Protein Design

Protein design generates sequences and/or structures intended to fold, bind, or function as specified, including de novo backbones and sequence design for a target structure.

One conditional view is:

$$s\sim p_{\theta }(s\mid X,c)$$

where $s$ is a sequence, $X$ is a backbone or structural context, and $c$ is a design condition such as binding, fold, or function.

Protein design can also generate coordinates:

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

or joint sequence-structure objects:

$$(s,X)\sim p_{\theta }(s,X\mid c)$$

The claim changes depending on whether the model generates sequence, backbone, side chains, complex pose, or only ranks candidates.

Design Targets

Target	Generated Object	Evidence Boundary
Sequence design	sequence for a fixed backbone or fold	sequence recovery and predicted structure are not enough for function
Backbone design	coordinates or fold scaffold	geometry plausibility does not prove foldability or expression
Binder design	sequence, backbone, or complex against a target	docking or predicted complex score is not experimental binding
Functional design	sequence or structure with desired activity	proxy model score is not functional validation
Motif scaffolding	scaffold around constrained residues or motif	motif placement must be separated from global stability

Objective Patterns

Autoregressive sequence design can be written as:

$$p_{\theta }(s\mid X,c)=\prod _{i=1}^L{p}_{\theta }(s_i\mid s_{<i},X,c)$$

Diffusion or flow-style structure design often models a path over coordinates:

$$X_t\to v_{\theta }(X_t,t,c)\to X_0$$

where $X_t$ is a noisy or intermediate structure, $v_{\theta }$ is a denoising or velocity field, and $c$ is the design condition.

Many workflows add filtering or guidance:

$$\tilde{x}\sim p_{\theta }(x\mid c),x^{\star }=\arg \underset{\tilde{x}\in \mathcal{C}}{\max }{r}_{\psi }(\tilde{x},c)$$

where $r_{\psi }$ can be a folding model, structure predictor, docking score, energy model, classifier, or human-defined rule. The reported claim must state whether quality comes from the generator, the filter, or both.

Claim Ladder

Claim	Evidence needed
sequence is plausible	composition, language-model score, diversity, training similarity
structure is plausible	backbone geometry, clashes, local quality, fold confidence
sequence folds to intended structure	independent structure prediction or experimental structure
binder is plausible	interface geometry, negative controls, target specificity checks
function is plausible	assay-relevant model or experiment, not only fold confidence
experimentally validated	wet-lab evidence with conditions and failure denominator

Evaluation Contract

Check	Why
Training similarity	near-template or homologous designs can look novel only superficially
Structural validity	backbone geometry, clashes, side-chain packing, and fold confidence test different failures
Conditional satisfaction	designs must satisfy the requested fold, motif, target, or function
Diversity	many high-scoring designs may be near-duplicates
Experimental status	in-silico validation must not be written as wet-lab success
Negative controls	generated designs should be compared with simple baselines or shuffled conditions
Candidate accounting	attempted, filtered, failed, and retained designs must be counted separately
Split boundary	training similarity, homologs, templates, and target-family leakage must be checked

Why It Matters

• Enables novel binders, enzymes, and scaffolds beyond natural proteins.
• Combines structure generation with sequence design and validation.
• Computational success must hold up against experimental reality.

Checks

• Are designs evaluated for foldability and structural consistency?
• Does the model exploit training data instead of generalizing?
• How are designs validated beyond in-silico metrics?
• Is the claim about sequence recovery, structure plausibility, binding, function, or experimental validation?
• Are generated samples filtered by a separate model that owns part of the final claim?
• Are failed or filtered designs included in the denominator?
• Is the validation model independent from the generator’s training signal?
• Does the design claim depend on sequence, structure, complex, or experimental evidence?

Related

• Molecular generation
• Diffusion model
• Flow matching
• Protein modeling
• Computational Biology paper claim patterns
• AI Computational Biology Math contract