RDKitSketcher
Draw and edit structures in the browser, then export to any common chemical format.
KetcherThe whole discovery cascade,
in one browser tab.
The technology
Artificial intelligence and physics,
in the same session.
Machine learning · narrows the search12 of the 31 modules
Structure read as a graph
SilicoScreen · SilicoSAR
Poses generated, not searched
SilicoDock · DiffuseBind
Sequence and peptide models
SilicoPeptify · Mutation Inducer
Generative design and ADMET
DeNovo-Generator · PharmK-AI
One chemical representation, one session, no export step
All-atom molecular dynamics
SilicoMD
End-point binding free energy
SilicoGBSA
Alchemical free energy
SilicoFEP
Density functional theory
DFT
Physics-based · tests what survives7 of the 31 modules
The remaining 12 modules are the preparation, visualisation and analysis steps that both sides share, from structure standardisation and pocket detection through to interaction profiling.
Graph neural networks
Structure read directly, not through descriptors
SilicoScreen passes each molecule through stacked graph isomorphism layers and a multi head graph attention layer, giving the expressivity to separate close analogues while triaging millions of compounds. SilicoSAR adds potency regression with nine feature selection routes and an applicability domain score on every prediction.
Diffusion models
Poses generated, not searched for
SilicoDock is our own SE(3) equivariant diffusion engine. Ligand poses are reverse diffused from Gaussian noise anchored at binding site grid points, with rigid body steric guidance and fragment aware torsion updates at each step, then ranked by a surface based mixture density scoring function. DiffuseBind brings the same class of method through DiffDock.
Transformers and language models
Sequence models across protein and peptide work
A protein language model predicts the stability change on point mutation through contextual sequence embeddings. SilicoPeptify embeds each residue from a 1875 dimensional descriptor vector and generates sequences under property, motif or complementarity constraints. Receptor specific transformer ensembles rescore poses for both protein and nucleic acid complexes.
Physics based simulation
The methods the field already trusts
All atom molecular dynamics with particle mesh Ewald electrostatics, end point MM-GBSA and MM-PBSA binding free energy, alchemical free energy perturbation across lambda windows, and Kohn Sham density functional theory for frontier orbitals and reactivity descriptors. Each is set up from a single form with defaults drawn from accepted practice.
End to end
A campaign should not be an exercise in file conversion.
A modern discovery programme passes through generative design, activity classification, pharmacokinetic filtering, several docking protocols, replicate molecular dynamics, free energy estimation and quantum chemical work. Each stage is well served by mature software. Chaining them is where the time goes.
01
Build and prepare
- Sketcher
- MolConv
- SilicoCG
- PocketFinder
02
Generate and triage
- DeNovo-Generator
- SilicoScreen
- SilicoSAR
- PharmK-AI
03
Dock and score
- LigandLock
- SilicoDock
- DiffuseBind
- Rescoring
04
Simulate and verify
- SilicoMD
- SilicoGBSA
- SilicoFEP
- DFT
05
Nominate
- BindingInspect
- Post-Docking
- Machine learning
- Physics-based
- Preparation and analysis
- No export step between stages
One chemical representation
All thirty one modules share the same internal data model. Nothing is exported, reformatted or prepared again at a stage boundary.
Parameters that are computed, not typed
Grid boxes come from a co crystal centroid or a detected pocket. Preparation happens inside the module rather than before it.
Every run reproducible
Each job stores its full parameter set, engine versions and logs beside its results, so any run can be inspected, repeated or shared.
At a glance
The periodic table of discovery.
All 31 modules on one screen, grouped by functional area and ordered the way a campaign runs. Hover over a tile for a one-line description.
How to read a tileDeveloped in-house
Number module, by area
Symbol short name used in workflows
Name
Bottom line engine it runs, or developed in-house
Specialist seriesSingle-module areas that extend SilicoX plore beyond small molecules.
- Preparation, inspection and modelling 8
- Docking and pose evaluation 9
- Ligand-based and machine-learning prediction 7
- Simulation, free energy and electronic structure 4
- Bioinformatics 1
- Peptide science 1
- Crystal structure prediction 1
Module directory
Every module, and what it is for.
AImachine-learning modelPhysicsphysics-based method
Preparation, inspection and modelling
Everything a structure needs before the science starts, from drawing and standardisation through to pockets and homology models.
RDKit
RDKitMolConv
Standardise every structure to canonical SMILES and InChI so one molecule stays one record.
RDKit
RDKit · ETKDGSilicoCG
Generate and rank 3D conformers from torsional knowledge, ready for docking and shape comparison.
ETKDG
RDKitSilicoPhysChem
Return molecular weight, LogP, TPSA and drug likeness descriptors straight from the input structure.
In house
PDB 1M17MolVisio
Inspect proteins and complexes in the browser with GPU accelerated cartoon, surface and stick views.
NGL
PDB 4EYLPocketFinder
Detect and rank binding cavities by volume, depth and druggability before you commit to a site.
fpocket
PDB 4EYLProtein Modelling
Build a three dimensional model from sequence and template, scored for quality before you use it.
MODELLER
PDB 3ERTMutation Inducer
Predict the stability change on point mutation and see which substitutions a protein will tolerate.
THPLMDocking and pose evaluation
Five docking engines under one interface, with protocol validation, rescoring and interaction analysis built around them.
PDB 1HSGLigandLock
Screen libraries of any size with Monte Carlo search and empirical scoring, from a grid box built for you.
AutoDock Vina
PDB 4TZKBINA
Pair empirical scoring with a convolutional network so pose selection reads the shape of the pocket.
GNINA
PDB 3ERTAntDock
Sample position, orientation and torsion angles together, scoring every pose as it is generated.
PLANTS
PDB 1M17DiffuseBind
Generate poses by reverse diffusion and rank them with a confidence model rather than a score function.
DiffDock
PDB 4EYLSilicoDock
Our own diffusion engine anchors sampling at pocket grid points and ranks poses by surface contact density.
In house
PDB 4TZKSelfDocking
Validate a protocol before you trust it by redocking the crystallographic ligand and reporting RMSD.
In house
PDB 1TUPRescoring
Rescore docked poses for protein and nucleic acid complexes using receptor specific transformer models.
In house
PDB 1HSGPost-Docking
Rank poses by binding energy and carry the most favourable one forward automatically.
In house
PDB 1M17BindingInspect
Identify every hydrogen bond, hydrophobic contact, salt bridge and metal coordination in the complex.
PLIPLigand based and machine learning prediction
Similarity, pharmacophores, potency, toxicity, ADMET and generative design, each served by its own trained model.
RDKitSimSearch
Search any library by fingerprint similarity, or invert it to find novelty and balanced subsets.
In house
RDKitPharmFrag
Compare pharmacophoric features against a reference so candidates keep the pattern that matters.
In house
RDKitSilicoSAR
Predict potency from structure with nine feature selection routes and an applicability domain on every value.
XGBoost, LightGBM
RDKitSilicoScreen
Triage millions of molecules with a graph network that reads structure directly, with no descriptors required.
In house GNN
RDKitSilicoTox
Flag likely toxicity early with a deep network trained on molecular property data.
In house
RDKitPharmK-AI
Run a full ADMET panel across absorption, distribution, metabolism, excretion and toxicity for large sets.
ADMET-AI
RDKit · BRICSDeNovo-Generator
Generate novel analogues from a single seed using a network trained on a million ChEMBL structures.
DeLA-DrugSimulation, free energy and electronic structure
Dynamics, end point and alchemical free energy and density functional theory, all set up from a single form.
PDB 1HSGSilicoMD
Run all atom dynamics without Linux and get RMSD, RMSF, radius of gyration, SASA and PCA back as finished plots.
GROMACS
RDKitSilicoFEP
Estimate relative binding free energy across lambda windows, with per window and cumulative profiles.
GROMACS
PDB 3ERTSilicoGBSA
Decompose binding free energy from your trajectory by MM-GBSA and MM-PBSA in a single pass.
gmx_MMPBSA
PySCF · B3LYP/6-31G*DFT
Compute frontier orbitals, electrostatic potential maps and reactivity descriptors from first principles.
Psi4, LibXCSpecialist series
Single module areas that reach beyond the small molecule space, from sequence quality control at the front of a campaign to peptide design and solid form prediction.
ENA ERR022075SilicoSeqCheck
Assess raw sequencing quality by Phred score, k-mer frequency, GC content and duplication.
FastQC
PDB 1L2YSilicoPeptify
Design peptides residue by residue against target properties, motifs or complementarity to a surface.
In house
COD 1548348SilicoCSP
Predict crystal packing with symmetry constrained generation scored by a materials graph network.
PyXtal, M3GNet
Boltz models, integrated through our collaboration
BoltzGenBoltz-2BoltzMol-1BoltzProt-1Boltz API
Built in house
Four engines we wrote because
nothing suitable existed.
Where a well validated implementation was available we integrated it. Where none was, we built our own and published the method.
Docking
SilicoDock
An SE(3) equivariant diffusion sampler that generates poses by reverse diffusion from noise anchored at binding site grid points, ranked by a surface based mixture density scoring function.
Classification
SilicoScreen
A hybrid graph isomorphism and graph attention network that reads molecular structure directly, giving large scale triage without descriptor engineering.
Regression
SilicoSAR
Bioactivity prediction with nine feature selection methods, five model families and an applicability domain score that flags predictions made outside the training space.
Peptides
SilicoPeptify
A four layer transformer encoder over physicochemical embeddings that generates sequences residue by residue under property, motif or complementarity constraints.
Validation
Measured against held-out data, not against itself.
Two things decide whether a technology of this kind is worth using: whether its models hold up on compounds they were never trained on, and whether its docking reproduces a structure that is already known. Both are published.
SilicoSAR held its accuracy on compounds it had never seen
Potency regression on pIC50. The external set was held out entirely, and performance tracks the internal test set.
Training set
n = 5,318
RMSE 0.28 log units
Test set
n = 1,400
RMSE 0.66 log units
External set
n = 759
RMSE 0.65 log units
Bars show R² and Pearson r, where higher is better. RMSE is in log units, where lower is better. Islam, Bhowmick & Chikhale, Comput Biol Chem 126, 109467.
View as a table
| Data split | Compounds | R² | RMSE | Pearson r |
|---|---|---|---|---|
| Training | 5,318 | 0.92 | 0.28 | 0.96 |
| Internal test | 1,400 | 0.58 | 0.66 | 0.77 |
| External validation | 759 | 0.61 | 0.65 | 0.78 |
Docking protocol validation
1.43 Å self-docking RMSD
Before a protocol is trusted, SelfDocking redocks the crystallographic ligand into its own structure and reports the deviation from the experimental pose. A deviation of 2.0 Å or less is conventionally taken as a reproduced pose.
Applied screening
Two published campaigns
EGFR: 36,000 natural products reduced to 8 candidates stable over 200 ns. NDM-1, a resistance enzyme with no clinical inhibitor: 752,436 generated structures reduced to 4. Each ran end to end inside one session.
36,000 natural products to 8 candidates, in one session
EGFR case study. Each stage ran in a named module, and no file was exported between them.
- NPAtlas natural products screened36,000SilicoSAR
- Predicted active, pIC50 above 616,000SilicoSAR
- Passed ADMET filtering2,035PharmK-AI
- Docked at 7.5 kcal/mol or better280LigandLock
- Taken into molecular dynamics20SilicoMD
- Stable over 200 ns of dynamics8SilicoGBSA, SilicoFEP
Bar lengths on a log scale. Islam, Bhowmick & Chikhale, Comput Biol Chem 126, 109467.
Peer reviewed
The methods are published, not just described.
The technology and its machine-learning core have each been through peer review, with the data and protocols available for inspection.
Journal of Molecular Graphics and Modelling
SilicoX plore: An integrated cloud platform coupling machine learning with physics-based modelling for end-to-end drug discovery, applied to the computational prioritisation of putative New Delhi metallo-β-lactamase-1 binders
The technology itself, the shared chemical representation across its modules, and the NDM-1 case study that narrowed 752,436 generated structures to four candidates carried to free energy.
Volume 149 (2026), article 109566 · Mangal, Shinde, Chikhale, Bhowmick & Islam
Computational Biology and Chemistry
SilicoSAR: an integrated machine-learning QSAR module within the SilicoX plore platform for QSAR-based virtual screening, demonstrated by the discovery of natural-product EGFR inhibitors
Potency prediction from structure, benchmarked on training, test and external sets, and applied to a screen of 36,000 natural products.
Volume 126, article 109467 · Islam, Bhowmick & Chikhale
Getting started
What you need to run it.
A browser
Your target and your data
A PDB entry, a sequence, a SMILES file or a library. Preparation is part of the task rather than a prerequisite for it.
Deployment that suits you
Hosted by us, deployed in your own cloud, or installed on your own servers for work that cannot leave the building.
Get started