AI in Science: From AlphaFold to Drug Discovery

1. AI moved from reading science to making it

2. The AlphaFold moment

note

What AlphaFold actually did

AlphaFold2, released by DeepMind in 2020, predicts protein structure from sequence. Its success at CASP14 marked a major milestone in computational biology.

Why the result mattered

A protein’s function depends on its shape. If you know the shape, you can often infer:

• where a ligand may bind
• which residues are likely active
• how a mutation may change function
• whether a protein complex is plausible

Important limits

AlphaFold is powerful, but not omniscient:

• it is less certain for flexible loops and disordered regions
• it does not fully solve protein complexes or dynamics by itself
• it predicts a likely structure, not a guarantee of biological behavior

diagram

illustration

note

The real breakthrough

The breakthrough was not only accuracy. It was scale. A structure that once took specialized effort could now be estimated for millions of proteins. That changed what questions scientists could ask on Monday morning.

equation

$$\text{GDT} = \frac{1}{N}\sum_{i=1}^{N} s_i$$

3. AI is now designing molecules and materials

4. Human-AI collaboration in the lab

5. Ethics, trust, and what counts as knowledge

Transcript

Welcome to Slate. Today we're looking at AI in Science: From AlphaFold to Drug Discovery. We'll cover How AI is accelerating drug discovery and materials science, The AlphaFold moment and what came after, Collaborative human-AI research workflows, and Ethical questions around AI-generated knowledge. Let's get into it.

For years, AI in science mostly meant search and summary. It could scan papers, extract entities, and rank results. That helped. But the bigger shift came when models started making useful predictions about the physical world. A protein sequence is like a sentence written in four letters. The old problem was: can we read that sentence and predict the folded shape it will become? In 2020, DeepMind’s AlphaFold2 did that with striking accuracy in the Critical Assessment of Structure Prediction, or CASP14. That was a turning point because structure is not trivia. Shape controls function. Enzymes, receptors, and antibodies all depend on it. Once a model can predict shape, it can help scientists ask new questions faster. Here’s the key idea in the visual: AI is no longer only a librarian. It is becoming a lab assistant that can propose structures, suggest candidates, and narrow the search space before expensive experiments begin. That matters because biology and chemistry have enormous design spaces. A typical small-molecule drug can have hundreds of atoms arranged in countless ways. Humans cannot test them one by one. AI helps turn that impossible search into a manageable shortlist.

AlphaFold is famous because it solved a problem people had worked on for decades: predicting a protein’s three-dimensional structure from its amino-acid sequence. At CASP14 in 2020, AlphaFold2 reached a median Global Distance Test score above 90 for many targets, which is roughly the level biologists consider near-experimental accuracy for many folds. In 2021, the AlphaFold Protein Structure Database launched with predictions for hundreds of thousands of proteins. By 2022, it covered more than 200 million proteins, including many from the UniProt database. That scale changed the daily workflow of structural biology. Before, a lab might wait months for X-ray crystallography or cryo-electron microscopy, and many proteins were still hard to solve. After AlphaFold, researchers could inspect a plausible structure in minutes and decide whether an experiment was worth pursuing. The important nuance is that AlphaFold did not end structural biology. It changed the starting point. Scientists still need to validate flexible regions, protein complexes, binding states, and effects of mutations. The model is strongest on single-chain structures; biology is often messier than that. But it gave researchers a map where there had been fog.

Once AI can score structures, it can help with design. In drug discovery, the goal is not just to find any molecule. It is to find one that binds the target, survives the body, avoids toxic effects, and can actually be made. That is a multi-objective problem. Generative models help by proposing candidate molecules instead of waiting for chemists to enumerate them manually. A model may search chemical space by optimizing properties such as binding affinity, solubility, and synthetic accessibility. This is where the analogy of a chess engine helps. A chess engine does not play one move at a time in the dark. It evaluates many possible futures and keeps the promising ones. AI molecule design works similarly, except the board is chemistry and the rules are physics, biology, and manufacturing. In materials science, the same logic applies. Researchers use AI to propose catalysts, battery materials, and alloys with desired properties. For example, models can screen thousands of candidate crystal structures far faster than human teams can synthesize them. The result is not magic. It is triage. AI helps scientists spend their limited lab time on better bets.

The best scientific workflows are collaborative. A scientist asks the question. The model proposes candidates. The lab tests them. Then the scientist decides what to trust. This loop is powerful because each side covers the other’s weakness. AI is fast, consistent, and good at ranking huge spaces. Humans are better at framing the problem, spotting artifacts, and deciding whether a result makes sense in context. In practice, a research team may use AI to suggest a protein mutation, then run molecular dynamics, then test binding in vitro, then revise the model based on the result. That is not a one-way pipeline. It is a conversation. The visual here shows the feedback loop clearly. The important point is that good AI science is not “human versus machine.” It is division of labor. The model can be wrong in systematic ways. It may overfit to training data, miss rare chemistry, or produce overconfident scores. A careful scientist checks those failure modes. The workflow succeeds when the model saves time without hiding uncertainty.

AI-generated science raises hard questions. If a model proposes a molecule, who is responsible if it fails in patients? If a model helps write a paper, how do we check that the claims are real? If a model learns from public data, how do we handle consent, attribution, and access? These are not abstract concerns. In drug discovery, a false positive can waste millions of dollars. A false negative can hide a useful treatment. In biology, a confident but wrong model can send a lab down the wrong path. So the standard has to be higher than “the model said so.” The evidence must be reproducible, transparent, and testable. That is why many groups insist on external validation, preregistered analyses, and clear provenance for training data. The next phase of AI in science will not be judged only by headline accuracy. It will be judged by whether it helps produce reliable knowledge. The best systems will make scientists faster without making them lazier. That is the real test: not whether AI can generate answers, but whether it helps us earn answers we can trust.