📝 AI_Drug_Discovery_Explainer_2.mdTheme:

AI and Drug Discovery: What Is Solid, What Is Overstated, What Is Open

Companion to the video transcript. Prepared 6 November 2026; revised the same day (version 2) to add section 2, which explains what "fold" means.

A note on sourcing. My attempts to open and read the source pages were stopped partway, so most figures below come from my own knowledge (reliable to about mid-2026) and from search-result headlines, not from pages I read in this session. Treat every number as approximate and check it before quoting it publicly. Company-reported results are labelled as such. A list of leads is at the end.

1. The short version

The video is broadly right that something real changed: computers can now propose new protein molecules that bind a chosen target, and a meaningful fraction of those proposals work when tested in a lab.
It is overstated in three ways. Drug discovery was never purely guessing. "One-shot" design applies mainly to protein drugs (antibodies and similar), not to the small-molecule pills in a typical medicine cabinet. And binding a target is only one of many hurdles between a molecule and a medicine.
The time saved so far is real at the earliest stage (months instead of years) and not yet demonstrated for the whole journey, because human clinical trials still take most of the time and most of the failures.

2. What "fold" means

The one-sentence version. A protein is built as a long, floppy chain, and folding is that chain crumpling itself, through its own chemistry, into one specific three-dimensional shape. "The fold" is the finished shape.

Why the word. Think of origami. A flat strip of paper does nothing useful; folded one way it becomes a crane, folded another way a box. Same strip, different shape, different object. A protein starts as a one-dimensional strip, a line of amino acids (the 20 kinds of building block) in an order dictated by a gene, and ends up as a three-dimensional object. The word covers both the act of crumpling and the shape that results.

What drives it. Nobody folds the chain: there are no hands and no machine. Each amino acid has a chemical character. Some are water-avoiding (hydrophobic, or "oily"); others are water-loving or carry an electric charge. Dropped into the watery cell, the chain collapses much as oil droplets merge in salad dressing. The oily parts bury themselves in the core, away from the water, while the charged parts stay on the surface. Weak attractions between particular pairs (hydrogen bonds, opposite charges) lock local patterns in place. The chain settles into the arrangement where all these pulls are best satisfied, which is its lowest-energy shape. This takes a fraction of a second to a few seconds. Some proteins get help from chaperones, helper proteins that stop the chain from clumping with its neighbours while it folds.

The four levels people mention

Level	Name	Plain meaning
1	Sequence (primary structure)	The order of amino acids. This is the "message".
2	Secondary structure	Short stretches that form repeating patterns: spiral staircases (alpha helices) and flat zigzag ribbons (beta sheets).
3	Tertiary structure	Helices, sheets and loops packed together into one compact three-dimensional body. This is usually what "the fold" means.
4	Quaternary structure	Several folded chains docking together into one machine. Haemoglobin, which carries oxygen in blood, has four.

Why shape equals function. The fold creates pockets, grooves and surfaces. An enzyme works because its fold forms a pocket that fits one particular molecule. An antibody works because its fold presents loops that match a target. A drug works because it fits a pocket in its target protein. The image is a lock and key, though a flexible one, since proteins flex and adjust when something binds. Change the fold and the function goes. That is why drug designers care so much about folds: the pocket they want to hit exists only in the folded state.

When folding goes wrong

Cystic fibrosis. The most common mutation makes the affected protein misfold, and the cell destroys it.
Alzheimer's and Parkinson's disease. Proteins (amyloid-beta and tau in Alzheimer's, alpha-synuclein in Parkinson's) misfold and clump into aggregates.
Prion diseases. The same sequence can exist in two folds, a normal one and a misfolded one that pushes other copies to misfold too. Same message, two meanings.

The "protein folding problem". In the 1960s Christian Anfinsen showed that a protein unravelled with chemicals folds back by itself once the chemicals are removed. So the sequence alone contains everything needed to determine the fold (Nobel Prize in Chemistry, 1972). The hard question was then whether a computer could work out the fold from the sequence. It was hard for two reasons.

Too many possibilities. A chain of 100 amino acids can bend into more arrangements than there are atoms in the observable universe. Trying them all would take longer than the age of the universe, yet real proteins fold in seconds. This is known as Levinthal's paradox. The resolution is that a real chain does not search at random; it rolls downhill on an energy landscape, like a ball finding the bottom of a funnel. Computers could not reproduce that roll accurately for about 50 years.
Measuring folds was slow. The methods are X-ray crystallography (shining X-rays through a crystallised protein), cryo-electron microscopy (imaging flash-frozen proteins with an electron beam) and nuclear magnetic resonance (NMR, which reads the protein's atoms by their magnetic behaviour). A single structure often took months to years of lab work. The Protein Data Bank, the public archive of solved structures, holds well over 200,000, while known protein sequences number in the hundreds of millions.

AlphaFold 2 (2020) predicted folds from sequence at close to experimental accuracy for many proteins, and DeepMind released predictions for around 200 million of them. That is what the video means by "cracked the shape of proteins".

Two directions: prediction and design

Forward (prediction): sequence, then fold. This is AlphaFold.
Backward (design): desired shape, then a sequence that will fold into it. This is called inverse folding, and ProteinMPNN (from Baker's group, 2022) is a well-known tool for it. Generative design as described in the video chains these together: propose a shape that fits the target site, find a sequence that folds into it, check it with a predictor, then test it in the lab. The "one-shot milestone" in the video is the backward direction working well.

Three caveats about "fold"

A fold is a snapshot of something that moves. Real proteins flex, breathe and change shape when they bind. A prediction usually gives a single pose.
Some proteins have no fixed fold. Intrinsically disordered proteins or regions stay floppy. That is where predictors can produce confident-looking but meaningless pictures (see section 3).
"Fold" also names a category. Scientists catalogue proteins into families by overall architecture, and nature uses on the order of a thousand or so distinct folds (the count depends on how you classify). New proteins mostly reuse known ones. In 2003 Baker's group designed Top7, a protein with a fold never found in nature, and it folded as designed. That is what "unnatural" means in section 6: the building blocks are the same 20 amino acids, but the overall shape is one that evolution never produced.

3. "Guessing" and "hallucinating": clearing it up

The commenter is mixing up three different things.

a) Chatbot hallucination. A language model writes fluent text that is false, and nothing in the system checks it against reality. That is a real weakness of chatbots.

b) "Hallucination" as a technical term in protein design. In 2021 David Baker's group at the University of Washington published a method in which a structure-prediction network is run backwards: start from a random amino-acid sequence, then adjust it until the network is confident it folds into a stable shape. The researchers called this hallucination on purpose. The output is a proposal, not a claim. Many of those proposals were then made in the lab, and many folded as predicted.

c) Why the protein case differs from the chatbot case. A designed protein is a hypothesis that physical reality tests. It is manufactured, mixed with the target, and measured. A wrong design costs one failed experiment. A wrong chatbot answer can be believed without anyone ever checking it.

Is it guessing? In the old screening approach, chemists tested vast libraries of existing compounds and kept the few that stuck, which is search by brute force. In the new approach, a model trained on the statistics of known proteins proposes candidates that are far more likely to work than random ones. That is informed proposal followed by testing. Hit rates are still well short of 100%. For example, Chai Discovery announced in 2025 that its Chai-2 model produced working antibodies for roughly 16% of designs tested, across dozens of targets (company-reported, not independently confirmed). Roughly one in six is excellent compared with screening, and it also means most designs fail.

Genuine failure modes that are real and worth knowing:

Out-of-distribution weakness. Models learn from known proteins, which skew towards well-studied, easy-to-crystallise ones. They are less trustworthy on unusual targets.
Disordered regions. Many proteins have floppy regions with no fixed shape. Structure predictors, including AlphaFold, can output confident-looking structures there that are not meaningful. A 2025 paper examined exactly this for AlphaFold 3.
Binding is not a drug. A molecule can bind beautifully and still be useless or harmful (see section 6).
Benchmark leakage. If test targets resemble training data, results look better than they will be on new problems.

4. What the video says, checked

Claim in the video	Verdict
"For basically all of history, we made drugs by guessing"	Overstated. Early drugs came from natural products and accident, but structure-based rational design has been standard since the 1980s (HIV protease inhibitors, imatinib for leukaemia and others).
Almost every drug works by binding a protein	Mostly true. Exceptions exist (some target RNA, DNA, membranes, or act chemically, like antacids).
Testing millions of candidates, billions of dollars, decades per drug	Roughly right for the whole process (10 to 15 years, about US$1 to 2+ billion per approval including failures). Finding a candidate takes closer to 3 to 6 years.
AlphaFold "cracked the shape of proteins"	Largely true for predicting the fold of single proteins. AlphaFold 2 (2020) and AlphaFold 3 (2024, adding DNA, RNA and small molecules) earned Demis Hassabis and John Jumper a share of the 2024 Nobel Prize in Chemistry. Baker received the other half for computational protein design.
Chroma, "Proficient" and Evolutionary Scale design binders from scratch	The names are real but the line blurs three things. Chroma is a protein-generation model from Generate Biomedicines. "Proficient" is most likely Profluent, known for the AI-designed gene editor OpenCRISPR-1. EvolutionaryScale built ESM3, a protein language model that designed a new green fluorescent protein, about half as similar to natural ones as natural ones are to each other, which the authors compared to roughly 500 million years of evolution. Only some of these are about binders.
"Generate the exact molecule you need"	Overstated. You generate candidates, test them, and iterate.

5. Where the focus is

Oncology (cancer) has the most programmes. It has the clearest molecular targets, abundant data, and the strongest commercial incentive. Industry analyses of AI drug companies consistently list it as the largest single area.

Other main areas:

Fibrosis. Insilico Medicine's rentosertib, aimed at idiopathic pulmonary fibrosis (scarring of the lungs), is the best-known AI-found target and AI-designed molecule together. In a 2025 Phase 2a trial of about 70 patients over 12 weeks, patients on the top dose showed a small gain in lung capacity (roughly +98 mL) against a decline on placebo (roughly −62 mL). It was small and early, and it is a promising signal, not proof (company and press-reported figures, from memory).
Immune and allergic disease. Generate Biomedicines has an AI-engineered antibody against TSLP (a signalling protein involved in asthma) in Phase 3 trials, with dosing proposed twice a year. To my knowledge it is among the furthest-advanced AI-designed protein drugs.
Antibiotics. Researchers have used AI to find or design new antibiotics against resistant bacteria such as MRSA (methicillin-resistant Staphylococcus aureus) and gonorrhoea. Commercially unattractive, scientifically exciting.
Gene editing and enzymes. Profluent's OpenCRISPR-1 is an AI-designed editing protein.
Vaccines and designed nanoparticles. The first approved computationally designed protein medicine appears to be South Korea's SKYCovione COVID-19 vaccine (2022), which uses a nanoparticle scaffold designed in Baker's institute.

Dementia and neurodegeneration are harder, and progress there is slower.

The blood-brain barrier blocks most large molecules, so protein drugs such as antibodies reach the brain poorly.
The underlying biology of Alzheimer's disease is still disputed, so choosing the right target matters more than binding it well, and AI does not yet fix that.
The approved Alzheimer's antibodies (lecanemab, donanemab) were not AI-designed.
AI is being used for finding targets, analysing patient data, and designing delivery methods, but I know of no AI-designed dementia drug that has succeeded in late-stage trials.

Status check. To my knowledge as of mid-2026, no drug whose discovery and design were primarily AI-driven has completed Phase 3 and gained full approval. Check this, because it may have changed. One adjacent example is zasocitinib (originally from Nimbus Therapeutics, which used physics-based computer design, now at Takeda), which reported positive Phase 3 results in psoriasis in 2025.

6. Concerns about novel, "unnatural" designed proteins

Medical concerns

Immunogenicity. The immune system may attack an unfamiliar protein and neutralise it or cause side effects. Designed proteins have no evolutionary history to lean on, and clinical data on fully de novo proteins is thin. Prediction tools exist but are not a substitute for human trials.
Cautionary precedent. Neoleukin's computer-designed interleukin-2 mimic (NL-201) was dropped in 2022 after early trials, showing that a clever design is not enough.
Off-target effects. A protein designed to bind one site may also bind similar sites elsewhere in the body.
Upside. Designed proteins are often unusually stable and easy to produce, which helps manufacturing and storage.

Biosecurity (dual-use) concerns

The same tools that design a therapeutic could, in principle, help redesign a harmful protein. In October 2025 a Microsoft-led team published in Science a test in which AI-rewritten versions of known toxins slipped past the software that DNA synthesis companies use to screen orders. The team shared the results with the companies first, and patches were deployed, which caught most (not all) of the variants. Headlines claiming that "up to 100%" evaded screening described the situation before the patches.
The practical chokepoint is DNA synthesis: designs have to be ordered or made physically, so screening orders matters. Many protein-design researchers signed voluntary commitments in 2024 on responsible development, and some model builders filter toxic and viral sequences out of training data.
Open question: whether voluntary measures and screening can keep up with improving models.

Regulatory

The FDA (Food and Drug Administration) judges an AI-designed protein by the same safety and efficacy standards as any other. It issued draft guidance in 2025 on how much to trust AI models in submissions; check whether that has since been finalised.

7. How much time is really saved

Stage	Traditional	With AI	Caveat
Target to candidate molecule	About 4 to 6 years	About 18 months (Insilico, rentosertib; company-reported)	One case, with an unusually favourable target
Design to lab-validated protein binder	Months to over a year	Weeks, in a single lab round (Chai-2; company-reported)	Not independently replicated at scale
Fastest headline	n/a	A 2019 Insilico and University of Toronto study designed and lab-validated kinase inhibitors in 46 days	A well-studied target, and preclinical only
Phase 1 trials (safety)	Historic success about 40 to 65%	About 80 to 90% in a 2024 analysis of roughly two dozen AI-derived molecules	Small sample, selected molecules, safe-ish targets
Phase 2 trials (does it work)	About 40%	About 40% in the same analysis (around ten molecules)	No improvement shown yet
Phase 3 and approval	Several years	No demonstrated change	Still human-time-limited
Whole journey	10 to 15 years	No demonstrated change yet	The honest summary

Reading the table. AI compresses the early, computer-friendly stage from years to months. About 90% of molecules that enter human trials still fail, mostly for lack of efficacy or unforeseen toxicity, and those failures reflect biology that models cannot yet predict. If AI eventually raises the Phase 2 success rate, that is where the large savings would come from. That has not been shown yet.

8. And on, in the Shannon direction

Claude Shannon's 1948 paper treated a message as symbols drawn from a statistical source, with entropy measuring how surprising the next symbol is. A protein is exactly that: a message in a 20-letter alphabet (the amino acids), written by evolution through a noisy process of mutation and selection.

A protein language model is trained to predict a masked or next amino acid, which is the same task that Shannon's 1951 experiment posed to people. In that paper he had human volunteers guess the next letter of English text to estimate how much redundancy the language carries.
Redundancy is what the model exploits: evolution keeps certain patterns, so the "text" is far from random, and a model that learns the statistics can generate new sequences that obey them.
Shannon had an earlier link to biology: his 1940 MIT doctoral thesis, An Algebra for Theoretical Genetics, applied algebra to Mendelian inheritance.
The limit Shannon would point to: a model can learn the statistics of proteins that exist, but whether a new sequence helps a sick human is a separate question the statistics cannot answer. That takes experiments.

9. Suggested reply to the commenter

In protein design, "hallucinating" is a technical term for proposing candidate sequences, and every candidate is then made and tested in a lab. A bad guess costs one failed experiment; it isn't an unchecked claim. The real limits are elsewhere: most designs still fail, and binding a target is far from being a safe, effective drug.