AI Computational Biology Math Contract

Use this contract when a paper, post, or project combines an AI method, a computational biology object, and a mathematical objective or metric.

The minimum useful record is:

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

If one part is missing, the note may still be useful, but the missing part should be marked to verify rather than inferred.

Axis Roles

Axis	Owns	Does Not Own
Computational Biology	molecule, ligand, protein, pocket, complex, conformer, genome region, assay, split unit	generic architecture definitions
AI	architecture, learning method, generative model, training protocol, inference behavior	biological label semantics by itself
Math	objects, indices, distributions, objectives, metrics, uncertainty, symmetry	paper-specific biological interpretation by itself
Evaluation	split, metric, baseline, selection rule, leakage, uncertainty, allowed information	headline score without protocol
Systems	compute, memory, throughput, artifacts, reproducibility, serving boundary	model quality claim without task evidence

Minimum Fields

Field	Required Question	Route
Object	What biological or chemical unit is modeled?	Computational Biology, Entities
Context	Does the claim depend on target, pocket, assay, receptor state, species, construct, or template?	Computational Biology Boundary
Representation	How does the raw object become tokens, graph, coordinates, embedding, or features?	Representation contract
Architecture	What inductive bias, parameter sharing, and complexity are used?	Architectures
Learning signal	What supervision, mask, contrast, denoising target, preference, or reward is used?	Learning methods
Objective	What loss, likelihood, score, velocity, estimator, or constraint is optimized?	Formula intake
Metric	What score is reported, and does it match the objective?	Objective-metric alignment
Split	What unit is held out: scaffold, protein family, complex pair, assay, source, time, or template group?	Data and evaluation
Evidence	Which benchmark, baseline, ablation, uncertainty, and failure mode support the claim?	Benchmark claim contract
Public boundary	Can the note be published without private data, internal paths, unpublished results, or collaborator details?	Publishing gate

Formula Pattern

Most mixed AI and computational biology papers can be rewritten as:

$$\hat{\theta }=\arg \underset{\theta }{\min }{\mathbb{E}}_{u\sim q(u)}\left[w(u){\mathcal{L}}_{\theta }\left(r(u),y(u),c(u)\right)\right]$$

• $u$: sampled unit, such as molecule, protein, complex, conformer, residue, atom, assay row, or sequence window.
• $q(u)$: training sampling distribution.
• $w(u)$: mask, class weight, importance weight, time weight, or balancing rule.
• $r(u)$: representation passed to the model.
• $y(u)$: label, target, generated object, score, velocity, noise, reward, or preference.
• $c(u)$: context such as target, pocket, assay, condition, source, or time.
• $\theta$: optimized model parameters.

This rewrite forces the note to separate biological object, computational representation, learning objective, and evidence.

Routing Decisions

Main Claim	Primary Home	Required Companions
New model block improves performance	AI architectures	object contract, objective, benchmark boundary
New pretraining or SSL signal helps transfer	Learning methods	representation unit, transfer protocol, split rule
New generative method samples valid objects	Generative models	validity metric, sampling procedure, utility metric
New docking, conformer, or structure workflow	Computational Biology	preprocessing, coordinate contract, leakage check
New formula or estimator explains the method	Math	symbol table, sampled distribution, evaluation link
New benchmark or leaderboard result is central	Evaluation	dataset card, split, baseline, uncertainty
New implementation makes the method practical	Infra	artifact availability, reproducibility, scaling boundary

Cross-Axis Failure Modes

Failure Mode	Symptom	Fix
model name hides object	”Transformer for molecules” without molecule state or label context	add object and representation contract
biology hides method	”docking model” without architecture, objective, or sampler	route method details through AI and Math
formula hides data	clean loss equation without sampled unit or split	add $u$, $q(u)$, preprocessing, and split unit
metric hides objective	training loss and reported metric measure different things	add objective-metric alignment
benchmark hides leakage	high score with row split or unclear allowed information	add split, leakage, and allowed-information check
public post hides uncertainty	Korean synthesis states paper claim as fact	mark claim level and evidence boundary

Claim Record

Use this compact record inside paper notes or posts:

claim:
  text: to verify
  primary_axis: ai | computational-biology | math | evaluation | systems | agents
  object: to verify
  representation: to verify
  model: to verify
  objective: to verify
  metric: to verify
  split: to verify
  evidence_level: mentioned
  public_boundary: to verify

This keeps the note from becoming a loose summary. Every major statement should be traceable to one claim record or to a reusable wiki definition.

Readiness Checks

• The title makes the primary axis clear.
• The biological object is not hidden behind only the model name.
• The model input is distinguished from the raw object.
• The objective is written with sampled unit, distribution, and optimized parameter.
• The reported metric and selection rule are named.
• The split unit matches the generalization claim.
• Coordinate and symmetry behavior are stated for 3D claims.
• Public artifact status is released, not released, to verify, or not applicable.
• Missing metadata stays to verify.
• The public note does not expose private infrastructure, paths, people, tasks, or unpublished results.

Related

• AI paper intake
• Computational Biology paper intake
• Formula intake
• Claim routing
• AI computational biology math paper template
• AI computational biology math readiness gate
• AI computational biology math post intake
• Coverage matrix