Module 11: De Novo & Generative Molecular Design
Every method so far has been a way of searching for molecules that already exist. This module inverts the question: instead of asking which compound to screen next, we ask which compound should we make next — and let the computer design it.
1. From Searching to Designing
Virtual screening (Module 8) is fundamentally a lookup: you enumerate a catalogue and rank it. That caps you at what vendors happen to stock — perhaps 10¹⁰ molecules in the largest make-on-demand libraries. But drug-like chemical space is estimated at around 10⁶⁰ molecules. Screening explores a rounding error of the possible.
Generative design constructs molecules from scratch, which reframes drug discovery from "what do we have?" to "what should we make?" — and slots directly into the Design–Make–Test–Analyse (DMTA) cycle that governs real medicinal chemistry projects.
The 10⁶⁰ figure is seductive and slightly misleading. Chemical space is discontinuous: a great many structures you can draw on a computer cannot be synthesised, and some cannot physically exist at all. A generative model that ignores this will happily hand you a beautiful, high-scoring, entirely unmakeable molecule. Everything in this module is, in one way or another, a defence against that failure.
2. The Three Pillars
Strip away the branding and every generative design system — from a 1990s rule-based grower to a 2026 diffusion model — is built from the same three components, resting on a cheminformatics foundation (Module 5):
Construction
How do we build a molecule?
The set of legal moves for assembling or modifying a structure — atom by atom, fragment by fragment, or reaction by reaction.
Scoring
How good is this molecule?
The function that collapses a chemical structure into a number to be optimized. Get this wrong and the search will optimize the wrong thing, perfectly.
Search
Where do we look next?
The strategy for traversing an astronomically large space you can never enumerate — balancing exploration against exploitation.
3. Construction: The Control ↔ Synthesizability Spectrum
The three construction strategies are not competing methods so much as points on a single trade-off. Every step you take toward control costs you synthesizability, and vice versa.
Atom-based
Maximum controlMinimum synthesizabilityAdd, delete, or substitute one atom or bond at a time on the molecular graph. The entire periodic table is available and you can place anything anywhere — which is exactly the problem. Nothing constrains the result to be makeable, and every edit must be chased with valence bookkeeping (add a methyl to benzene and you must also delete a hydrogen, or you have created pentavalent carbon).
Fragment-based
Moderate controlModerate synthesizabilityAssemble molecules from a library of building blocks — rings, linkers, and side-chain decorators — usually harvested by fragmenting known drugs. Three macro-strategies: growing (extend a fragment into unoccupied pocket space), linking (join two fragments bound in adjacent subpockets), and merging (fuse two fragments that share a common substructure). The building blocks are real, but nothing guarantees the assembled product is.
Reaction-based
Least controlMaximum synthesizabilityBuild molecules only by applying encoded reaction transformations (SMIRKS, Module 5) to real available building blocks — an amide coupling, a Suzuki, a reductive amination. You cannot place a halogen on a strained ring just because you want to, because no transformation will do it. In exchange, the output arrives with a synthetic route attached.
4. Scoring: Explicit, Implicit, and the Trap of Optimizing One Number
Scoring is the hardest pillar, because it is where the chemistry and biology of your project must be compressed into a number that a search algorithm will relentlessly maximize. Scores split into two kinds:
Explicit scores
Anything you can compute: a docking score (Module 6), an MM/GBSA or FEP free energy (Module 10), a QSAR prediction (Module 9), logP, TPSA, a QED drug-likeness score. Objective, reproducible, and cheap to state precisely.
Receptor-based (needs the target structure) vs ligand-based (descriptors and pharmacophores only).
Implicit scores
The medicinal chemist's judgement — "I don't like that group", "we'll never make that", "that will be a metabolic liability". Subjective, hard to encode, and routinely the thing that decides whether a designed molecule is ever synthesised.
Two molecules can share an identical calculated logP of 2.64 and be worlds apart in whether a chemist would touch them.
If you optimize a single objective, you will get it — and nothing else. Optimize docking score alone and the model will hand you enormous greasy molecules that score beautifully and dissolve in nothing, because more atoms means more favourable contacts. The score was not wrong; it was incomplete. This is why real design is always multi-objective.
Interactive Playground: The Pareto Frontier
A batch of 16 designed molecules, scored on two objectives that genuinely fight each other: potency and synthesizability. Slide the weight to change what your project cares about, and watch which molecule wins. The lesson is in what never wins.
C
5 of 16 molecules are non-dominated. Sweep the weight from 0 to 1 and you will find the winner is always one of them.
The other 11 are dominated — some other molecule is at least as potent and at least as makeable. No weighting will ever select them. Synthesising one is strictly wasted effort.
- Multi-objective optimization does not return an answer — it returns a menu. The frontier is the set of rational choices; picking a point on it is a project decision (how much potency will you trade for a route your chemists can actually run?), not a computation.
- Dominated molecules are the real output of a bad scoring function. They are not merely suboptimal, they are never optimal for any preference whatsoever.
- Collapsing objectives into one weighted number hides the trade-off. Fix the weight at 0.5 and you would only ever see molecule C — and never learn that a slightly different preference buys you a very different molecule.
5. Search: Exploration vs. Exploitation
Search is what makes the space tractable. Consider reaction-based design with ~4,500 available transformations: a mere four-step synthesis already implies on the order of 4,500⁴ ≈ 4 × 10¹⁴ possible trajectories. Exhaustive enumeration is not slow — it is impossible. Every search algorithm is therefore a policy for sampling, and each one answers the same question differently:
Do I exploit — refine the best thing I have found so far — or do I explore — try something unfamiliar that might be better, or might waste the step?
| Strategy | How it decides | Character |
|---|---|---|
| Systematic enumeration | Try every legal move in order | Fully deterministic; only viable for tiny spaces |
| Monte Carlo | Pick moves at random; accept good ones, accept bad ones occasionally (Module 6) | Stochastic; escapes local optima by design |
| Genetic algorithm | Keep a population; crossover and mutate the fittest molecules | Stochastic; good at recombining partial solutions |
| ε-greedy | Take the best-known move with probability 1−ε, a random one with probability ε | The simplest explicit exploration/exploitation dial |
| Monte Carlo Tree Search | Build a tree of moves, expanding branches that look promising and under-explored | Balances the trade-off adaptively; strong for multi-step synthesis |
| Reinforcement learning | Train a generator so that high scores become more probable outputs | The scoring function reshapes the generator itself |
Push exploitation too hard — especially with reinforcement learning — and the generator discovers one high-scoring scaffold and produces it forever, in thousands of trivial variations. The score goes up; the chemistry goes nowhere. Diversity is not a nice-to-have in generative design, it is a constraint you must actively defend.
6. 3D Pocket-Conditioned Diffusion
The most powerful structure-based generators do not build a SMILES string and dock it afterwards — they generate directly inside the pocket in 3D. Models such as Pocket2Mol and DiffSBDD treat the protein pocket coordinates as boundary conditions and iteratively denoise a random cloud of atoms into a molecule.
Continuous denoising diffusion
Start from a random swarm of atoms in the pocket. Over many denoising steps the model refines their Cartesian coordinates and atom identities (C, N, O…) toward a chemically sensible, electrostatically complementary structure — then infers the bonds.
Why 3D conditioning helps
Designing in the pocket means hydrogen bonds, hydrophobic surfaces, and steric walls are satisfied during generation rather than checked afterwards. It fuses the construction and scoring pillars into a single learned step.