PolyPharm-AI – Prognica Labs

PolyPharm – AI

Every molecule tells a story.
We teach AI to read it.

A multi-task graph neural network trained on 3M+ compound-target interaction data points. Predicts binding affinities, selectivity profiles, and polypharmacology signatures with state-of-the-art accuracy – at the speed your discovery timelines demand.

3M+ training interactions
94.7% binding affinity accuracy
Graph Transformer architecture
ADMET endpoint prediction
500K+ compound library
Polypharmacology profiling
Off-target liability detection
ChEMBL-validated benchmarks
94.7% binding affinity accuracy
Multi-task learning
Selectivity scoring
GNN + Attention layers

3M+

Compound-target interaction training data points

97.4%

Binding affinity prediction accuracy (AUC-ROC)

1.2s

Median inference time per compound across all tasks

850+

Validated protein targets in the prediction library

17x

Faster than traditional HTS screening workflows

Neural Architecture

Three layers. One unified
molecular intelligence

Traditional docking simulations treat compounds as rigid 3D objects. PolyPharm-AI treats them the way biology does. Its multi-task architecture simultaneously predicts binding affinity, selectivity, off-target liabilities, and ADMET properties from a single SMILES input, giving medicinal chemists a complete molecular portrait in under one second.

PolyPharm-AI is a three-layer learned architecture where each layer’s output feeds the next, creating a cascaded representation that captures molecular structure, binding geometry, and clinical-stage properties simultaneously.

Layer 01

Graph Transformer — Molecular Representation

The compound enters the model as a molecular graph: atoms as nodes (atomic number, hybridisation, formal charge, chirality, aromaticity), bonds as edges (bond type, conjugation, ring membership, stereo configuration). A 12-head Graph Transformer with residual connections and layer normalisation learns a rich embedding for each atom in the context of the full molecular environment. Unlike fingerprint-based methods, this representation captures long-range electronic relationships across the molecule.

Graph Transformer

12-head Attention

Molecular Graph Input

Atom Embeddings

SMILES / InChI

Layer 02

Attention-Based Binding Site Prediction

The molecular embedding is cross-attended against protein pocket representations derived from 3D structural data (PDB) and predicted structures. A binding site attention module learns which atoms in the compound interact with which residues in the target pocket, producing per-atom interaction scores that serve as both a prediction output and an interpretability signal. Trained on 3M+ experimentally validated compound-target pairs from ChEMBL, BindingDB, and PDBbind.

Cross-Attention

AlphaFold2 Structures

PDBbind

Binding Affinity (pKd/pKi)

Interpretable Scores

Layer 03

Multi-Task ADMET Property Forecasting

A shared molecular embedding branches into 24 parallel task-specific heads, each predicting a different ADMET property, from aqueous solubility and CYP450 inhibition to hERG liability and BBB penetration. Multi-task learning exploits correlated property signals: a compound’s metabolic stability predictions improve its solubility estimates and vice versa. Each head is calibrated against held-out experimental data, producing probability distributions rather than point estimates, quantifying prediction uncertainty.

24 ADMET Tasks

Multi-Task Learning

Uncertainty Quantification

CYP450

hERG

BBB

Benchmark Performance

Binding Affinity (AUC-ROC) - 0.947

Selectivity Index (Pearson r) - 0.91

ADMET Solubility (RMSE) - 0.62 log

hERG Liability (AUROC) - 0.93

CYP3A4 Inhibition (AUROC) - 0.90

Off-Target Prediction (F1) - 0.87

Inference Latency (p50) < 1.2s

Validating Against

ChEMBL v33

PDBbind 2023

BindingDB

MoleculeNet

TDC Benchmarks

Prediction Modules

Six things PolyPharm-AI
tells you about every
molecule

Predicts pKd, pKi, and pIC50 values for a compound against any target in the 850+ protein library. Input a SMILES string and a UniProt ID — get a calibrated binding affinity estimate with a 95% confidence interval. Validated against PDBbind 2023 with Pearson r = 0.91 on the core test set.

Generates a full selectivity index across paralog proteins and off-target kinases. A compound designed for EGFR may also hit ERBB2, ERBB3, and 14 other kinases — PolyPharm-AI™ maps the entire selectivity landscape before you run a single assay, prioritising compounds that hit your target with minimum collateral activity.

Computes a polypharmacology fingerprint — a ranked vector of all significant predicted off-target interactions. For repurposing, this is the core signal: the “accidental” target profile of a drug often explains both its side effects and its potential in a new indication. DrugNet™ integration maps each off-target interaction to disease associations.

24 simultaneous ADMET predictions: aqueous solubility, lipophilicity (logP/logD), plasma protein binding, metabolic stability (microsomal), CYP1A2/2C9/2C19/2D6/3A4 inhibition, P-glycoprotein substrate, hERG liability, BBB penetration, oral bioavailability, half-life, and human intestinal absorption. Each with uncertainty estimates.

Screens compounds against PAINS (Pan-Assay Interference) filters, REOS alerts, structural toxicophores, and a proprietary clinical toxicity model trained on post-market withdrawal data. Flags compounds likely to cause hepatotoxicity, cardiotoxicity, or mutagenicity – before any animal studies are conducted.

Given a lead compound, PolyPharm-AI generates structure-activity relationship (SAR) maps and suggests 10–50 synthetic analogues predicted to improve potency, selectivity, or metabolic stability – while staying within synthetic accessibility constraints. Each analogue is pre-scored across all prediction modules.

Want to know how PolyPharm-AI helps you get a complete profile for your first molecule?