gpu_resultApr 10, 2026· SMA Research Platform
Seven New LIMK2-Selective Hits from Overnight DiffDock Campaign (Total: 14)
#2026-04-10#LIMK2#DiffDock#PocketXMol#selectivity#drug-discovery#ROCK-LIMK-CFL2
TL;DR
Seven additional molecules from the PocketXMol DFG-out Type II library passed our LIMK2 selectivity filter overnight (2026-04-09 → 04-10). Combined with the previous session's seven, we now hold 14 computational LIMK2-selective candidates — all with selectivity margin ≥ 0.30 against LIMK1 and ROCK1.
New hits (overnight chunk)
| # | mol_id | LIMK2 | LIMK1 | ROCK1 | Margin |
|---|---|---|---|---|---|
| 1 | 84_0 | −0.17 | −0.85 | −1.23 | 0.68 |
| 2 | 851_0 | −0.05 | −0.73 | −4.03 | 0.68 |
| 3 | 293_0 | −0.68 | −1.48 | −1.22 | 0.54 |
| 4 | 987_0 | −0.21 | −0.73 | −0.67 | 0.46 |
| 5 | 1019_0 | −0.51 | −0.97 | −1.87 | 0.46 |
| 6 | 307_0 | −0.65 | −1.04 | −1.87 | 0.39 |
| 7 | 434_0 | −0.86 | −1.19 | −2.05 | 0.33 |
Selectivity margin = LIMK2 confidence − max(LIMK1, ROCK1). Higher = more selective.
DiffDock confidence is log-odds of pose correctness, not binding affinity, so ranking is comparative, not absolute.
Combined candidate pool (n = 14)
- Session 1 (pre-2026-04-09): 7 hits from the first PocketXMol DFG-out + DiffDock screen
- Session 2 (overnight 2026-04-09 → 04-10): 7 new hits from batches 1, 2, 4 (this finding)
- Total: 14 candidates passing margin ≥ 0.30 against LIMK1 and ROCK1
Next steps
- ADMET triage — drop anything failing BBB/DILI/hERG/CYP3A4
- POCKET_FIXED re-dock on top 10 to eliminate near-miss poses
- 10 ns MD contact persistence on top 3–5 survivors vs LIMK2/LIMK1/ROCK1
- MMPBSA ddG vs BMS-5 and LIMKi3 reference runs
- ChEMBL/PubChem novelty check (Tanimoto > 0.85 = known → discard)
- Present top 1–3 to Simon with full evidence package — NOT preliminary scores
Caveats
- DiffDock confidence is pose quality, not binding strength
- Several hits contain unusual SMILES valence patterns (
c=2,=c1, aromatic N artefacts) — RDKit sanitization + manual inspection required before purchase - Selectivity only tested against LIMK1 + ROCK1. JAK2, ROCK2, kinome-wide panels still open
- This is a computational prioritization list, NOT a therapeutic claim
Compute: 4 × Vast.ai GPUs overnight, ~8 h, ~$8 total. CC-BY-4.0. Full finding at /findings/2026-04-10/FINDING_2026-04-10_new_7_selective_hits.md.