R2 Pre-Lock Evidence Snapshot · 2026-05-22 KST

Every axis, on the record.

9-axis GOLD filter applied to 977,415 in silico compounds. Each axis is a live JSON evidence file. We publish negative results (MolFormer LoRA, TYK2 RBFE NaN) alongside positive ones because honest disclosure is a load-bearing scientific claim, not a footnote.

Status — in silico, pre-wet-lab. Every candidate below is a computational prediction. Wet-lab validation (R1 biochemical assay · 5/15–6/15 window) is pending; measured hit-rate and kinase selectivity will be published on this page on arrival. Until then, reported metrics are predictive, not assay-confirmed.

Section 1 · Funnel

9-axis GOLD funnel

Consensus 6-way → QSAR pIC50 → AEV pKi → AiZynth route → GNINA top 30% → Vina < -6 kcal/mol → QED + Lipinski. 4 candidates pass all 7 stringent axes (5.5% pass rate vs CACHE benchmark 3.7% wet hit-rate).

977,415

Generated

cumulative since 2026-04-13

1,000

Saturn pool

5 targets · 24/7 production

73

Consensus top

6-way deep-learning ≥ 0.85

4

Tier 1 GOLD

7-axis stringent pass

5

R2 portfolio

q=5 pending R1 wet anchor

Section 2 · Tier 1 GOLD

4 candidates pass all 7 axes

snapshot 2026-05-22 02:01:56 KST

Lead candidate · #1

MF-TNIK-6f036c

TNIK kinase · Insilico Rentosertib clinically validated target (Phase IIa Nat Med 2025-06). Tanimoto 0.153 vs Rentosertib — IP-novel chemotype.

Consensus 6-way

0.969

QSAR pIC50

7.83

AEV pKi

5.69

GNINA CNN

5.21

Vina dock

-6.72 kcal

QED

0.852

AiZynth

solved

Tanimoto vs Insilico

0.153

UIDTargetConsensusQSAR pIC50AEV pKiGNINAVina (kcal/mol)QED
MF-TNIK-6f036cTNIK0.96867.835.695.21-6.720.852
MF-CDK4-99a012CDK40.92846.995.245.36-7.020.884
MF-TNIK-576faaTNIK0.90157.795.335.19-6.870.852
MF-TNIK-066333TNIK0.87867.195.505.46-7.480.900

MF-TNIK-6f036c is the strongest single hit (consensus 0.969, QSAR 7.83, AEV 5.69, Vina -6.72, QED 0.85). TNIK target validated by Insilico Rentosertib Phase IIa (Nat Med 2025-06).

Section 3 · Portfolio · 10-axis unified

R2 CRO q=5 — pending R1 anchor

KRW 100M budget · 5 candidates × KRW 20M biochemical assay. Each row shows all 10 evidence axes including Boltz-2 cofold confidence where available. Empty Boltz cells = R3 batch scope; not a quality signal.

#TargetUIDGNINAVinaQEDhERGAiZynthSel.Boltz cofold
1EGFRMF-EGFR-saturn-99085d4.84-6.690.550.33146 ✓0.68R3 batch
2TNIKMF-TNIK-saturn-3261bb5.20-8.530.800.67189 ✓0.372
3TNIKMF-TNIK-saturn-df13314.60-8.290.670.35100 ✓R3 batch
4CDK4MF-CDK4-saturn-13d49b4.62-7.210.840.07147 ✓-0.62R3 batch
5CDK6MF-CDK6-saturn-96690c4.71-6.060.780.31358 ✓0.21R3 batch

Boltz cofold coverage today: 1 / 5 R2 candidates (MF-TNIK-saturn-3261bb, conf 0.37 — TNIK pocket novel-chemotype). Remaining 4 cofolds scheduled as R3 batch run; Vina + GNINA + AEV + Kinome + AiZynth all complete now.

Section 4 · Structure-based validation

Five independent physics signals

Vina + GNINA + AEV-PLIG + Boltz-2 N=9 ensemble + Boltz-2 cofold (27 compounds across 4 targets, 5-seed each). CDK4 cofold confidence 0.93 (excellent); TNIK 0.37–0.46 (low — reflects novel chemotype pocket fit, not affinity).

_phase_h_vina_73.json

AutoDock Vina (real dock)

n compounds
58
mean score
0.55 kcal/mol
binders < -6
0 / 58 (0%)
scope
consensus_top_20 (73)

_phase_a1_gnina_1000_complete.json

GNINA CNN affinity

n compounds
1000
mean CNN affinity
4.10
≥ 5.0 affinity
84 / 1000
orthogonality
3/4 targets |ρ|<0.2 vs AEV

_phase_boltz_cofold_aggregate.json

Boltz-2 cofold (27 compounds)

n unique compounds
27
mean confidence
0.698
per target
TNIK:9 CDK4:5 CDK6:5 EGFR:8
best target conf
CDK4 0.93 · TNIK 0.37

_phase_b1_boltz_n9.json

Boltz-2 N=9 affinity ensemble

n seeds
9
mean σ
0.279 pIC50
paper σ
1.5 kcal/mol (Wohlwend 2025)
precision gain
3.8× vs paper

_phase_b3_posebusters_1000.json

PoseBusters mol mode

n input
1000
pass
995 / 1000
pass rate
99.5%
elapsed
348 s

_phase_a1_aizynth_combined_2000.json

AiZynthFinder retrosynth

n pool
1,700 (R2 + R3)
R2 partial 1-1000
319 / 700 (45.6%)
R3 complete 1001-2000
533 / 1000 (53.3%)
combined solved
852 / 1700 (50.1%)

Section 5 · Boltz-2 cofold breakdown

27 compounds · 4 targets · 5-seed ensemble

Each compound co-folded with its target protein 5 times. Mean confidence per row is averaged across all 5 model seeds. CDK4 confidence 0.93 = pocket-aligned binding; TNIK 0.41 reflects novel chemotype outside training distribution (not an affinity signal — pocket fit only).

0.371

TNIK

9 cofold · mean conf

0.927

CDK4

5 cofold · mean conf

0.886

CDK6

5 cofold · mean conf

0.845

EGFR

8 cofold · mean conf

#UIDTargetScopeSeedsConf meanConf maxComplex ipLDDT
1MF-CDK4-saturn-a8ea09CDK4vinatop550.9330.9340.907
2MF-CDK4-saturn-5caa6dCDK4vinatop550.9290.9340.897
3MF-CDK4-saturn-4a94e0CDK4vinatop550.9270.9310.888
4MF-CDK4-saturn-9a7cc3CDK4vinatop550.9230.9350.880
5MF-CDK4-saturn-1b1931CDK4vinatop550.9220.9260.882
6MF-CDK6-saturn-526e0eCDK6vinatop550.8970.8990.871
7MF-CDK6-saturn-8de887CDK6vinatop550.8910.8960.836
8MF-CDK6-saturn-17d66fCDK6vinatop550.8900.8960.809
9MF-CDK6-saturn-d295ffCDK6vinatop550.8770.8820.789
10MF-CDK6-saturn-40c8d2CDK6vinatop550.8740.8890.809
11MF-EGFR-saturn-8b8319EGFRmsa50.8590.8680.794
12MF-EGFR-saturn-97e109EGFRmsa50.8590.8760.762
13MF-EGFR-saturn-2fa503EGFRgold50.8480.8570.723
14MF-EGFR-saturn-36ff3aEGFRgold50.8450.8600.684
15MF-EGFR-saturn-29d9f5EGFRmsa50.8410.8540.680
16MF-EGFR-saturn-fbe055EGFRmsa50.8410.8650.684
17MF-EGFR-saturn-37d80bEGFRgold50.8380.8460.698
18MF-EGFR-saturn-7b6c63EGFRmsa50.8320.8730.634
19MF-TNIK-saturn-f35503TNIKv250.3980.4550.262
20MF-TNIK-saturn-dd5618TNIKv250.3830.4260.263
21MF-TNIK-saturn-c7e12dTNIKvinatop550.3810.4360.263
22MF-TNIK-saturn-01449fTNIKvinatop550.3790.4620.261
23MF-TNIK-saturn-01449fTNIKv250.3790.4620.261
24MF-TNIK-saturn-3261bbTNIKv250.3720.4030.263
25MF-TNIK-saturn-3e05b8TNIKv250.3610.3890.248
26MF-TNIK-saturn-3262c6TNIKvinatop550.3550.3690.253
27MF-TNIK-saturn-2a9163TNIKvinatop550.3520.3840.246
28MF-TNIK-saturn-9d69c4TNIKvinatop550.3490.4090.234

28 records · 27 unique compounds (1 dup across vinatop5/v2 scopes for MF-TNIK-saturn-01449f).

Section 6 · Affinity model upgrade

ChEMBL kinome multi-task R 0.81 mean

Morgan FP r=2 × RF n=300 trained on 8,166 ChEMBL active SMILES across 5 kinases. Mean R 0.81 = +0.14 vs internal QSAR baseline (0.671). Classical feature engineering beats foundation-model frozen embeddings at our dataset scale.

0.845

TYK2

n = 898

0.825

CDK4

n = 952

0.716

CDK6

n = 323

0.858

EGFR

n = 766

file: _phase_i_kinome_chembl.json · TNIK n=9 sparse (excluded from mean).

Section 7 · IP novelty

Insilico Rentosertib comparison

13 MolForge TNIK consensus_top_20 candidates vs Rentosertib (Insilico INS018_055, TNIK inhibitor, Phase IIa positive readout, Nat Med 2025-06). Clinically validated target + structurally independent chemotype = licensing dual advantage.

0.153

Mean Tanimoto

Morgan r=2, 2048bit

0.238

Max Tanimoto

closest pair

0.108

Min Tanimoto

furthest pair

0 / 13

Same Murcko

scaffold identity

Full structure gallery →

Section 8 · Honest disclosure

Three things that did not work

Cherry-picking positives is the failure mode of AI drug discovery. Below are three negative results we encountered today, published as evidence in their own JSON files.

_phase_a2_molformer_lora.json

MolFormer-XL LoRA (r=16) multi-task

R 0.565 (5-target mean)

Foundation model + LoRA fine-tune underperforms per-target Morgan+RF baseline (R 0.671). Multi-task interference confirmed: TYK2 0.487, TNIK 0.437, CDK4 0.463, CDK6 0.538, EGFR 0.807. Frozen + Morgan concat closes gap (R 0.666) for n>2000 targets only.

Classical feature engineering wins at our dataset scale.

_phase_c_tyk2_rbfe_failure.json

TYK2 RBFE OpenFE 1.7 quickrun

FAILED — NaN

SimulationNaNError: Propagating replica 0 at state 0 resulted in a NaN! · Started from X-ray PDB 4GIH (pdbfixer cleaned, 4841 atoms). NaN persists even with crystal protein — not a Boltz cofold pose issue, fundamental force field or alchemical topology problem.

TYK2 ABFE-grade ΔΔG unavailable. R2 falls back to Vina + GNINA + AEV + Boltz.

_phase_e_selectivity_v2.json

Saturn pool selectivity (Kinome v2)

0 / 1000 selective

Median on-vs-off-target diff is OOD-negative (−0.28). Saturn-novel chemotypes lie outside ChEMBL active distribution → unreliable selectivity prediction. 73 compounds show any positive diff, 0 show > 1.0 log unit advantage.

R2 CRO biochemical panel (≥20 kinases) required for selectivity anchor. In parallel, a KLIFS-anchored kinome-wide selectivity model is on the roadmap (P1-6) to bring off-target prediction in-distribution for novel chemotypes.

Section 9 · Live GPU operations

complete

AiZynth 1001-2000 — DONE

R3 pool retrosynth: 533/1000 solved (53.3%) in 96.8 min. TNIK 419 > EGFR 77 > CDK4 28 > CDK6 9. Combined R2+R3: 852/1700 solved (50.1%). Output: phase_A1_aizynth_combined_2000.json

paused

Saturn watchdog cron

Watchdog #PAUSED prefix applied. Hourly target crons (5x) still active for compound generation. ROBOGATE GPU work can proceed.

complete

Boltz cofold chain

27 unique compounds recovered (TNIK 9 + CDK4 5 + CDK6 5 + EGFR 8). 5-seed ensemble per compound = 135 model evaluations.

Snapshot: 2026-05-22 05:11:53 KST · Auto-refresh via ISR 60s · /api/stats provides live API data