it.