AI & Digital Ethics Grade 12: AI in Science Drug Discovery Climate Modelling

Grade 12 – AI & Digital Ethics Topic: AI in Science: Drug Discovery, Climate Modeling

1. The Driving Question

How can a computer that doesn’t understand chemistry or physics invent a new medicine or predict a hurricane’s path years in advance—and why do we still need human scientists if it can? If AI can simulate millions of drug molecules overnight, does that mean the next penicillin will come from a server farm instead of a petri dish? And when the model gets it wrong, who’s responsible: the programmer, the scientist, or the algorithm itself?

2. The Core Idea – Built, Not Listed

Imagine a pharmaceutical lab in Cambridge, Massachusetts, where a robot arm pipettes tiny drops of liquid into 384-well plates while a nearby supercomputer runs molecular docking simulations—like a video game where drug molecules are Tetris blocks trying to fit into a protein’s "active site." The protein is SARS-CoV-2’s spike protein, and the computer isn’t just testing one molecule at a time; it’s screening 100 million virtual compounds in a weekend, something that would take human chemists decades. This isn’t sci-fi—it’s how Moderna’s COVID-19 vaccine was optimized in record time.

Now shift to a climate research center in Boulder, Colorado. A neural network trained on 40 years of satellite data and ocean temperature grids is predicting the 2030 Atlantic hurricane season. It doesn’t "know" meteorology, but it’s spotted patterns in jet stream shifts and sea surface temperatures that human forecasters missed. When it warns of a Category 5 storm hitting Miami in August 2030, policymakers have to decide: evacuate now, or wait for human models to confirm?

In both cases, AI isn’t replacing scientists—it’s amplifying them. It handles the brute-force work (testing millions of molecules, crunching terabytes of climate data) while humans ask the why questions: Why did this molecule bind? Why did the model predict a storm here? The AI provides hypotheses; humans test them in the real world.

Key Vocabulary: - Generative AI (in science): Definition: AI models that create new data (molecules, climate scenarios) by learning patterns from existing datasets, rather than just classifying or predicting. Example: Google DeepMind’s AlphaFold generates 3D protein structures from amino acid sequences—like predicting how a folded origami crane would look from its crease pattern. College shift: In grad school, you’ll debate whether generative models are discovering new science or just interpolating between known examples (the "black swan" problem).

• High-Throughput Virtual Screening (HTVS): Definition: Using AI to rapidly test millions of virtual compounds for drug potential by simulating their interactions with target proteins. Example: The AI Atomwise screened 10 million compounds in 2020 to find potential treatments for Ebola—something that would take a lab team 100 years. College shift: In computational chemistry, you’ll learn how HTVS trades speed for accuracy (false positives are common, so wet-lab validation is still critical).

• Explainable AI (XAI): Definition: Techniques to make AI’s decision-making transparent, so scientists can audit why a model predicted a molecule would work (or a storm would hit). Example: Climate models now use attention maps to show which data points (e.g., Gulf Stream temperatures) most influenced a hurricane prediction—like a weather forecaster showing their work. College shift: In AI ethics, you’ll study how XAI clashes with proprietary models (e.g., if a pharma company’s AI finds a drug, do they have to reveal how it worked to get FDA approval?).

• Bias-Variance Tradeoff (in scientific AI): Definition: The tension between a model that’s too simple (missing real patterns) and too complex (fitting noise in the data). Example: A climate model that’s too simple might ignore El Niño effects; one that’s too complex might predict a storm based on a single anomalous buoy reading. College shift: In machine learning, you’ll learn how this tradeoff applies to overfitting—where a model memorizes training data but fails in the real world.

3. Assessment Translation

AP Computer Science Principles / SAT Subject Test (Math Level 2) / College Admissions Essays: This topic appears in three key formats:
1. Free-response questions (AP CSP): Analyze a scenario where AI accelerates scientific discovery (e.g., "Explain how generative AI could design a new antibiotic, and identify one ethical concern"). - Proficient response: Names a specific AI tool (e.g., AlphaFold), describes its method (e.g., "predicts protein folding via neural networks"), and links to a real-world outcome (e.g., "helped design COVID-19 antivirals"). - Developing response: Vague ("AI helps scientists") or conflates AI with automation (e.g., "robots mix chemicals faster").

1. SAT Math (Data Analysis): Graph interpretation where AI-generated climate models are compared to historical data. Example: "A climate model predicts a 2°C temperature rise by 2050 with 90% confidence. Which of the following would most weaken this prediction?"
2. Distractors: (A) "The model was trained on 50 years of data" (irrelevant—more data could help or hurt); (B) "The model’s error bars overlap with another model’s" (doesn’t weaken, just shows agreement).

3. Correct answer: "The model assumes linear warming but recent data shows accelerating trends" (violates the model’s core assumption).

4. College admissions essays (e.g., MIT, Stanford): Prompts like "Describe a time you used technology to solve a real-world problem." A strong response might describe:

5. Using Google’s TensorFlow to analyze local air quality data and propose policy changes.
6. Critiquing a news article about AI drug discovery (e.g., "The article claimed AI ‘invented’ a drug, but it was actually a repurposed molecule—here’s why that matters").

Model Proficient Response (AP CSP Free-Response): Prompt: "Explain how AI is used in drug discovery, and identify one limitation of this approach." Response: "AI accelerates drug discovery through high-throughput virtual screening (HTVS). For example, the AI Atomwise simulates how millions of virtual molecules bind to a target protein (like SARS-CoV-2’s spike protein) by calculating their 3D shapes and chemical interactions. This narrows down candidates from billions to a few hundred, which human chemists then test in labs. However, a key limitation is false positives—the AI might predict a molecule will bind strongly based on simulations, but in reality, the molecule could be toxic or unstable in the body. This is why wet-lab validation is still essential. Additionally, AI models can inherit biases from their training data; if the dataset lacks diverse chemical structures, the AI might miss promising drug classes."

4. Mistake Taxonomy

Mistake 1: Overstating AI’s Autonomy Prompt: "How has AI transformed the process of drug discovery?" Common wrong response: "AI invents new drugs by itself, replacing human scientists." Why it loses credit: - Misrepresents the process: AI generates hypotheses, not final drugs. The response ignores the role of human validation (e.g., lab testing, clinical trials). - Lacks specificity: Doesn’t name a tool (e.g., AlphaFold, Atomwise) or describe how it works. Correct approach:
1. Name a specific AI tool and its function (e.g., "AlphaFold predicts protein structures").
2. Explain the human role (e.g., "Scientists use these predictions to design drugs, then test them in labs").
3. Note a limitation (e.g., "AI can’t yet predict side effects or toxicity").

Mistake 2: Ignoring Data Bias in Climate Models Prompt: "A climate model predicts a 30% increase in hurricanes by 2040. What’s one reason this prediction might be unreliable?" Common wrong response: "The model is wrong because AI isn’t perfect." Why it loses credit: - Vague: Doesn’t specify a source of bias (e.g., training data, assumptions). - Misses the assessment format: AP/SAT questions expect you to critique how the model works, not just say "it’s flawed." Correct approach:
1. Identify a specific bias (e.g., "The model was trained on satellite data from 1980–2020, which doesn’t account for recent Arctic warming trends").
2. Explain the impact (e.g., "This could underestimate hurricane intensity, since warmer oceans fuel stronger storms").
3. Suggest a fix (e.g., "Incorporating real-time buoy data could improve accuracy").

Mistake 3: Confusing Correlation with Causation in AI Outputs Prompt: "An AI model finds that regions with more wind turbines have lower asthma rates. Should policymakers install more turbines to reduce asthma? Why or why not?" Common wrong response: "Yes, because the AI found a pattern, so turbines must cause lower asthma." Why it loses credit: - Logical fallacy: Assumes correlation = causation without considering confounding variables (e.g., wealthier areas have both more turbines and better healthcare). - No evidence: Doesn’t propose a way to test the hypothesis (e.g., controlled studies). Correct approach:
1. Acknowledge the pattern (e.g., "The AI identified a correlation between turbine density and asthma rates").
2. Propose alternative explanations (e.g., "Wealthier areas may have both more turbines and better air quality due to other policies").
3. Suggest next steps (e.g., "A randomized study could test whether installing turbines in low-income areas reduces asthma").

5. Connection Layer

1. Within subject (AI & Digital Ethics): [AI in drug discovery]-[AI in criminal justice]

2. Why it matters: Both fields use AI to make high-stakes predictions (e.g., "this molecule will work" vs. "this defendant will reoffend"), but drug discovery has clear validation steps (lab tests), while criminal justice lacks transparent ways to audit bias. Understanding one exposes the risks of the other.

3. Across subjects (Science-Statistics): [Climate modeling]-[Bayesian inference]

4. Why it matters: Climate models update their predictions as new data comes in—just like Bayesian statistics updates probabilities with new evidence. The "confidence intervals" in climate reports (e.g., "90% chance of 2°C warming") are Bayesian in nature.

5. Outside school (Real world): [Generative AI in science]-[AI-generated art and music]

6. Why it matters: Tools like DALL·E and Stable Diffusion generate images by remixing training data—just like AlphaFold generates protein structures by remixing known folds. The same debates apply: Is this creation or remixing? Who owns the output? Next time you see an AI-generated album cover, you’ll recognize the same ethical questions as in drug discovery.

6. The Stretch Question

If an AI discovers a new antibiotic that cures a previously untreatable infection, but the model can’t explain how it works, should doctors prescribe it? Why or why not?

Pointer toward the answer: This isn’t just a science question—it’s a clash between utilitarian ethics (save lives now) and deontological ethics (follow rules like "do no harm"). The FDA currently requires drug mechanisms to be understood, but some argue this is outdated in the AI era. Consider: - Precedent: Many drugs (e.g., lithium for bipolar disorder) were used before their mechanisms were known. - Risk: If the AI’s "discovery" is a fluke (e.g., a false positive in virtual screening), the drug could fail in real patients—or worse, cause harm. - Transparency: If the AI is a "black box," how can doctors trust it won’t have long-term side effects? On the other hand, if we wait for full explainability, are we condemning patients to die while we debate?

The answer likely lies in hybrid validation: use the AI to generate candidates, but require rigorous lab testing before approval—even if the mechanism remains unclear. But this raises another question: Who gets to decide when the evidence is "good enough"?