The failure boundary of drug discovery.
Built on MolForge-Gen — our proprietary 25.4M-parameter molecular transformer trained on 1.6M drug-like molecules.
Powered by MolForge-Gen · 25.4M Params · 97.5% Validity · 5 Targets Fine-Tuned
Try Live Demo →AI-designed JAK-family kinase inhibitors, confirmed in the lab.
On-target (catalytic) potency independently confirmed by a third-party CRO — predictions made before the assay. These compounds were seeded from JAK-family kinase data and cross-bind TYK2; they are validated potent JAK-family / pan-JAK catalytic binders, not yet a proven TYK2-selective clinical candidate. A TYK2/JAK selectivity panel (JAK1/2/3) is in progress.
STEP 01 — GENERATE
Generate 3,211,243+ structurally diverse compound variants, property-optimized for data-rich validated targets.
01STEP 02 — OPTIMIZE
Rank candidates by 7 criteria: potency, safety, synthesizability, novelty, and more.
02STEP 03 — VALIDATE
Every prediction ships with a conformal uncertainty interval — no black-box point estimates.
03PIPELINE LOGS
Live Processing Queue| ID | Compound | Target | pIC50 | hERG Risk | Tier | |
|---|---|---|---|---|---|---|
| 0001 | MF-TYK2-0001 | TYK2 | 9.32 | MED | Tier 2 | → |
| 0002 | MF-TYK2-0002 | TYK2 | 9.08 | LOW | Tier 2 | → |
| 0003 | MF-TYK2-0003 | TYK2 | 8.93 | MED | Tier 2 | → |
| 0001 | MF-TNIK-0001 | TNIK | 8.41 | LOW | Tier 2 | → |
| 0002 | MF-TNIK-0002 | TNIK | 8.27 | LOW | Tier 2 | → |
| 0009 | MF-TNIK-0009 | TNIK | 8.15 | MED | Tier 2 | → |
Engine
MOLFORGE-GEN ENGINE
Proprietary GPT-2-based molecular transformer. Pre-trained on 1.6M ChEMBL drug-like molecules, then fine-tuned on 5 targets with property-conditioned generation.
25.4M
Model Parameters
8 layers · 512 dim · 8 heads
1.6M
Training Molecules
ChEMBL drug-like SMILES
97.5%
Generation Validity
RDKit-parseable output
5
Targets Fine-Tuned
TYK2 · CDK4 · CDK6 · TNIK · EGFR
R2 Pre-Lock Snapshot · 2026-05-22 KST
9-AXIS EVIDENCE
Every claim under R2 (q=5 portfolio) is backed by a live JSON file in /public/data. 4 candidates pass all 7 stringent axes — TYK2 RBFE and MolFormer LoRA negative results published alongside.
4
Tier 1 GOLD
all 7 axes pass
27
Boltz cofold
TNIK 9 / CDK4 5 / CDK6 5 / EGFR 8
0.81
Kinome R
+0.14 vs baseline
0.279
Boltz σ
3.8× paper precision
852/1700
AiZynth solved
R2+R3 combined · 50.1%
99.5%
PoseBusters
1000 Saturn pool
Section · /evidence
9-axis GOLD funnel + 4 Tier 1 candidates
Vina dock real (73, mean -7.13), GNINA CNN 1000, Boltz N=9 ensemble σ, ChEMBL kinome R 0.81, PoseBusters 99.5%, selectivity v2.
Section · /r2-candidates
R2 CRO portfolio q=5
KRW 100M budget · 5 candidates × KRW 20M. R1 validated (3/3 JAK-family binders @ 2 nM, cross-bind TYK2) — R2 locked to the confirmed series; TYK2/JAK selectivity panel in progress.
Section · /preprint
bioRxiv v0.2 + KIPO + figures
Methods (12 sections) + Discussion + Conclusion drafts, KIPO 5-patent ROBOGATE outline, 5 figures, TNIK licensing narrative.
Research
LATEST INSIGHTS
April 2026
TNIK: The Next Frontier Beyond GLP-1 in Obesity
Why TNIK inhibitors could solve the muscle loss problem that limits current obesity drugs.
Read →
April 2026
The Case for TNIK + GLP-1 Combination Therapy
How combining TNIK inhibitors with GLP-1 agonists could create the first obesity treatment that preserves muscle mass.
Read →
March 2026
TYK2 Case Study: From Zero to Gate A in 14 Days
How MolForge generated 22,910 drug-like variants, filtered to 5,509 ADMET-passed candidates in 14 days.
Read →
Partnership
PARTNER
WITH US.
Request Data Room Access →Top 20 dossiers · Full ADMET profiles · Failure boundary heatmaps · Under NDA
DigitalBiology Co., Ltd.
Pareto Rank-1 is a front-membership count, not a quality ranking — near-saturation means it does not discriminate quality. Meaningful selection uses consensus + wet-lab.
© 2026 DigitalBiology Co., Ltd.