Can AI replace Structural Biologists?
Navigating the Intersection of Artificial Intelligence and Structural Biology
One afternoon, I was sitting at my computer running Coot, painstakingly building an atomic model from a ~ 3.5 Å (Ångström)-ish electron density map of a protein I had just spent nearly a month refining from a massive cryo-EM dataset. I was hooked, and my curious mind was completely locked into the puzzle — trying to trace the backbone and visualize the crucial sites I was interested in.
Out of nowhere, a colleague (friend) of mine walked up, stared over my shoulder at my Mac, and asked: “Have you seen what Google DeepMind just released? Don’t you think your career is over? If AI can solve any protein structure in a few minutes, why would anyone still hire structural biologists?”
I looked back at him, smiled, and said: “Yeah, I have seen, and I don’t think our career is over”.
But before I tell you why I smiled, let me take you behind the scenes of “blobology” for a moment — and introduce you to Coot, the tool that helps turn those mysterious blobs into something that actually makes sense. Coot is a classic piece of software with a bird logo that structural biologists use to build atomic models into 3D density maps. After months and years of hard wet-lab work and tedious data processing, you finally get what looks like a floating blob. It’s true that a 3D electron-density map can look like a mysterious cloud, but inside that cloud is the backbone of your protein you have been trying to understand for months to years — so, here comes Coot to resolve your protein structure — “It’s like building a string of amino acids, analogous to weaving a chain with beads”.
Now, let’s get back to my colleague and decode my ‘smile’ at him — why AI isn’t going to make us obsolete anytime soon.
First, let’s see what AlphaFold 3 brings to the Table
When AlphaFold 2 came out in 2021 (led by John Jumper and Demis Hassabis), it felt like science fiction; a real fantasy came true! Suddenly, researchers could predict millions of protein structures with some sort of accuracy, which wasn’t possible in the past with other known software.
Later in 2024, AlphaFold 3 (AF3) took another giant leap (again led by John Jumper). It doesn’t just predict individual proteins anymore — now it can model molecular complexes, including ligands, DNA, RNA, ions, glycans, etc. Definitely this was a major breakthrough! His AF3 protein structure prediction helped him win the Nobel Prize in Chemistry in 2024 along with Demis Hassabis and David Baker.
Simply remarkable! No doubt that it’s an extraordinary achievement.
But despite all the excitement, one question keeps resurfacing: Will AI replace structural biologists?
The short answer is “NO”.
In fact, AlphaFold 3 highlights something interesting: predicting a three-dimensional structure is only one step toward understanding how biology may work.
What Makes AlphaFold 3 Unique?
Instead of juggling different tools for different tasks, AlphaFold 3 can manage everything in one place — it can now predict how proteins, DNA, RNA, and drug molecules interact with each other.
Basically, its biggest upgrades come down to three clever engineering shifts:
- Diffusion-Based Generation: Instead of assembling a structure frame-by-frame, AF3 starts with random atomic noise and gradually denoises it into a coherent 3D shape.
- A Streamlined Core: It replaces the heavy Evoformer from AF2 with a lighter Pairformer module, making it less reliant on massive sequence alignments (MSAs) and more focused on atomic physics.
- A Unified Molecular Representation: Working directly on atomic coordinates allows AF3 to model proteins, DNA, RNA, drug molecules, glycans, and chemical modifications all in one go.
Instead of combining several specialized tools, scientists get a single, adept platform that handles complex biological systems in one step. But such impressive predictions are not the same as how biology works in our cell.
Biology Doesn’t Stand Still
One of the biggest misconceptions about protein structures is that they may behave like rigid sculptures. The reality is ‘they don’t’.
Proteins are dynamic molecular machines, constantly changing shape as they interact with other molecules, carry out chemical reactions, connect with cellular partners, and respond to their surroundings.
1. Proteins Exist in Different Conformations
Proteins are rarely static — they constantly move and form different conformations.
A dynamic membrane protein naturally alternates between open and closed states depending on whether a ligand is bound or other stimuli. When researchers tested AF3 on a particular protein, it consistently predicted only the closed conformation, failing to reproduce the experimentally observed other states like open, occluded, or semi-occluded.
Because of the dynamic nature of these proteins, the AF3 model isn’t necessarily ‘wrong’ — rather, it simply favors one structural snapshot instead of capturing the full conformational landscape of that protein in solution.
Here is where the fun role of a structural biologist comes in — resolving alternative states that depend on a multitude of factors by identifying the exact conditions, such as pH, salt, detergent, lipids, and ligands. These conditions can give us clues about the mysterious function of a protein, backed by experimental validations of those alternative conformations. This is often where the real biology happens.
2. AI Can Hallucinate Structure
Generative AI models are incredibly good at recognizing patterns. Sometimes they’re too good.
Get Suhaila Rahman’s stories in your inbox
Join Medium for free to get updates from this writer.
In flexible or intrinsically disordered regions, AlphaFold 3 can generate well-defined structural features where no stable structure actually exists in nature.
Although newer training strategies reduce these errors, distinguishing genuine biological architecture from AI-generated artifacts still requires human eyes — experimental evidence and expert interpretation.
The real truth is—“A beautifully folded prediction may not automatically be a genuine protein conformation.”
3. Physics and Chemistry Still Matter
Unlike molecular simulations that explicitly enforce physical force fields, AlphaFold 3 learns stereochemistry from data patterns.
That approach occasionally produces physically unrealistic structures. As a result, it can occasionally output physically impossible geometry — like inverted chiral centers or severe atomic clashes where protein chains interpenetrate each other.
These errors remind us that even highly accurate AI models don’t fully obey the laws of chemistry.
4. Antibodies Are Inherently Unpredictable
Antibodies are among the trickiest molecules for AI to model accurately. The regions that actually recognize and bind to antigens — the complementarity-determining regions (CDRs) — can change rapidly through somatic hypermutation. As a result, they often don’t carry the evolutionary signals that prediction tools like AlphaFold typically rely on.
Finding a convincing antibody–antigen model can be a bit like searching through a haystack. Researchers may generate hundreds — or even thousands — of predictions using different random seeds before they find a structure that looks biologically plausible.
The “Black Box” Debate
One reason AlphaFold 2 transformed structural biology was its openness. Researchers could use the codes and improve their data.
AlphaFold 3 took a different path. Instead of a fully open-source release, it initially became available through a web server with usage limits and commercial restrictions. That decision sparked a lot of debate within the scientific community.
At its core, science depends on being able to reproduce and verify results. If researchers cannot independently examine or reproduce an AI-generated prediction, it becomes much harder to test, improve, or fully trust the results.
No matter how sophisticated the AI is, a prediction is still just a hypothesis until experiments prove it.
AI Is Only as Good as Its Data
Perhaps the biggest irony is that AlphaFold’s success is built on decades of work by experimental structural biologists.
AlphaFold learned from the Protein Data Bank (PDB), a vast collection of protein structures that researchers spent years and decades — and countless experiments — determining with techniques such as X-ray crystallography, NMR spectroscopy, and cryo-EM.
Every structure in the PDB represents years of experimental effort.
Without scientists continuing to solve new structures, AI would eventually run out of fresh samples to learn from.
Novel protein folds, transient intermediate states, membrane protein complexes, and unusual post-translational modifications still require experimental determination before they can become part of future AI training datasets.
In other words, AI doesn’t replace structural biology — it feeds on it.
Co-Pilot, Not Autopilot
Rather than replacing structural biologists, AlphaFold is becoming a useful partner. It helps us explore different possibilities, test ideas, and move faster, while we still make the final call based on the science and the experimental evidence.
It can speed up initial model building in tools such as Coot, help us design better truncated constructs for gene expression, generate hypothesis-driven models of drug binding, and identify key residues to target in functional assays.
The modern research workflow isn’t Human vs. AI; it’s a collaborative loop.
Without experimental validations, AI predictions are just a hypothesis. Without AI, structural biologist will struggle little longer to get the desired structures.
AlphaFold 3 hasn’t ended the jobs of structural biologists — it has accelerated and opened up its most exciting chapter yet to come. Although AI can predict what’s possible, it still takes structural biologists to delve into it and prove what’s real.
And that’s exactly why our career and this field are just getting started — we’re gonna make a good team!
The Bigger Picture
AlphaFold 3 has changed the game in structural biology. It gives scientists a new way to explore protein structures and opens the door to questions we couldn’t easily tackle before. But biology isn’t straightforward, nor easy.
Since proteins change shapes all the time in our body, such structural snapshots from AF3 doesn’t tell us much about the biology — which is the main goal of doing any kind of science. The cellular environment is complicated, and it constantly reshapes its molecular machinery in response to countless signals. Capturing that complexity requires much more work than an algorithm — it requires functional data, validations, careful interpretation, and, ultimately scientific judgement. AI can give us possibilities, but it takes a structural biologist to figure out which ones actually hold up in the real world.
That’s why I don’t see AlphaFold 3 or any similar AI tool as the end of structural biology. I see it as the beginning of a new chapter — one where AI and experimental scientists work together, each bringing something the other cannot, to answer questions that neither could solve alone.
