Technical Deep Dive
METHODOLOGY
A detailed technical reference for how MolForge generates, validates, and ranks drug candidates across the full discovery pipeline.
01
Generative Chemistry: MolForge-Gen
MolForge-Gen is a proprietary 25.4M-parameter Transformer model (GPT-2 architecture) for conditional molecular generation. Unlike off-the-shelf tools, MolForge-Gen natively understands target context, property constraints, and patent boundaries as first-class generation conditions.
- Architecture: Decoder-only Transformer, 8 layers, 512dim, 8 heads (25.4M params)
- Pre-trained: 1.6M drug-like SMILES from ChEMBL (Validity 97.5%)
- Multi-target: Single model generates for TYK2/CDK4/CDK6/TNIK/EGFR via target tokens
- Property-conditioned: QED, hERG, pIC50, MW ranges as input tokens
- FTO-aware: Tanimoto distance constraint learned during RL (100% FTO safe)
Model Specifications
| Parameters | 25.4M |
| Architecture | GPT-2 Decoder-Only |
| Pre-train Data | 1.6M SMILES (ChEMBL) |
| Pre-train Validity | 97.5% |
| Pre-train Uniqueness | 99.8% |
| Fine-tune Targets | 5 (TYK2/CDK4/CDK6/TNIK/EGFR) |
| TNIK Accuracy | 99.0% |
| RL FTO Safe | 100% |
| RL Avg QED | 0.872 |
02
ADMET Prediction: Dual Model Screening
Every generated compound passes through ADMET-AI (Swanson et al., 2024) for 104-property prediction. Compounds failing any of the five critical thresholds are eliminated before further analysis.
| Property | Model | Threshold | Action |
|---|---|---|---|
| hERG Cardiotoxicity | ADMET-AI (ChEMBL) | > 0.7 | Eliminate |
| AMES Mutagenicity | ADMET-AI (ChEMBL) | > 0.5 | Eliminate |
| HIA (Oral Absorption) | ADMET-AI (TDC) | < 0.5 | Eliminate |
| Molecular Weight | RDKit | > 500 Da | Eliminate |
| PAINS Alerts | RDKit SMARTS | > 0 | Eliminate |
| QED | RDKit | < 0.3 | Eliminate |
03
Conformal Prediction: Uncertainty Quantification (Calibration Insufficient — n=3)
pIC50 predictions carry a split-conformal prediction interval (Vovk et al., 2005). In theory split conformal gives a distribution-free coverage guarantee on exchangeable data — but that guarantee is only meaningful with an adequate, in-distribution calibration set, which we do not yet have. Our current wet-lab calibration sample is n=3, the observed marginal coverage is ~0.78 (below the 0.88 target), and the TNIK prospective points fell outside the predicted interval (out-of-distribution). We therefore report these intervals as uncalibrated / insufficient-n / OOD, not as a validated 90% calibrated CI on every prediction.
“Split conformal is a sound framework, but a coverage number is only as trustworthy as its calibration set. With n=3, ~0.78 empirical coverage, and OOD TNIK behavior, we treat our current intervals as provisional and uncalibrated — not as a guaranteed 90% CI.”
Validation Status (Insufficient Calibration)
| Target Coverage | 90% (nominal, not met) |
| In-silico holdout coverage | 86.5% (scaffold-split, not wet-lab) |
| Wet-lab calibration n | 3 (insufficient) |
| Wet-lab marginal coverage | ~0.78 (below 0.88 target) |
| TNIK prospective points | OUTSIDE interval (OOD) |
| Status | Uncalibrated / insufficient-n / OOD |
| Method | Split Conformal (Mondrian) |
| Interval Width | Adaptive (per-compound) |
04
ROBOGATE: Failure Boundary Mapping
ROBOGATE uses adaptive sampling to explore the multi-dimensional property space around a target. Instead of optimizing for a single objective, it maps the Pareto frontier across 7 competing criteria: potency (pIC50), cardiac safety (hERG), mutagenicity (AMES), oral absorption (HIA), drug-likeness (QED), synthesizability (SA Score), and patent novelty (Tanimoto distance).
7
Optimization Axes
500
Pareto Rank-1 Count (TYK2)
33,755
TNIK Compounds Mapped
Note: Pareto Rank-1 count is a front-membership count, not a quality ranking. Because the front is near-saturated (most candidates land on it), this number does not discriminate quality — meaningful selection relies on consensus scoring + wet-lab validation, not front membership.
05
Boltz-2: Structure Prediction & Docking
Top-ranked compounds are submitted to Boltz-2 (MIT/Recursion, open-source) for protein-ligand co-folding. Compounds achieving ligand_ptm ≥ 0.85 are promoted to Tier 2 (Structure Verified). Boltz-2 provides binding affinity predictions approaching FEP accuracy at 1,000× the speed.
~15s
Per Complex
0.85+
ligand_ptm Threshold
100%
Success Rate (TYK2)
Tier 2
Promotion Level
06
MolForge Score: 7-Axis Compound Ranking
The MolForge Score is a composite metric that ranks compounds across 7 axes using Pareto optimization. Unlike single-objective scoring, it identifies compounds that are not dominated on any axis — the true drug-like sweet spot.
| Axis | Source | Weight | Direction |
|---|---|---|---|
| pIC50 (Potency) | QSAR GNN | High | Maximize |
| hERG (Cardiac Safety) | ADMET-AI | High | Minimize |
| AMES (Mutagenicity) | ADMET-AI | High | Minimize |
| QED (Drug-likeness) | RDKit | Medium | Maximize |
| SA Score (Synthesizability) | RDKit | Medium | Minimize |
| HIA (Oral Absorption) | ADMET-AI | Medium | Maximize |
| Tanimoto (Novelty) | RDKit Morgan FP | Low | Maximize distance |
07
Freedom-to-Operate Analysis
Every compound is compared against known patented structures using Morgan fingerprint Tanimoto similarity. For TNIK specifically, all compounds are compared against Insilico Medicine's rentosertib (the only clinical-stage TNIK inhibitor).
TNIK FTO Status:All 33,755 ADMET-validated TNIK compounds have Tanimoto similarity < 0.21 vs rentosertib. Industry standard for structural novelty is Tanimoto < 0.85. MolForge compounds are 4× more structurally distinct than the minimum required threshold.
Model Roadmap
MolForge-Gen v1 → v2
v2 training begins Q2 2026 on RTX 5090 (32GB). Targets ~2x model capacity, 5x pre-train data, and explicit reasoning-style multi-step generation. v1 → v2 metrics will be tracked publicly.
| Spec | v1 (Current) | v2 (Q3 2026 Target) | Δ |
|---|---|---|---|
| Parameters | 25.4M | ~50-100M | 2-4× |
| Architecture | GPT-2 decoder, 8L × 512d × 8h | GPT-2 decoder, 12L × 768d × 12h | Deeper + wider |
| Pre-train Data | 1.6M ChEMBL drug-like | ~5-10M (ChEMBL + PubChem subset) | 3-6× |
| Validity | 97.5% | ≥99% target | +1.5% |
| Novelty (vs ChEMBL) | 96.8% | ≥98% target | +1.2% |
| Property-Conditioning | QED, hERG, pIC50 ranges | + multi-objective (Pareto-aware) | Native multi-objective |
| Targets Fine-Tuned | 5 (TYK2, CDK4, CDK6, TNIK, EGFR) | 20+ (Kinome roadmap) | 4× |
| Compute (training) | RTX 4090 / ~3 days | RTX 5090 / ~5-7 days | Same generation |
Status
Architecture finalized
Data preprocessing in progress
Training start
2026-05
Estimated
First public benchmark
2026-Q3
v2 vs v1 comparison report
Open-Source Tool Stack
MolForge-Gen
Generative Chemistry
Proprietary
Boltz-2
Structure Prediction
MIT
ADMET-AI
ADMET Prediction
MIT
RDKit
Cheminformatics
BSD
AutoDock Vina
Molecular Docking
Apache 2.0
PyTorch
GNN Training
BSD
Open Targets
Target Validation
Apache 2.0
Conformal Pred.
Uncertainty Quantification
Internal