Validation & evidence

Lead with external spot-checks you can verify against ChEMBL literature, then reproducible generated libraries. RL generation step count affects how long Pathfinder optimizes new molecules — it does not retrain the affinity model. Absolute pKd on every marketed drug is harder than rank-ordering within your own series; both views are documented here.

Reproduce in the app: Open Pathfinder demo for a pre-scored 10-compound EGFR library (200 RL steps), or paste any SMILES into the workbench. Scores use VectaBind API v1.0.0 · same Stage 6 model for all targets. EGFR scoring pocket: PDB 2ITK (PDBBind co-crystal); 3D viewer may show PDB 1M17.

External spot-check: known JAK2 inhibitors

Literature pKd values from ChEMBL bioactivities (JAK2 biochemical assays; rounded). VectaBind predictions scored live against the JAK2 pocket embedding (PDB 2b7a, PDBBind co-crystal), June 2026. These are scorer checks on approved drugs — independent of any RL generation run.

Compound	Class	ChEMBL	Literature pKd	VectaBind pKd	\|Δ\|
Ruxolitinib	JAK1/2	CHEMBL214485	8.52	8.92	0.40
Baricitinib	JAK1/2	CHEMBL3301582	8.70	8.94	0.24
Upadacitinib	JAK1 selective	CHEMBL3989951	9.00	8.85	0.15
Filgotinib	JAK1 selective	CHEMBL3989956	8.52	8.82	0.30
Tofacitinib	Pan-JAK	CHEMBL221959	7.70	8.54	0.84
Fedratinib	JAK2 selective	CHEMBL3989953	8.40	9.04	0.64
Momelotinib	JAK1/2	CHEMBL3989955	8.00	8.77	0.77
Pacritinib	JAK2	CHEMBL3989954	7.70	8.77	1.07

How to read this: On JAK2 pocket 2b7a, approved JAK inhibitors track literature closely — MAE ~0.55 pKd across eight marketed compounds. This reflects target/chemotype fit of the pre-trained scorer, not how many RL steps were used in a separate generation run.

External spot-check: known EGFR inhibitors

Same scorer, different target. Literature pKd from ChEMBL kinase assays (rounded). Scored live against EGFR pocket 2ITK, June 2026. Viewer may show PDB 1M17. Use for transparency, not as a global MAE claim.

Compound	Scaffold	ChEMBL	Literature pKd	VectaBind pKd	\|Δ\|
Erlotinib	Quinazoline Type I	CHEMBL553	9.00	5.74	3.26
Gefitinib	Quinazoline Type I	CHEMBL939	9.00	5.65	3.35
Lapatinib	Dual kinase	CHEMBL554	8.52	7.29	1.23
Osimertinib	Covalent 3rd-gen	CHEMBL3353410	9.70	7.42	2.28
Afatinib	Irreversible	CHEMBL1173655	9.30	7.96	1.34
Vandetanib	Multi-kinase	CHEMBL941	7.40	7.47	0.07
Neratinib	Irreversible	CHEMBL209386	8.70	5.56	3.14
Dacomitinib	Irreversible	CHEMBL3989959	8.82	7.54	1.28
Canertinib	Quinazoline Type I	CHEMBL3035	8.43	5.79	2.64

How to read this: Stratified on this table, the quinazoline subset (erlotinib, gefitinib, canertinib, neratinib) shows MAE ~3.1 pKd with compression to ~5.6–5.8; the non-quinazoline subset is closer (~1.3 pKd MAE). This gap is target/chemotype-specific — compare to the JAK2 table above, not to RL step count.

Generated libraries (Pathfinder exports)

These tables show RL-optimized candidates ranked by the same Stage 6 scorer. RL steps (200 vs 500 here) control how long generation searches — not how the affinity model was trained. Higher top pKd after 500 steps partly reflects more optimization time, not a different model. Do not compare raw top pKd across runs with different step counts as proof one target is “better.”

EGFR library · 200 RL steps

Default Pathfinder demo — 10 compounds from a 200-step generation run, batch-scored and ranked by predicted pKd on pocket 2ITK.

Rank	ID	pKd	QED	LogP	Lipinski	ChEMBL top sim	Confidence
1	GEN-16	8.18	0.92	1.82	yes	—	high
2	GEN-11	8.16	0.94	2.95	yes	—	high
3	GEN-18	8.09	0.92	1.97	yes	61%	high
4	GEN-20	8.09	0.92	2.37	yes	—	high
5	GEN-13	8.07	0.94	2.26	yes	—	high
6	GEN-12	8.06	0.94	3.06	yes	80%	high
7	GEN-15	7.86	0.94	2.44	yes	—	high
8	GEN-14	7.85	0.95	2.52	yes	—	high
9	GEN-19	7.80	0.94	3.85	yes	—	high
10	GEN-17	7.67	0.94	3.15	yes	—	medium

GEN-12 ChEMBL top hit: CHEMBL551595 (80% Tanimoto).

Spotlight: GEN-12 · EGFR hit #6

Example of orthogonal signals med chem teams expect — not a single opaque score.

EGFR · 200 RL steps · rank #6 Strong binder

GEN-12 · Gen-EGFR

Predicted pKd8.06

QED0.94

ChEMBL similarity80%

LipinskiCompliant

LogP3.06

ConfidenceHigh

Drug-like · MW 317 · bind 76.5% · verify analogs in ChEMBL yourself

JAK2 library · 500 RL steps

Separate 500-step generation run — same scorer, more RL optimization time. Top pKd ~9.25 vs ~8.18 on EGFR (200 steps) is not apples-to-apples; use external spot-checks above for cross-target comparison.

Rank	ID	pKd	QED	LogP	Lipinski	ChEMBL top sim	Confidence
1	GEN-1	9.25	0.94	2.56	yes	100%	high
2	GEN-2	9.15	0.95	3.31	yes	64%	high
3	GEN-3	9.14	0.94	2.17	yes	—	high
4	GEN-6	9.14	0.94	2.39	yes	—	high
5	GEN-7	9.12	0.94	3.14	yes	—	high
6	GEN-8	9.12	0.94	3.71	yes	90%	high
7	GEN-4	9.10	0.94	3.32	yes	62%	high
8	GEN-5	9.04	0.94	3.18	yes	90%	high
9	GEN-9	9.00	0.94	3.18	yes	100%	high
10	GEN-10	8.92	0.95	3.00	yes	76%	high

Top hit GEN-1 ChEMBL analog: CHEMBL175030 (100% Tanimoto). Scoring pocket: PDB 2b7a.

Within-library vs external benchmark

External spot-checks test the scorer on approved drugs. Generated libraries test rank-ordering within your series on a fixed pocket — the workflow VectaBind is built for. A higher top pKd after 500 RL steps does not mean JAK2 is a better target than EGFR; it means generation had more steps to climb the same scorer. Quinazoline EGFR drugs can look weak in absolute terms while your GEN series still ranks usefully relative to each other.

Signal provenance

Every column in the workbench traces to a named source — not a black-box number.

Signal	Source
pKd / bind probability	VectaBind Stage 6 EGNN + calibration
QED, MW, LogP, Lipinski	RDKit (in API + app)
ChEMBL similarity	ChEMBL REST API (browser-side lookup)
MPO composite	Weighted pKd + QED + Lipinski (app workflow)
3D pocket	Scoring: EGFR PDB 2ITK · JAK2 PDB 2b7a · Viewer: PDB 1M17 (EGFR display)
CSV export	Full scored row from workbench session

Model benchmark (secondary)

Internal PDBBind 2020 validation MAE is reported in Methods. For screening workflows, rank order within your library and ChEMBL-grounded analog checks are more actionable than a single global MAE.

Metric	Value	Notes
PDBBind 2020 val MAE	0.20 pKd	Model-selection split · see Methods
Intended use	Relative ranking	Compare compounds on the same target in your library