Comprehensive and Unique Datasets of High Quality
Tens of millions of drug-like molecules and protein–ligand complexes generated through CADD methods
Deeply exploring experimental electron density (ED) data, pioneering its use in AI model training
18 million 2D small molecule library
6-million torsion energy data of common small molecule fragments
4 million 3D protein-ligand complex library
300,000 noncovalent interactions (NCIs) energy data of ligand-residue pairs from the Protein Data Bank (PDB)
We firstly use the ED data for modeling of noncovalent interactions (NCIs) for protein–ligand complexes.
Our methodology is based on the reduced electron density gradient (RDG) theory describing intermolecular NCIs by ED and its first derivative.
We identified about 200,000 NCIs from over 12,000 protein–ligand complexes for the development of the AI model.
Advanced Generative Model Architecture
By innovatively combining the Transformer framework with geometric deep learning, our model delivers molecules with validated 2D chemical structures and accurate 3D conformations.
Pretrain
Inference
Fine-tune
The model was pretrained on large-scale molecular datasets to learn fundamental chemical rules.
Experimental electron density (ED) grids are integrated during sampling to guide accurate molecule generation.
We adapted the model with pocket–ligand interaction data to capture binding dynamics and improve structural relevance.
Reinforcement Learning From Human Feedback (RLFHF)
· Reinforcement learning from chemistry cognition/computational results
· Reinforcement learning from wet-lab results
Industry-Oriented 3-Tiered Evaluation Framework
A stepwise evaluation strategy rigorously guides model iteration, avoiding the misleading phenomenon of non-drug-likeness in molecular computational metrics
Challenge
The gap between pure computational metrics and real activity  
Testing whether the model can generate active compounds reported in patents
Our Approach
Challenge
Undruggable-molecule contamination
Our Approach
Stage-by-stage evaluation so that the molecules with low drug-likeness are filtered out in the early stage and cannot contaminate metrics in the following stages
Validity of chemical structure & molecular conformation
Generating known active compounds
Generating novel active compounds
· New scaffold, old binding mode
· New scaffold, new binding mode
· High diversity

Generating patent reported active compounds by only knowing historical information
2D Structure Validity:
· QED, SAS
· Rarely observed fragments
3D Conformational Validity:
· Bond lengths, bond angles;
· Torsion energy
SBDD:
· Generating known active compounds in new pockets
LBDD:
·
Powered by comprehensive and unique data assets, advanced generative model, and an industry-oriented evaluation framework, our AI platform, IMOLX, accelerates drug discovery from pocket to candidate, empowering researchers to explore vast chemical space and unlock breakthrough innovations.
Follow Us
Address: 7 temasek boulevard,#12-07 suntec tower 1, singapore 038987
BD Inquiries: bd@imolx.com
Technical Support: support@imolx.com
Media Inquiries: press@imolx.com
General Inquiries: info@imolx.com