Gains:
- Ability to evaluate AlphaFold structure predictions with the pLDDT confidence score and avoid claims based on low confidence regions
- Ability to interpret docking score as relative ranking, not absolute affinity, and verify candidates with experimental affinity measurement
- Ability to filter productive molecule suggestions through synthesizability, stability, and biosafety and link critical predictions to wet testing
To understand what a protein does, it is necessary to know its three-dimensional shape; because structure determines function — the catalytic pocket of an enzyme, the binding surface of a receptor always arises from its three-dimensional shape. For decades, unraveling the structure of a protein took months of experiments (X-ray crystallography, cryo-EM). AlphaFold and similar AI models have changed this picture: they can predict the three-dimensional structure of a protein from a sequence of amino acids with high accuracy in minutes. This is revolutionary for bioengineering; But knowing the limits of the revolution is the condition for using it safely.
In this unit, you will learn how to use AI in structure prediction, molecular docking (calculating how and how strongly a small molecule binds to a protein) and molecule design workflows; At each step you will learn where the prediction ends and the experiment begins.
AlphaFold: power and limit
AlphaFold (a deep learning model that predicts 3D structure from a protein sequence) produces a confidence score (pLDDT — a value from 0 to 100 indicating how confident the model is for each amino acid) with each prediction. High pLDDT (>90) is generally reliable; Low regions (<50) are often disordered regions—flexible portions of protein that do not have a fixed structure—or areas where the model is uncertain. The limits AlphaFold has to know are that it gives a single static structure (not the movement of proteins), it does not model ligand (small molecule binding to a protein) or metal ion effects, it does not always capture the subtle effect of mutations, and it does not guarantee whether a complex will form.
Caution: AlphaFold is a predicted hypothesis, not an experimental construct. Claiming a mechanism based on regions of low pLDDT or identifying a drug pocket in these regions is misleading. Critical structures must be confirmed by experimental method (crystallography, cryo-EM).
Docking: predicting docking
The fundamental question in drug and enzyme engineering is: "Does this molecule bind to this protein, and how strongly?" Docking calculates this and ranks it with a scoring function — an approximate mathematical model that estimates binding strength. But docking scores are relative and do not predict true binding affinity; The false positive rate is high. New AI-based methods make docking faster and more accurate, but the result is still a scanning sequence — verified by experimental affinity measurement (such as SPR, ITC).
Tip: Use docking to sequence thousands of molecules before experimentation, not for an absolute "bind/no-bind" decision. Synthesizing and testing the 50 highest-scoring molecules is much more efficient than testing the 50 at random; But experience decides.
Molecular design and generative models
Generative models—AI that can “design” new protein sequences or small molecules—can suggest candidate molecules with desired properties: a more stable enzyme, a new drug scaffold that binds to a target. This is powerful, but it has two pitfalls: the model may suggest unsynthesizable (chemically impossible to produce) molecules or unrealistic sequences. That's why productive outputs are always filtered through synthesizability, stability, and security.
three mini cases
Case 1 — Structure guided the estimate. An enzyme engineering team was working without an experimental structure of the target enzyme. AlphaFold pointed to the predicted active site—the area where the enzyme carries out the chemical reaction; The team designed 3 mutations in this region. In the experiment, one increased activity by 2.3 times. But two mutations didn't work — prediction led, experiment decided.
Case 2 — Low trust zone trap. A student wanted to attach medication to a "pocket" in the AlphaFold structure. Looking at the pLDDT map, he saw that the area was a low-safe, possibly disorderly area with a score of 38. Moved docking to safe zone; wasting months on a wrong target was prevented.
Case 3 — Unsynthesizable proposal. A generative model suggested 200 molecules that bind to a target with a high score. When the chemist filtered it, he found that 40% of it could not be practically synthesized or would be unstable. Once the remaining candidate pool became realistic, 12 were synthesized.
Four copyable templates
1) AlphaFold output comment:
Your role: structural biologist. I'll give you the pLDDTscore summary of an AlphaFold prediction. Distinguish which regions are reliable (>90) and which are suspicious (<50). Do not establish any mechanism claims based on low confidence zones. Write in your printout that this is not an experimental build and how to verify it.
2) Docking workflow plan:
Your role: computational chemist. Describe the small molecule docking workflow for [target protein]: protein preparation, binding pocket definition, ligand preparation, scoring, sequencing. Explain that docking score is NOT absolute affinity and by what experimental method (SPR/ITC) it will be verified.
3) Mutation design justification:
I would like suggestions for mutations to increase the stability of an enzyme. For each recommendation: target position, rationale (structural logic) and possible risk (loss of activity). Point out that these are guesses and each must be tested by experiment. Give up to 5 suggestions and prioritize.
4) Productive output filtering:
Give me the criteria for evaluating molecules suggested by a generative model: synthesizability, chemical stability, toxicity signatures (structural caveats), molecular weight, and solubility. Tell me the threshold at which I "should eliminate" for each criterion. This is a pre-filter; The final decision belongs to the wet laboratory.
Weak prompt / Strong prompt
Weak prompt:
Design a drug that binds to this protein.
No security, no synthesizability, no verification; AI produces unrealistic, unverifiable output.
Powerful prompt:
Your role: computational chemist. Propose a SCREENING strategy (docking-based) for small molecule scaffolds that have the potential to bind to the [target] protein. Don't produce concrete molecules; instead describe step by step which library, which pocket, which scoring and which filters to use, and how to validate each candidate with experimental affinity.
The difference: strategy focus, realistic scope, explicit validation and security awareness.
Structured workflow and validation
step
AI contribution
border
verification
Structure prediction
AlphaFold 3D model
Static, without ligand
Cryo-EM/crystallography
Trust assessment
pLDDT comment
Low zone unclear
Independent model comparison
docking
quick sort
Score ≠ affinity
SPR/ITC measurement
mutation design
Candidate recommendation
Impact estimate
activity experiment
molecule production
new candidate
Synthesis/stability
Chemist + wet test
Molecular dynamics: activating structure
AlphaFold gives you a single, frozen photo; However, proteins are machines that are constantly moving, vibrating, their pockets opening and closing. Molecular dynamics simulation (MD—a computer calculation of the movement of a molecule's atoms over time according to Newton's laws of physics) makes this movement visible: whether a drug pocket is actually open, how a mutation changes flexibility, whether a binding remains stable. AI helps in planning an MD setup (force field selection, water box, equilibrium steps) and summarizing the results, but the physical validity of the simulation (enough time, correct force field) is expert judgment. A short or improperly constructed simulation produces a result that appears real but is physically meaningless.
Tip: Before trusting the result of an MD simulation, ask two questions: is the simulation long enough (has the system reached equilibrium) and does it give the same behavior when repeated (with a different starting seed)? A single, brief simulation shows a trend, not evidence.
Common mistakes
- Assuming the structure is certain without looking at pLDDT. Claims based on low safety zones are misleading.
- Mistaking the docking score as an affinity measurement. Score is relative; It requires experimentation.
- Mistaking a static structure for a dynamic reality. Proteins move; One structure does not tell the whole story.
- Accepting suggestions for non-synthesizable molecules. Productive output must pass through the chemical filter.
- Security/dual-use blindness. Requests to design toxins or harmful agents should be rejected.
In summary
AlphaFold and docking are powerful AI tools that accelerate protein engineering and drug discovery; Structure prediction leads the way, docking sorts thousands of molecules, generative models suggest new candidates. But the output of all is hypothesis: read pLDDT confidence scores, don't mistake docking score for affinity, don't confuse static structure with dynamic reality, and verify each critical candidate experimentally. Synthesizability, safety and dual-use awareness are indispensable in molecule design.
Application task
Find a structure prediction for a publicly available protein sequence from the AlphaFold database (or an online server). Examine the pLDDT map and separate safe and suspicious regions. Then have the AI comment on the active site of this protein and compare each claim it makes with the pLDDT map: is the claim based on a low-confidence region? Evaluate the result in one paragraph.
checklist
- [ ] I evaluated the structure prediction together with the pLDDT confidence score.
- [ ] I have avoided claims based on low confidence regions.
- [ ] I interpreted the docking score as relative ranking, not absolute affinity.
- [ ] I observed that the static structure does not represent dynamic behavior.
- [ ] I filtered productive molecule suggestions through synthesizability and safety.
- [ ] I linked each critical prediction to an experimental verification plan.