Computational Biology Paper Claim Patterns

Computational biology papers often look different on the surface but reuse a small set of claim patterns. Use this page before writing a paper note when the abstract mentions molecules, proteins, docking, conformers, complexes, or genome sequence modeling.

The generic pattern is:

$$\text{claim}=(\text{object},\text{task},\text{representation},\text{objective},\text{evidence})$$

Do not route a paper only by model name. A Transformer for protein sequences, a GNN for molecular graphs, and a diffusion model for ligand poses need different object and evaluation contracts.

Pattern Map

Pattern	Object	Typical Claim	Primary Route
Molecular property prediction	molecule or ligand	better scalar/class prediction	Molecular and Ligand Modeling, Property prediction
Target-conditioned activity	molecule, target, assay	better bioactivity or affinity prediction	Interaction modeling, Target-assay-label contract
Protein representation learning	protein sequence or structure	better transfer to downstream protein tasks	Sequence-Based Modeling, Self-supervised learning
Protein structure prediction	protein sequence, MSA, template, structure	better coordinates or contact geometry	Protein modeling concepts, Coordinate modeling contract
Docking or pose generation	protein-ligand complex	better pose, ranking, or enrichment	Docking, SBDD concepts
Structure-conditioned generation	pocket, ligand, complex	valid and useful molecules under context	Generative models, Structure-based modeling
Conformer generation	molecule, conformer ensemble	plausible low-energy 3D structures	Conformer, Geometry for Structure Modeling
Protein design	sequence, backbone, fold, binder, function	generated designs satisfy structural or functional constraints	Protein design, Protein modeling
Genome sequence modeling	sequence window, variant, region	better sequence or variant-level prediction	Sequence-Based Modeling, Genome modeling concepts

Property Prediction

The basic supervised form is:

$$\hat{y}=f_{\theta }(r(m)),\hat{\theta }=\arg \underset{\theta }{\min }\frac{1}{n}\sum _{i=1}^n\mathcal{L}\left(f_{\theta }(r(m_i)),y_i\right)$$

• $m_i$: molecule or ligand.
• $r(m_i)$: SMILES, graph, fingerprint, descriptor, conformer, or embedding.
• $y_i$: property label with unit, endpoint, and direction.
• $\mathcal{L}$: classification, regression, ranking, or censored-label objective.

Minimum evidence:

Required	Why
Chemical state policy	salt, tautomer, protonation, stereo, and conformer choices can change identity
Split unit	random split may overstate scaffold or series generalization
Baseline	fingerprint/tree or simple descriptor baselines catch weak representation claims
Metric alignment	RMSE, AUROC, PR-AUC, enrichment, and calibration answer different questions

Target-Conditioned Activity

Bioactivity rows usually depend on target and assay context:

$$\hat{y}=f_{\theta }(r(m),r(t),r(a))$$

• $m$: molecule or ligand.
• $t$: target, protein, pocket, family, or construct.
• $a$: assay, endpoint, condition, source, threshold, or unit.
• $y$: measured activity, affinity, binary hit, ranking label, or censored value.

Do not collapse this into only molecule -> label unless the target and assay are truly fixed.

Minimum evidence:

Required	Why
Target-assay-label contract	same molecule can have different labels under different targets or assays
Split on both axes	ligand-only or protein-only split may not test pair generalization
Assay harmonization	pooled assay values can mix incompatible measurement processes
Applicability domain	prediction reliability can fail on new chemotype, target family, or assay source

Protein Representation Learning

Many protein papers optimize a pretraining objective and evaluate transfer:

$$z_i=g_{\theta }(s_i),{\hat{y}}_i=h_{\varphi }(z_i)$$

• $s_i$: protein sequence, MSA, residue graph, or structure-derived representation.
• $g_{\theta }$: pretrained encoder.
• $h_{\varphi }$: probe, fine-tuned head, retrieval scorer, or task-specific model.
• $y_i$: downstream label such as family, function, stability, contact, or binding.

Minimum evidence:

Required	Why
Representation unit	residue, sequence, domain, chain, complex, and pocket embeddings are not interchangeable
Adaptation budget	linear probe, kNN, frozen head, and full fine-tuning test different claims
Protein split	random sequence splits can leak homologous families
Downstream metric	representation quality is task- and evaluator-dependent

Docking, Pose, and Screening

Pose claims depend on coordinates and inference-time information:

$$S_{\theta }(p,l,X)\to \text{rank or pose decision}$$

• $p$: protein, receptor, or pocket context.
• $l$: ligand.
• $X$: generated or evaluated ligand coordinates in the pocket.
• $S_{\theta }$: score used for pose selection, affinity ranking, or virtual screening.

Minimum evidence:

Required	Why
Pocket definition	ligand-defined pockets can leak evaluation information
Coordinate source	experimental, predicted, apo, holo, and template-derived structures support different claims
Pose metric	RMSD, clash, interaction recovery, and plausibility measure different failures
Ranking metric	pose success does not imply affinity ranking or enrichment
Template and analog leakage	close structures or analogs can make the test task easier

Generative Molecules and Complexes

Generative claims should separate model objective from utility evaluation:

$$x\sim p_{\theta }(x\mid c),\text{utility}(x,c)\ne \log p_{\theta }(x\mid c)$$

• $x$: molecule, conformer, pose, sequence, backbone, or complex.
• $c$: condition such as target, pocket, property, scaffold, fold, or function.
• $p_{\theta }$: autoregressive, diffusion, flow, VAE, GAN, or search distribution.

Minimum evidence:

Required	Why
Validity rule	syntactic validity, chemical validity, structural plausibility, and biological utility differ
Novelty and diversity	high average score can hide near-duplicates or mode collapse
Conditional satisfaction	generated samples must satisfy the stated condition, not only look plausible
Filtering policy	post-generation filters can own the final performance claim
Utility metric	docking score, affinity proxy, property predictor, or experimental assay has a scope boundary

Protein Design

Protein design may generate sequence, structure, or both:

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

• $s$: amino-acid sequence.
• $X$: backbone, side-chain, complex, or generated coordinate structure.
• $c$: fold, motif, target, binding site, function, or stability condition.

Minimum evidence:

Required	Why
Design target	fold, binder, enzyme, scaffold, or sequence recovery are different claims
Structure validation	predicted foldability is not the same as experimental function
Novelty check	training-template or family similarity can inflate design success
Experimental boundary	in-silico scores should not be written as wet-lab validation

Genome Sequence Modeling

Genome notes in this site are limited to sequence, region, k-mer, and variant-level modeling:

$$\hat{y}=f_{\theta }(r(g_{a:b}),c)$$

• $g_{a:b}$: genomic sequence window or region.
• $r(\cdot )$: tokens, k-mers, embedding, annotation features, or variant context.
• $c$: organism, coordinate system, strand, annotation source, or task condition.
• $y$: sequence label, region label, or variant-effect target.

Minimum evidence:

Required	Why
Coordinate system	region definitions depend on reference build and indexing convention
Split unit	random windows can leak nearby sequence or homologous regions
Annotation source	labels can reflect database and pipeline artifacts
Scope boundary	this site does not expand into broad omics or clinical interpretation by default

Final Routing Checklist

• Does the paper’s strongest claim belong to object/domain, AI method, Math objective, benchmark, systems, or agent workflow?
• Is the biological object explicit before the model name?
• Is the raw object separated from model-ready representation?
• Is the loss written with sampled unit and distribution?
• Is the evaluation metric tied to split, baseline, uncertainty, and failure mode?
• Are generated or predicted structures checked under the correct coordinate and symmetry contract?
• Are public artifacts and missing metadata marked honestly?

Related

• Computational Biology paper intake
• AI Computational Biology Math contract
• Data and evaluation
• AI paper intake
• Formula intake
• Claim routing
• AI Computational Biology Math paper template
• Coverage matrix