Skip to main content
Case Study

Overcoming KRAS G12C Resistance: From Prompt to Lead Compounds in 6 Hours

A single natural-language prompt drove an autonomous drug discovery campaign that identified three lead compounds targeting sotorasib-resistant KRAS G12C — compressing 4–6 weeks of traditional work into a single afternoon.

0
Compounds screened
0
Qualified hits
0
Tier 1 leads
0 hrs
Total elapsed
The Challenge

KRAS G12C is the most actionable oncogene in non-small cell lung cancer, affecting ~13% of patients. Two approved covalent inhibitors — sotorasib and adagrasib — have transformed treatment, but acquired resistance mutations (Y96D, R68S, H95Q) now limit durability of response in the majority of patients.

Identifying next-generation inhibitors that maintain potency against these resistance variants requires screening large chemical libraries against the mutant binding pocket — a process that traditionally takes weeks of computational chemist time to set up, execute, and interpret.

The Input

One natural-language prompt. No manual configuration, no scripting, no queue management.

Prompt
“Discover next-generation KRAS G12C inhibitors that overcome acquired resistance mutations (PDB: 8AZX) from small molecule covalent inhibitors. Screen for selectivity against wild-type KRAS and common resistance mutants Y96D, R68S, and H95Q. Disease context: non-small cell lung cancer with sotorasib resistance. Max 300 compounds.”
How the Platform Worked

Five autonomous phases, one prompt

1

Research & Intelligence

Target dossier from PubMed, ChEMBL, PDB. KRAS G12C biology, Switch II pocket, resistance mechanisms, 100 known actives.

3 min1–2 weeks
2

Library Design & Acquisition

500 adagrasib-class compounds retrieved. 398 passed quality gates with 91% warhead coverage.

5 min3–5 days
3

Covalent Docking

1,408 conformers docked via GNINA covalent mode across 4 warhead groups with auto-classified SMARTS patterns.

5.5 hrs1–2 weeks
4

Physics-Based Validation

26 top compounds rescored with MM-GBSA + 1 ns molecular dynamics to separate genuine binders from artifacts.

45 min2–3 days
5

Multi-Agent Interpretation

4 parallel agents: screening quality, binding mode, ADMET triage, and SAR synthesis.

<10 min1 week
Total elapsed time
~6 hoursvs4–6 weeks
Compression
~40×
Autonomous Intelligence

Three expert-level decisions the platform made automatically

These domain-specific judgement calls would normally require an experienced medicinal chemist familiar with the KRAS G12C literature. They directly determined the quality of the output.

Disabled Brenk structural alerts

The Michael acceptor warhead IS the pharmacophore for covalent Cys12 engagement, not a liability

Raised MW threshold to 1,350 Da

KRAS G12C inhibitors require large, sp3-rich scaffolds to span the shallow cryptic pocket (sotorasib: 560 Da, adagrasib: 604 Da)

Reweighted scoring: 70% physics / 30% CNN

Large covalent inhibitors fall outside CNN training domains — physics-based methods are more reliable here

Results at a Glance

Campaign output

398
Compounds screened
1,408 conformers
173
Qualified hits
score < −5 kcal/mol + CNN > 0.4
147
Unique scaffolds
chemical diversity
34
Composite score > 0.9
high-confidence hits
−16.55
Best docking score
kcal/mol
−82.9
Best MM-GBSA energy
kcal/mol
0
PAINS-positive
clean library
3
Tier 1 leads
recommended for synthesis
Key Findings

Three lead compounds identified

CPD_0038

Best physics-validated binder
MM-GBSA
−82.9 kcal/mol
RMSD
1.0 Å
Warhead
Fluoroacrylamide

Slower, more selective covalent reactivity than standard acrylamide

Predicted poor oral bioavailability — requires formulation assessment or IV dosing

CPD_0035

Best overall profile
Composite
0.964
Bioavail.
>80%
Warhead
Acrylamide

Fluoropyridyl aryl wing predicted to tolerate Y96D resistance mutation. No AMES flag.

Immediate candidate for biochemical IC₅₀

CPD_0042

Tri-metric convergence
MM-GBSA
−17.9 kcal/mol
Docking
−10.21 kcal/mol
CNN Affinity
8.819
Warhead

Strong convergence across all three scoring methods

Commercially purchasable for immediate testing

Structure-Activity Relationships

Two dominant scaffold series

Series A

126 compounds

Fluoropyridyl-quinazoline

  • Chiral fluoromethyl pyrrolidine linker — sp3-rich 3D shape
  • Better ligand efficiency (LE 0.23–0.29) and oral bioavailability
  • Predicted more tolerant to Y96D due to compact aryl wing
  • Both acrylamide and fluoroacrylamide warheads represented

Series B

108 compounds

Chlorofluoronaphthalene pyridopyrimidine

  • Bridged bicyclic (DABCO-type) ether linker
  • Higher raw CNN affinity but lower ligand efficiency (LE 0.21–0.26)
  • Extended naphthalene provides deeper P2 pocket occupation
  • May clash with Y96D aspartate — resistance risk
ADMET Profiling

Class-obligate vs. addressable liabilities

Class-obligate

Shared with approved drugs — accepted risks:

  • Michael acceptor warheads
  • Lipinski MW violations
  • DILI risk

Addressable

Compound-specific — require optimization:

  • Pan-CYP inhibition (planar aromatics)
  • hERG risk (basic amines)
  • 7 compounds AMES-positive — hard rejected
Timeline Compression

What this demonstrates

The platform didn't just run faster — it made domain-expert decisions that would normally require an experienced medicinal chemist familiar with the KRAS G12C literature.

Step
Traditional
AiDDA
Literature review & target assessment
1–2 weeks
3 minutes
Library design & compound sourcing
3–5 days
5 minutes
Docking setup, execution, post-processing
1–2 weeks
5.5 hours
MM-GBSA rescoring & analysis
2–3 days
45 minutes
Hit triage, SAR, ADMET review
1 week
10 minutes
Total
4–6 weeks
~6 hours
Outcome

A prioritised, tiered portfolio of lead compounds with full SAR context, ADMET profiling, resistance mutation analysis, and a concrete experimental roadmap — ready for a project team meeting the same day the campaign was launched.

Full Report

Campaign report

The complete AiDDA-generated report including target intelligence, screening results, safety assessment, SAR analysis, and experimental plan.

Download PDF

Run a campaign on your target

AiDDA runs the full virtual screening pipeline — from a single prompt to a ranked compound portfolio — in hours. See what it finds for your protein target.

No credit card required