Public Roadmap · 2026 Q2–Q4

What We're Shipping, What We're Not

Every priority below is indexed by the licensing objection it kills, not by the technology stack it adopts. We publish this roadmap so due-diligence teams can pre-empt their own questions. Realistic durations. Named kill criteria. No 12-month fantasy timelines.

Single developerRTX 5090 · 32GBBudget: meaningful work only

Priority 0

Ship within 30 days

5 events. 3 new integrations + 2 release-ready artifacts. Every item externally verifiable.

P0-1

AiZynthFinder retrosynthesis integration

EXECUTED

Kills licensing objection

Generated compounds may not be synthesizable

Result (as of today)

TYK2 top-10: 30% fully solved, 100% partial routes (USPTO + ZINC, 2026-04-18)

Duration

Weekend

Scope: Small

P0-2

TDC + Polaris leaderboard submissions (Mondrian CP framing)

PARTIALLY SUBMITTED

Kills licensing objection

Zero externally-verified benchmarks

Result (as of today)

TDC ADMET Group 22/22 tasks SUBMITTED 2026-04-22 (via Google Form, AgentAI Labs reproducing ADMET-AI v2.0 + Mondrian CP, 5 independent runs). Top PR-AUC: HIA 0.999, Pgp 0.965, DILI 0.956, BBB 0.950, CYP3A4 0.931, AMES 0.930, hERG 0.911. TDC leaderboard URL pending review. Polaris submission still outstanding.

Duration

2 weeks

Scope: Small

P0-3

PLINDER + LP-PDBBind leakage-free re-evaluation

EXECUTED

Kills licensing objection

Pearson R=0.562 may depend on leakage

Result (as of today)

PLINDER 500 kinase scaffold-leakage audit (2026-04-20): random 5-fold CV Pearson R=0.95 vs Murcko scaffold-split R≈0.21 (gap 0.74) on Morgan FP + XGBoost baseline. Only 45 unique scaffolds across 500 systems, top 1 covers 67.6%. Random-CV is intra-dataset memorization; scaffold-split is defensible OOD number. Full re-evaluation at /data/plinder_scaffold_reval.json.

Duration

2–4 weeks

Scope: Small/Medium

P0-4

bioRxiv preprint v1.0 submission

DRAFT V0.1 COMPLETE

Kills licensing objection

No peer-review or citable artifact

Duration

2 weeks

Scope: Small

P0-5

CRO contract + Gate B schedule lock

RFQ PACKAGE READY

Kills licensing objection

in silico only, no wet-lab validation — #1 licensing objection

Duration

3–4 weeks

Scope: Medium

Status snapshot — 2026-04-19

P0-1 AiZynthFinder is live (3/10 TYK2 top compounds fully solved on ZINC stock, 100% with partial routes). P0-3 PLINDER infrastructure WIRED — 1.36M PLI index locally cached, 500 kinase subset (158 TYK2/JAK) extracted. P0-2, P0-4, P0-5 packages are all zero-coding releases waiting on human execution (submit to portals, send RFQs). Their licensing-credibility impact exceeds the combined effect of every other P0/P1 item.

Priority 1

Gate B Parallel Track · 2–9 months

Runs in parallel with CRO wet-lab cycle. Each has a named entry gate — nothing merges without criterion pass.

P1-6

Kinome-wide selectivity model

PLANNED

Kills licensing objection

Only TYK2 validated; kinase generalization unclear

Duration

6–9 months

Scope: Medium

Proprietary Moat
P1-7

MolCRAFT + DecompDiff ensemble addition

PLANNED

Kills licensing objection

Generators stuck at 2023 SOTA

Gate criterion (entry)

chains_ptm ≥ 0.85 AND scaffold diversity +20%

Duration

1–2 months

Scope: Medium

P1-8

OpenFE FEP pilot (TYK2, 10 ligands)

PLANNED

Kills licensing objection

No physics validation; Boltz-2 ΔΔG unbacked

Gate criterion (entry)

Gate B PASS + Boltz-2/FEP MAE < 1 kcal/mol

Duration

2–3 months

Scope: Small/Medium

Priority 2

Gate B PASS Conditional

Expansion items. Gated on wet-lab Gate B outcome — not started prematurely.

P2-9

Chai-1 + Protenix co-folding ensemble

GATE B CONDITIONAL

Kills licensing objection

AF3 commercial license blocker

Duration

1 month

Scope: Small

P2-10

Link-INVENT PROTAC extension (CDK4/6)

GATE B CONDITIONAL

Kills licensing objection

Small-molecule only, cannot access PROTAC market

Duration

1–2 months

Scope: Medium

Explicitly Out of Scope

What We're Not Doing

Discipline = saying no publicly. These are technically interesting but dilute the kinase-focused, Gate-B-driven licensing narrative.

R-1

ESM3-open / RFdiffusionAA / Chai-2 antibody

Separate business axis, narrative dilution

R-2

JUMP-CP / RxRx3-core phenomics

No interpretation infrastructure; no licensing hook

R-3

scGPT / Geneformer target triage

PandaOmics imitation; dilutes TYK2 Gate B focus

3-Month Forecast

Where we should be by 2026-07-18

Already Executed (2026-04-19)

  • ✓ AiZynthFinder integrated — TYK2 top-10: 30%, top-100: 32% solved
  • ✓ AiZynth pre-filter backtest (top-200): 95.5% synthesizable
  • ✓ PLINDER 500 kinase scaffold-leakage audit: random CV R=0.95 vs scaffold-split R≈0.21 (gap 0.74, 45 unique scaffolds) — see /data/plinder_scaffold_reval.json
  • ✓ 2-way Boltz-2 + Chai-1 co-folding ensemble (top-50)
  • ✓ 3-way consensus (MolForge + AEV-PLIG + Chai-1) re-ranking
  • ✓ SynFormer TYK2 analogs: 79 generated, 100% synth-guaranteed
  • ✓ TDC ADMET real inference: 22/22 tasks, 16,082 predictions
  • ✓ bioRxiv preprint v0.1 + 6 figures + LaTeX + Zenodo plan
  • ✓ TDC ADMET 22/22 tasks SUBMITTED 2026-04-22 (Google Form). Polaris + CRO 10-CRO RFQ still outstanding

Expected by 2026-07-19

  • ◐ Actual Polaris submission (human action; TDC already SUBMITTED)
  • ◐ bioRxiv DOI public after ORCID + Zenodo account setup
  • ◐ CRO Tier 1/2 contracts signed, Gate B wet-lab active
  • ◐ CP recalibration after 6-compound CRO verdicts
  • ◐ Kinome selectivity model (30% initial training)

If this state is achieved, MolForge becomes licensable — Series A or first partnership conversations begin. If not, the roadmap is wrong and we say so publicly.

Last reviewed: 2026-04-18 · Supersedes prior methodology roadmap

Partnership inquiries →