DDLS Week 4 · Seminar 4 · 01 / 06
Can this KRAS model support residue-level design?
The owner wanted five residues that could guide design of a KRAS inhibitor series. The starting model looked confident, but I needed to determine whether it actually supported that residue-level decision.
Dr Yuki Tanaka · medicinal chemistry KRAS G12D assay construct
02 / What I built
I turned an interview into a testable brief.
The workflow separated three questions: what protein is this, how confident is the local model, and what can I honestly claim?
Owner interviewdesign decision
→
→
Identity checksequence + boundaries
→
Confidenceper-residue pLDDT
→
Claim
Reliable local shape in the pocket neighbourhood.
Model
Construct-matched KRAS G12D 1–169 fold.
Boundary
Model support is not experimental validation.
03 / The right confidence · visual centrepiece
For a local claim, use local confidence.
The main evidence is per-residue pLDDT in the actual KRAS G12D 1–169 fold — not global pTM.
CHAIN A · B-FACTOR = pLDDT · AUTO-ROTATE
Requested regions
10–17 has the consistently higher local confidence. This is model evidence, not experimental validation.
Full sequence · 1–169
1255075100125150169
■ selected residues · shaded = 10–17
G1087.70
V1489.92
G1590.79
K1687.06
S1787.03
04 / The trap
The confident AlphaFold model was not the exact assay construct.
AlphaFold DB reference
KRAS P01116
G12 · 1–189
canonical reference
Owner's assay construct
KRAS G12D
D12 · 1–169
one chain · no tag
G12 / G12D mismatch
plus the AlphaFold-only C-terminal overhang: residues 170–189
I therefore checked the construct-matched G12D 1–169 structure rather than relying on the canonical model alone. I did not treat the canonical model as sufficient evidence for residue-level contacts in the assay construct.
05 / The answer
Five anchors — one high-confidence local region.
G10 · V14 · G15 · K16 · S17
Model-supported design anchors
Gly1087.70pLDDT
Val1489.92pLDDT
Gly1590.79pLDDT
Lys1687.06pLDDT
Ser1787.03pLDDT
These residues are supported as structural design anchors by the construct-matched model.
Not experimentally validated ligand contacts.
The shortlist supports hypothesis generation and design prioritisation — not proof of binding.
06 / Limitations + AI use
What I can say — and what I still cannot know.
What the model supports
- Construct-matched KRAS G12D 1–169 fold was checked.
- Region 10–17 has higher local pLDDT.
- Five residues can guide design as model-supported anchors.
What it does not establish
- Experimental ligand contacts.
- Residue-level mutational validation.
- The ligand-bound conformational state.
The predicted structure does not establish the ligand-bound conformational state or experimental residue-level contacts.
Next step
Compare an appropriate experimental structure, or use an appropriate complex prediction and evaluate interface/structural confidence.
AI use disclosure
Agent A — owner interview and biological question.
Pi — structural analysis, viewer and presentation draft.
I verified the construct/sequence, residue numbering, pLDDT values, selected residues, structure match and limitations.