Open source · Free for academic use

Molnovi

End-to-end computational drug discovery
pipeline infrastructure.

An open, modular pipeline that takes a protein target and produces ranked drug candidates with full safety profiles. We automate the structural analysis, generation, and screening so you can focus on the science.

The Problem

You know this pain.

The computational workflow to find active molecules traditionally involves chaining together 5-6 separate open-source tools, each with different file formats, dependencies, and no shared infrastructure. Academic labs and biotechs do this manually with custom scripts. Things go wrong consistently.

Wasted Cycles: Teams repeat experiments that have already been tried because there is no structured record of what was run, what failed, and why. There is no memory across campaigns.
Silent Attrition: Molecules fail months later in synthesis or assay for reasons that were computationally detectable upfront — toxicity flags, metabolic liabilities — but nobody checked because the pipeline didn't enforce it.
Expert Bottleneck: Computational chemists spend 70%+ of their time on pipeline plumbing: format conversion, job management, results aggregation, debugging tool incompatibilities. The science drowns in engineering.

What We're Building

The End-to-End Pipeline

Input: a protein structure file (PDB). Output: a ranked shortlist of candidate drug molecules, each with a full safety and toxicity profile, ready for a medicinal chemist to review and decide what to synthesize.

Pocket Detection

ML-based identification of druggable binding sites on the protein surface.

P2Rank

Molecule Generation

Creates candidate molecules either from chemical building blocks or from scratch using a diffusion AI model conditioned on the 3D pocket shape.

TargetDiff, RDKit

Screening & ADMET

Filters for drug-likeness (Lipinski rules, QED, PAINS) then runs 104-property toxicity/pharmacokinetic profiling (hERG, AMES, etc).

ADMET-AI

Docking

Physics-based binding affinity estimation, scoring how well each molecule physically fits the pocket.

AutoDock Vina

Reporting

Produced ranked candidates with per-molecule scorecards, ADMET profiles, explicit limitations, and full provenance tracking.

SQLite

Validated Results

Proven on Real Targets

The pipeline independently finds molecules scoring in the same range as approved drugs, without any prior knowledge of those drugs.

EGFR (1M17)

Lung Cancer

Known Drug: Erlotinib

Pocket Accuracy

2.7 Å from drug site

82% residue overlap

Best Docking Score

-9.32 kcal/mol

BCR-ABL (2HYY)

Leukemia

Known Drug: Imatinib

Pocket Accuracy

2.7 Å from drug site

92% residue overlap

Best Docking Score

-12.59 kcal/mol

BRAF V600E (6P3D)

Melanoma

Known Drug: Ponatinib

Pocket Accuracy

3.1 Å from drug site

89% residue overlap

Best Docking Score

-11.20 kcal/mol

The pipeline independently finds molecules scoring in the same range as approved drugs, without any knowledge of those drugs. The diffusion model (TargetDiff) generates structurally novel molecules with excellent ligand efficiency (0.38-0.39) and drug-likeness (QED >0.88).

What Molnovi is NOT

It is not a drug. It produces ranked hypotheses for experimental validation. It accelerates the work of computational chemists — it does not replace their judgment.

We are honest about limitations: docking scores correlate with real binding at only r ≈ 0.4-0.6, ADMET predictions degrade on novel scaffolds, and every result must be validated experimentally.

Why It Matters

Different By Design

We are not competing with Schrödinger or Recursion. We are building accessible infrastructure for an underserved segment of a $2.5-5B market growing at 20-30% CAGR.

100% Open & Free: Open source and fully free for academic use. Researchers shouldn't be gated by $50K+ per-seat software licenses.
Full Integration: No more fragile pipelines. Stop chaining together six different tools with custom Python scripts that break on edge cases.
Comprehensive ADMET: Every generated candidate immediately receives a 104-property toxicological and pharmacokinetic profile via ADMET-AI.
Honest Limitations: Every result includes confidence context and caveats, not just a score. We tell you when the data is weak or uncertain.
Reproducible and auditable: Every decision logged to SQLite with full provenance tracking. No more lost parameters or untraceable runs.