Introduction to SEARCH-ML
SEARCH-ML bridges the gap between high-throughput chemical screening and precise structural biology. By integrating deep learning architectures with physics-based molecular modeling, it offers a robust pipeline for predicting binding affinities.
Bi-Directional Screening
Supports Forward Screening (one target vs. library) for hits, and Reverse Screening (one molecule vs. proteome) for target ID.
Stacked Ensemble
Aggregates XGBoost, CatBoost, LightGBM, and Random Forest models via a Ridge Regression meta-learner.
Data Preparation
Accuracy begins with high-quality input. SEARCH-ML parses .pdb files by distinguishing between protein backbones and reference ligands.
Input Data Format (PDB)
Ensure your PDB file contains ATOM records for the protein and HETATM records for the ligand.
ATOM 1 N PRO A 2 31.242 3.064 39.284 1.00 39.90 N
ATOM 2 CA PRO A 2 31.195 2.392 37.963 1.00 31.96 C
... (Protein Backbone Data) ...
TER 1632 GLN A 209
HETATM 3265 N VWW A 210 15.088 10.798 23.547 1.00 14.90 N
HETATM 3266 CA VWW A 210 15.010 9.987 24.792 1.00 20.92 C
HETATM 3267 C VWW A 210 16.115 8.924 24.830 1.00 21.55 CDetailed PDB structure information:
View PDB Format Guide
Important: HETATM tags are critical. The system scans these to detect co-crystallized ligands for site-directed docking. We recommend removing bulk water molecules before upload.
Forward Virtual Screening
Map chemical libraries against a single protein structure to identify high-affinity binders.
Target Upload & Visualization
Upload your PDB file. The integrated 3D viewer allows you to verify the structural integrity (secondary structures, chains) before screening.
Defining the Active Site
Select a reference ligand from the dropdown. SEARCH-ML calculates a 3D bounding box (red grid) around this ligand to restrict the docking search space.
Modes & Filters
Screening Modes
Submission & Tracking
Upon submission, you receive a unique Job ID. Results are persistent and can be retrieved later.
Reverse Virtual Screening
Identify potential biological targets for your small molecule by screening it against specific proteomes.
Molecule Input
Target Library Selection
Select from pre-processed Organism proteomes (e.g., Homo sapiens, M. tuberculosis) or broad Taxonomic groups.
Interpreting Results
PBE (kcal/mol)
Gibbs Free Energy (∆G). More negative = stronger binding. E.g., -10.5 is better than -7.2.
pKi / pKd
Negative log of dissociation. Higher is better. A value > 7 usually indicates a solid hit.
PBE (nM)
Nanomolar concentration required for effect. Lower is better. < 100nM is ideal.
Ensemble Confidence
We provide individual scores from XGBoost, CatBoost, RF, and LightGBM. When all four models agree (Consensus), the prediction reliability is significantly higher.