Conditional Generation

Conditional generation models a distribution over outputs $x$ given a condition $c$.

The goal is:

$$x\sim p_{\theta }(x\mid c)$$

$c$ can be text, class label, image, sequence, graph, structure, property target, scaffold, pocket, or another modality.

Conditional Objective

Most conditional generative objectives optimize a conditional distribution or conditional denoising target:

$$\mathcal{J}(\theta )={\mathbb{E}}_{(x,c)\sim p_{\mathrm{data}}}\left[-\log p_{\theta }(x\mid c)\right]$$

or, for diffusion/flow-style models:

$$\mathbb{E}\left[{\ell }_{\theta }(x_t,t,c,x)\right].$$

The condition $c$ is part of the data contract. A paper should state whether $c$ is observed at inference, predicted by another model, retrieved, manually specified, or derived from the target output.

Conditioning Mechanisms

• Concatenate condition tokens or features to the input.
• Cross-attend from generated tokens to condition representations.
• Condition a score, noise, velocity, or decoder network.
• Use classifier or classifier-free guidance during sampling.
• Constrain decoding or generation with validity rules.

Condition Boundary

Condition Type	Example	Leakage Risk
label or class	molecule property bin, protein family	label may be unavailable or noisy at deployment
text or prompt	desired function, natural-language instruction	prompt may encode benchmark answer
structure or pocket	binding pocket, partial coordinates	pocket may be ligand-defined or template-derived
scaffold or motif	fixed substructure, sequence motif	constraint can make novelty claims narrower
retrieval context	nearest examples, database hits	retrieval corpus may include test or homologous examples
predicted condition	upstream model output	error propagation must be measured

Evaluation Axes

Conditional generation should be evaluated along several axes:

$$\text{quality}\ne \text{condition satisfaction}\ne \text{diversity}\ne \text{novelty}$$

Axis	Question
validity	is the sample syntactically, chemically, physically, or structurally valid?
fidelity	does it satisfy $c$?
diversity	are samples varied under the same condition?
novelty	are samples not memorized or near-duplicates?
utility	does the sample help the downstream task?
cost	how much sampling, guidance, filtering, or reranking was used?

Molecular Modeling Examples

• Generate a molecule conditioned on target property.
• Generate a ligand conditioned on a pocket.
• Generate protein sequence conditioned on fold or function.
• Complete a structure conditioned on partial coordinates.

Checks

• What is the condition $c$ and where does it enter the model?
• Is the condition available at deployment time?
• Does stronger conditioning reduce diversity?
• Is controllability evaluated separately from sample quality?
• Could condition labels leak target information across splits?
• Is $c$ raw, predicted, retrieved, or derived from the target output?
• Are filtered invalid samples counted in the evaluation?

Related

• Guidance
• Sampling
• Diffusion model
• Flow matching
• Multimodal learning
• Protein-ligand interaction
• Generation evaluation