⚔️ Discovery Arena: How to Play

✅ Chemical validity Is this even a real molecule?

Before any score means anything, the structure has to be chemically possible. This is the first thing checked and the cheapest to check.

• Does the SMILES string parse?
• Are the bonds possible?
• Are the atom valences correct?
• Is the molecule a duplicate?
• Is stereochemistry defined?
• Is the charge reasonable?

🚦 Hard gate. An invalid molecule is rejected immediately, no matter how attractive its other numbers look.

🧲 Binding affinity How strongly does it stick to the target?

Affinity is reported as Kd, Ki, or IC50 — or their log forms pKd, pKi, pIC50. A lower Kd usually means stronger binding, and because pKd = −log10(Kd), a higher pKd means stronger binding.

Kd, Ki, and IC50 are not interchangeable — a model must state which value it predicts. Three related but separate ideas sit next to affinity: docking (how the molecule physically fits the 3D binding site), potency (how much is needed to produce an effect), and efficacy (how large an effect it can produce). A molecule can bind tightly and still produce only a small biological effect.

⚠️ A docking score is not a measured binding affinity, and binding is one score — never the verdict.

🎯 Selectivity Does it hit only the target it is meant to?

A molecule that also binds many unrelated proteins produces off-target interactions, and those are a common source of side effects.

🎯 The goal is not maximum binding. It is strong, useful binding to the correct target, with limited harmful binding everywhere else.

⚖️ Drug-likeness Does it resemble medicines that have worked before?

QED (Quantitative Estimate of Drug-likeness) gives a score between 0 and 1, combining molecular size, lipophilicity, polarity, flexibility, hydrogen bonding, and structural complexity.

Lipinski’s Rule of Five examines oral candidates using molecular weight, lipophilicity, hydrogen-bond donors, and hydrogen-bond acceptors.

📏 These are guidelines, not laws — some successful drugs break them. A high QED only suggests the molecule looks conventional; it never proves it is effective or safe.

💧 ADME What does the body do to the drug?

Add toxicity and the term becomes ADMET. Context decides what counts as good: crossing the blood–brain barrier is desirable for a brain drug and a liability for one meant to stay out of the brain. A short half-life may force frequent doses; a very long one may cause accumulation.

💧 A molecule cannot work if it never arrives in a useful amount — or if it is destroyed before it gets there.

🛡️ Toxicity Could it harm the patient?

Toxicity is not one property. Predictions are made separately for many kinds of harm:

• Hepatotoxicity — liver damage
• Nephrotoxicity — kidney damage
• Cardiotoxicity — heart damage
• Neurotoxicity — nervous-system damage
• Genotoxicity / mutagenicity — damage to genetic material
• Carcinogenicity — possible contribution to cancer
• Immunotoxicity — harm to the immune system
• Reproductive toxicity — harm to fertility or development

hERG risk: the hERG ion channel helps control the heart’s electrical rhythm, and blocking it can cause dangerous rhythm problems.

🚦 Hard gate. A candidate with high predicted hERG risk may be rejected even when it binds beautifully to the intended target.

🏭 Synthesis Can chemists actually make it — and make a lot of it?

Synthetic accessibility asks whether the molecule can realistically be produced at all, weighing complexity, number of rings, unusual structures, stereochemistry, available starting materials, and known reactions.

Retrosynthesis plans backwards from the target molecule to simpler starting materials, intermediates, reagents, and reaction steps. A route needing four reliable steps is far more practical than one needing twenty difficult ones.

Manufacturability goes further: can it be made reliably, repeatedly, at large scale, at acceptable cost and purity, without extreme hazards? Yield, raw-material availability, purification, batch consistency, and environmental impact all count.

🏭 Making one milligram in a laboratory is a very different problem from manufacturing kilograms for patients.

💊 Developability Can it become a medicine someone can actually take?

Chemical stability, shelf life, formulation, dose size, route of administration, crystal formation, solubility at useful concentrations, and storage requirements all decide whether a promising compound survives into a real product.

Therapeutic window: the range between a dose that produces a useful effect and a dose that causes unacceptable toxicity.

📈 A wider therapeutic window is usually safer — there is more room between helping and harming.

🌱 Novelty & diversity Is it actually new?

AI models happily generate the same idea over and over, so candidates are checked for similarity to the reference molecule, similarity to known compounds, scaffold novelty, diversity across the candidate set, and outright duplicates.

🌱 A new SMILES string is not necessarily a new molecule — and chemical novelty does not automatically mean the molecule can be patented.