Targeting KRAS — A Tutorial for High School Biology

Why this matters

One single protein, mutated in about a quarter of all human cancers, was considered impossible to target with drugs for over thirty years. Then, in 2021, the first medicine that blocks it was approved. This tutorial walks through how that happened — and what it has to do with the tiny GTP-switching machine inside your cells called KRAS.

By the end you'll be able to explain:

What KRAS does in a healthy cell
What goes wrong when KRAS mutates
Why drug developers called it "undruggable" for decades
How a clever new approach finally cracked it
How modern AI tools like Boltz are speeding up the next generation of drugs

What is KRAS?

Picture a single cell in your body. It's surrounded by other cells, all sending chemical signals like "grow," "stop," "divide," or "die." The cell has to decide what to do — and inside, a relay team of proteins passes those signals from the outside membrane to the nucleus, where genes get switched on.

KRAS sits right at the start of one of the most important relay teams: the RAS–MAPK pathway. It's a small protein attached to the inside of the cell membrane, and it works like an on/off switch.

Figure 1. KRAS flips between an OFF state (bound to GDP) and an ON state (bound to GTP). When it's ON, it tells the cell to grow and divide.

The switch works by binding small fuel molecules. When KRAS holds a molecule called GTP, it's ON and shouting "grow!" to the next protein in the chain. When it chemically clips GTP down to GDP, it's OFF and quiet. Normally, this happens in quick pulses — flick on, flick off — only when the cell genuinely needs to divide.

Go deeper: KRAS is a GTPase

KRAS belongs to a family of proteins called small GTPases. The "-ase" ending means enzyme, and the enzymatic job is to hydrolyze GTP — that is, snip a phosphate off GTP using water to turn it into GDP. This hydrolysis is what flips the switch back to OFF. KRAS isn't very good at this on its own, so it gets help from accessory proteins called GAPs (GTPase-Activating Proteins) that speed up hydrolysis, and GEFs (Guanine nucleotide Exchange Factors) that help kick off the old GDP so a fresh GTP can bind.

What goes wrong in cancer

Cancer is, at its core, a disease of cells that won't stop dividing. Now imagine KRAS's on/off switch gets jammed in the ON position. Every moment of every day, the cell is hearing "grow, grow, grow," even when no real signal is coming in. The cell divides when it shouldn't. Its daughter cells inherit the same jammed switch. A tumor begins.

A single mutation in the KRAS gene — often a swap of one amino acid for another at position 12 of the protein (written as something like G12C, meaning glycine became cysteine) — is enough to break the switch. The mutated KRAS can still bind GTP, but it loses much of its ability to chop GTP back down to GDP. The protein gets stuck in the ON state.

The scale of the problem. KRAS mutations are found in roughly 90% of pancreatic cancers, 30–40% of colorectal cancers, and 25–30% of lung adenocarcinomas. Across all human cancers, RAS-family genes (KRAS, HRAS, NRAS) are mutated in about a quarter of cases. That makes KRAS one of the single most important cancer-driving proteins ever discovered.

Go deeper: why "G12"?

Position 12 in the KRAS protein sits right next to the spot where GTP is held and chopped. Even a small change here — glycine, the smallest amino acid, swapped for something bulkier like cysteine, valine, or aspartate — physically blocks the chopping reaction. The most common variants are G12C, G12D, and G12V. Each one is a slightly different jammed switch, and each one needs its own drug.

Why KRAS was called "undruggable"

KRAS was discovered as a cancer-causing gene in the early 1980s. For the next three decades, drug companies poured billions of dollars into finding a molecule that could block it. They all failed. By the 2000s, KRAS had earned a reputation as the "undruggable" protein. Why?

Two main reasons:

1. GTP binds way too tightly. Most drugs work by sitting in a pocket on a protein and blocking whatever normally fits there. For KRAS, that "whatever" is GTP, and GTP is present inside cells at very high concentrations and binds KRAS with extreme affinity. Any drug would have to out-compete a flood of GTP — almost impossible.

2. The surface is too smooth. When researchers solved the 3D structure of KRAS, they found that outside the GTP-binding site, the protein's surface was almost featureless. There were no deep grooves or pockets for a small drug molecule to grab onto. It was like trying to find a handhold on a billiard ball.

Figure 2. The "smooth surface" problem. KRAS's GTP site is occupied and inaccessible, and the rest of the protein offers no obvious place for a drug to bind.

Go deeper: why pocket shape matters

Drug molecules are typically small organic chemicals (a few dozen atoms). For one to bind a protein with useful strength, it needs to fit snugly into a pocket — making contact with the surrounding amino acids on multiple sides at once, the way a key sits in a lock. A flat, featureless surface gives the drug nothing to hold onto, so it just floats away. This is why finding "druggable pockets" is step one in modern drug discovery, and why a protein without obvious pockets is a nightmare target.

The breakthrough: targeting G12C

The wall finally cracked around 2013. A research group led by Kevan Shokat at UCSF tried something different. Instead of attacking KRAS in its busy ON state, they looked at the rarer OFF state (bound to GDP) — and they zoomed in on the G12C mutant specifically.

G12C is the version of KRAS where glycine at position 12 has mutated to cysteine. Cysteine is special: it has a reactive sulfur atom (–SH) sticking out of it that no other amino acid has. The Shokat group designed a drug that would chemically bond to that one specific sulfur, locking the protein shut. Because regular healthy KRAS doesn't have a cysteine at position 12, the drug would ignore it and only attack the cancer-driving mutant.

There was one more surprise. When the drug snapped onto the cysteine, it pulled open a small hidden pocket on the protein — a pocket nobody had noticed before — called the switch-II pocket. This was the "handhold" everyone said didn't exist.

Figure 3. The drug latches onto the mutant cysteine at position 12 and wedges into the newly-revealed switch-II pocket, locking KRAS in the OFF state.

From that initial chemistry, two drugs were developed and tested in patients with lung cancer driven by KRAS G12C:

Sotorasib (brand name Lumakras), approved by the U.S. FDA in May 2021, became the first ever KRAS-targeting medicine. Adagrasib (Krazati) followed in late 2022. Both are pills. Both shrink tumors in many — though not all — patients with KRAS G12C-mutant lung cancer.

What this proves. A protein being "undruggable" usually just means we haven't been clever enough yet. The KRAS story shows that targeting a specific mutation, in a specific protein state, with chemistry tailored to that one mutated amino acid, can crack a problem that resisted brute-force screening for thirty years.

Go deeper: covalent vs. non-covalent drugs

Most drugs bind their targets non-covalently — they sit in a pocket but can drift back out. Sotorasib and adagrasib are covalent drugs: they form an actual chemical bond with the cysteine, permanently. The protein is destroyed (so to speak) for the rest of its life. The cell eventually makes new KRAS molecules, which is why patients have to keep taking the drug, but each individual KRAS protein hit by the drug is one-and-done. Covalent drugs used to be considered too dangerous because they can attack the wrong protein, but designing them to react only with a mutation-specific cysteine sidesteps that problem.

Go deeper: what's left to do

G12C is only one of the KRAS mutations. G12D, which is more common (especially in pancreatic cancer), has no reactive cysteine to grab onto, so it needs a totally different chemical approach. As of 2025 there are G12D inhibitors and even broader "pan-KRAS" inhibitors in clinical trials, several using non-covalent designs that mimic the shape of the switch-II pocket. Also, tumors often become resistant to G12C drugs after months of treatment — they evolve new mutations that block the drug from binding. The next generation of KRAS drugs has to anticipate and stay ahead of that resistance.

Where Boltz and AI fit in

Designing a drug like sotorasib used to require years of painstaking X-ray crystallography — freezing proteins into crystals and shooting X-rays at them to figure out their 3D shape. Without a 3D structure, you can't design a molecule to fit it. Without a lot of structures (the protein with the drug, without the drug, with mutations, in different states), you can't really understand the system.

In the last few years, AI has flipped that. Tools like DeepMind's AlphaFold and the open-source Boltz model can predict a protein's 3D structure from its amino acid sequence alone, in minutes, with accuracy that often matches experimental methods. They can also predict how proteins, drugs, DNA, and other molecules fit together.

What this means for KRAS hunters. A researcher can now type in a sequence for, say, a KRAS G12D mutant and get back a predicted 3D structure showing where the new pocket might form. They can virtually dock thousands of candidate drug molecules into that pocket overnight, before ever ordering a chemical. Boltz also predicts how the protein flexes and changes shape — which matters because, as we saw, the switch-II pocket only opens up when something pulls on the cysteine. Modeling these motions is what makes "undruggable" targets approachable.

The Boltz lab page you shared (standard-biotech-division-vzI1) is the kind of workspace where scientists run those predictions today: feed in a sequence, get out a structure, ask "where could a drug bind here?" and iterate. The KRAS story took thirty years of trial and error. The next undruggable target might take three.

Go deeper: how structure-based drug design actually works

The workflow looks roughly like this: (1) get the structure of the protein you want to target — experimentally or, increasingly, from a model like Boltz. (2) identify a pocket on the protein where a drug could plausibly bind. (3) computationally "dock" libraries of candidate small molecules into that pocket and score how well they fit. (4) take the top scorers and actually synthesize and test them in the lab. (5) iterate: improve the chemistry, retest, repeat. AI structure prediction massively accelerates step 1 and makes steps 2–3 cheaper and faster, which means more shots on goal.

A guided tour of the Boltz interface

This section walks through the actual screens a researcher uses on lab.boltz.bio to set up and run a virtual drug screen — the same kind of workflow that scientists use to hunt for new KRAS inhibitors. Each diagram below shows a real screen with arrows pointing at every important button.

The big picture. A Boltz project moves through four screens: pick a target (the protein you want to drug, like KRAS), set up an experiment (the screen itself), browse the candidate molecules the AI returned, and either triage them one by one or design your own to test.

Screen 1 — Workspace home (your list of projects)

This is what you land on after logging in. Each card is one drug-discovery project. For a high schooler thinking about KRAS, the mental model is: one project per protein you want to drug. You'd start a new project called "KRAS-G12C" and everything that follows — targets, experiments, candidate molecules — would live inside it.

Screen 2 — Inside a project

A project holds two main things: the target (the protein you're trying to drug) and one or more experiments (the screens you've run against that target). The example shown here is for BRD-4 — a protein involved in some cancers — but for KRAS you'd see exactly the same layout. You'd add a KRAS-G12C target up top, then launch an experiment to virtually screen a chemical library for molecules that bind to the switch-II pocket.

Go deeper: what's a chemical library?

The card mentions "Enamine REAL." That's a real catalog of about 50 billion synthesizable molecules — meaning each one can actually be made by Enamine if you order it. A virtual screen takes that giant library and uses AI to ask, for each molecule, "does this fit the pocket on my protein?" The 1,002 candidates in the tutorial are the molecules that scored well enough to be worth looking at by a human.

Screen 3 — Table view (browsing 1,002 candidates)

The four tabs at the top are how you move through the data: Overview shows summary statistics across all 1,002 candidates. Table (shown) is a sortable spreadsheet view. Triage walks you through molecules one at a time. Design lets you draw your own.

Filters and sort chips in the upper left let you slim down 1,002 hits to a manageable shortlist — e.g., "show only molecules with Binding Confidence above 0.9 and no chemistry alerts."

Alert Level and triage buttons on each row let a human reviewer quickly flag (⚐), thumbs-up (↑), or thumbs-down (↓) each candidate. This is how a 1,002-molecule list gets whittled down to the ~50 worth synthesizing in the lab.

The 3D viewer on the right shows the target protein with the currently-selected molecule docked in. For KRAS this is where you'd literally see your candidate drug sitting in the switch-II pocket.

The properties chart at the bottom is a "parallel coordinates" plot — every line is one molecule, traced across every score at once (binding confidence, drug-likeness, size, lipophilicity, etc.). Lines that stay high across the board are your best shots.

Go deeper: what do those score names mean?

Binding Confidence: the AI's estimate (0–1) that the molecule actually binds the pocket. Structure Confidence / pLDDT / ipTM: how sure the AI is about its predicted 3D structure overall and at the binding interface. CLogP, TPSA, Lipinski: classic chemistry rules of thumb for whether a molecule is likely to behave like a real drug (oral absorption, etc.) — these come from the Lipinski "rule of five" you may have seen in chemistry class. Alert Level: a flag for known-problematic substructures (e.g., chemical groups linked to toxicity).

Screen 4 — Triage view (one molecule at a time)

Triage view is where the chemist actually looks at each molecule. You see the structure, the 3D pose of how it sits inside the target protein, all the AI-predicted scores on the left, and all the drug-likeness descriptors on the right. Then you make a call: thumbs-up to keep, flag to discuss with the team, thumbs-down to reject. Arrow keys move to the next molecule. For a KRAS screen you'd be looking at hundreds of these to find the few that look like real drug candidates.

Screen 5 — Design view (draw your own molecule)

This is the most fun screen for a high schooler. You can draw any molecule you want — pick atoms (H, C, N, O, S, halogens) and bond them together, or drag in pre-built rings like benzene. When you submit it, Boltz predicts how (and how well) it would bind your target. So if you were studying KRAS, you could literally design your own candidate inhibitor, click Submit Molecule, and within minutes the AI would tell you whether the switch-II pocket would accept it.

The whole loop, in one sentence. Boltz lets you take a protein you want to drug, hand it a library of millions of molecules, get back a manageable list of the best-scoring ones with predicted 3D poses, filter and triage them down to a final shortlist — and even draw your own molecules to test new ideas the AI didn't think of. That's the workflow behind essentially every modern drug discovery program, KRAS included.

Go deeper: from screen to lab

A passing score in Boltz is just an AI prediction — it's still a hypothesis. The molecules that survive triage get ordered (or synthesized), then tested in actual biochemical assays (does it actually bind purified KRAS protein in a test tube?), then in cells (does it shut off the growth signal?), then in animal models, then in patients. The whole pipeline takes years, but tools like Boltz make the very first stage — "which of these billions of molecules are worth even looking at?" — go from years to minutes.

Discussion prompts

Use these in pairs, small groups, or as a whole-class conversation. There are no "right" answers — the point is to apply what you just learned.

The switch analogy. KRAS is described as an on/off switch for cell growth. What other on/off switches can you think of in biology? Why do you think cells use switches rather than just continuous signals?
Mutation specificity. Sotorasib only works on KRAS G12C, not on G12D or G12V. Why is that a clinical problem? Why might it also be a clinical advantage?
"Undruggable" reconsidered. If KRAS was undruggable for thirty years and then suddenly wasn't, what does that tell you about how scientific knowledge actually progresses? Can you think of other examples in science of a problem that seemed impossible until it wasn't?
AI in biology. Boltz and AlphaFold can now predict protein structures in minutes that used to take years to determine experimentally. What are the upsides for medicine? What might be lost, or what new risks might emerge, when AI replaces hands-on lab work?
Resistance. Cancer cells treated with sotorasib often develop new mutations that make the drug stop working. How is this similar to bacteria becoming resistant to antibiotics? What strategies could researchers use to stay ahead of resistance?
Cost and access. Sotorasib costs roughly $17,000 per month in the United States. Who should decide who gets access to expensive precision medicines like this? Is it fair that the same drug costs different amounts in different countries?
Your turn. If you were a researcher with a Boltz account and unlimited compute time, which disease or protein would you try to target next? Why?