Term validator misses GO/UBERON slot misuse due missing Descriptor term bindings

[cmungall]

Problem

We are seeing ontology-slot misplacements pass just validate / just validate-terms, e.g.:

• UBERON:0002107 (liver) placed under biological_processes
• GO Molecular Function terms (e.g. GO:0003824, GO:0022857) placed in biological_processes

These should fail curation QC, but currently pass in several PR branches.

Repro

Example from CPS1 branch/PR #407:

• kb/disorders/Carbamoyl_Phosphate_Synthetase_I_Deficiency.yaml
• Node Impaired mitochondrial ureagenesis had UBERON:0002107 under biological_processes.

Running:

• just validate kb/disorders/Carbamoyl_Phosphate_Synthetase_I_Deficiency.yaml
• just validate-terms kb/disorders/Carbamoyl_Phosphate_Synthetase_I_Deficiency.yaml

Both passed before manual file fix.

Root Cause

In src/dismech/schema/dismech.yaml, descriptor classes lack term binding constraints:

• BiologicalProcessDescriptor.slot_usage.term has no bindings to BiologicalProcessTerm.
• AnatomicalEntityDescriptor.slot_usage.term has no bindings to AnatomicalEntityTerm.

So the validator checks term existence/label but not ontology branch constraints (reachable_from GO:0008150, reachable_from UBERON:0001062) for these descriptor usages.

Proposed Fix

Add explicit bindings in descriptor slot usages, analogous to CellTypeDescriptor and GeneDescriptor:

• BiologicalProcessDescriptor.term -> range: BiologicalProcessTerm
• AnatomicalEntityDescriptor.term -> range: AnatomicalEntityTerm
• (Optionally also tighten other descriptor classes where intended.)

Then add/adjust a regression test asserting that:

• GO MF term in biological_processes fails
• UBERON term in biological_processes fails

Context

Raised while curating metabolic PRs where these were repeatedly flagged by reviewer but not by validator. This issue captures the validator-side gap so these become machine-caught.

Reported by Codex.

[github-actions]

Summary & Analysis

This is a schema validator gap issue: two descriptor classes in src/dismech/schema/dismech.yaml lack bindings that would enforce ontology branch constraints, allowing nonsensical cross-slot placements to pass QC silently.

---

Root Cause (Confirmed in Schema)

The CellTypeDescriptor and GeneDescriptor classes have explicit bindings that enforce ontology branch membership:

CellTypeDescriptor:
  slot_usage:
    term:
      bindings:
        - binds_value_of: id
          range: CellTypeTerm      # reachable_from CL:0000000
          obligation_level: REQUIRED

GeneDescriptor:
  slot_usage:
    term:
      bindings:
        - binds_value_of: id
          range: GeneTerm
          obligation_level: REQUIRED

But BiologicalProcessDescriptor and AnatomicalEntityDescriptor have no such bindings — just a description string:

BiologicalProcessDescriptor:
  slot_usage:
    term:
      description: Optional GO biological process term reference   # ← no enforcement

AnatomicalEntityDescriptor:
  slot_usage:
    term:
      description: Optional UBERON anatomical entity term reference  # ← no enforcement

The target enum types already exist in the schema:

• BiologicalProcessTerm — reachable_from GO:0008150 (biological_process)
• AnatomicalEntityTerm — reachable_from UBERON:0001062 (anatomical entity)

The fix is straightforward: add bindings to both descriptors, mirroring CellTypeDescriptor.

---

Important Tension with Issue #361

This issue has a direct conflict with the open issue #361 — "Use of GO Molecular Function terms in biological_processes field".

Issue #361 documents that 59 GO Molecular Function (MF) terms are currently used across ~20 disorder files in the biological_processes field — legitimately and with maintainer endorsement (cmungall suggested using MF terms like GO:0004722 phosphatase activity instead of BP terms for molecular activities). Examples:

• GO:0004672 (protein kinase activity)
• GO:0004713 (protein tyrosine kinase activity)
• GO:0005088 (Ras GEF activity)

If BiologicalProcessDescriptor is bound strictly to BiologicalProcessTerm (reachable_from GO:0008150), all those legitimate MF-term usages would fail validation. This is a pre-existing design tension that needs resolution before or alongside this fix.

Options:

1. Strict fix (BP-only) → Add BiologicalProcessTerm binding, then migrate the 59 MF terms to a new molecular_functions slot (as Option 2 in Use of GO Molecular Function terms in biological_processes field - schema design options #361)
2. Permissive fix → Create a broader GOTerm binding (allowing BP and MF subtrees), giving partial protection while preserving existing data
3. Two-phase → Fix AnatomicalEntityDescriptor now (no data conflict), and defer BiologicalProcessDescriptor until Use of GO Molecular Function terms in biological_processes field - schema design options #361 is resolved

---

Concrete Implementation (for the non-controversial part)

AnatomicalEntityDescriptor binding is unambiguous — UBERON terms in biological_processes are clearly wrong and this fix has no data compatibility risk:

AnatomicalEntityDescriptor:
  is_a: Descriptor
  description: A descriptor for anatomical locations, bindable to UBERON
  slot_usage:
    term:
      description: Optional UBERON anatomical entity term reference
      bindings:
        - binds_value_of: id
          range: AnatomicalEntityTerm
          obligation_level: REQUIRED

---

Recommended Next Steps

1. Fix AnatomicalEntityDescriptor immediately — add binding to AnatomicalEntityTerm. No existing data conflicts.
2. Resolve Use of GO Molecular Function terms in biological_processes field - schema design options #361 to decide whether biological_processes allows GO MF terms. If yes, either rename the field or use a broader GO-term binding.
3. Fix BiologicalProcessDescriptor after Use of GO Molecular Function terms in biological_processes field - schema design options #361 decision — either bind to BiologicalProcessTerm (strict) or a broader GO root.
4. Add regression tests asserting UBERON-in-biological_processes and GO-MF-in-biological_processes fail validation.

@dragon-ai-agent please implement the fix for this issue — add bindings to BiologicalProcessDescriptor (using a broad GO root to avoid breaking existing MF term usage) and AnatomicalEntityDescriptor, and add regression tests. See the tension with issue #361 noted in the summary comment.