In-House Drug Discovery Automation
We build and operate our own computational drug discovery pipeline. Part of it is assembled from open scientific software; part of it we write from scratch. The result is a single connected line of work rather than a collection of disconnected scripts.
What the platform does
A project starts from a protein target and ends with analysed simulation data, without leaving our stack.
- 01
Target and binding site.We take a protein identifier, obtain or model a usable structure, predict candidate cavities, and rank them using biological evidence (sequence conservation, annotated functional and binding residues, domain context, known ligand occupancy) rather than cavity geometry alone.
- 02
Molecule design.We generate new ligands directly inside a defined pocket, decorate an existing core, swap scaffolds while preserving binding vectors, or grow from a fragment. Generation is structure-aware, not SMILES-blind enumeration.
- 03
Preparation and docking.Ligands are standardized, protonated and conformer-generated; receptors are repaired, protonated and reviewed. Docking runs across both classical scoring-function engines and deep-learning pose predictors, with independent rescoring.
- 04
Property filtering.Candidate molecules are scored for physicochemical profile, absorption, blood-brain barrier behaviour, metabolic liability, distribution, excretion and toxicity. We run two independent prediction engines in parallel and flag disagreements instead of averaging them away.
- 05
Ligand optimization.Aligned ligand series are converted into 3D field-based models (steric, electrostatic, hydrophobic and hydrogen-bonding contributions) through an end-to-end 3D-QSAR pipeline, from ligand preparation and alignment to validated predictive models, to relate structure to measured activity. Alongside these models, we use AI-based models for side-chain enumeration, core hopping and fragment growing to propose optimized analogues and guide the next design round.
- 06
Molecular dynamics.Selected complexes are parameterized, solvated, ionized, minimized, equilibrated and run to production. The same system supports mixed-solvent simulation for cosolvent site mapping and free-energy workflows.
- 07
Trajectory analysis.Full post-processing: structural stability and flexibility metrics, compactness and solvent exposure, dimensionality reduction and free-energy landscapes, interaction fingerprinting over time, secondary-structure evolution, pocket dynamics, binding-free-energy estimation.
Methods we develop ourselves
Alongside the integrated workflow, we are building our own methods where the available options fall short.
- Pocket prediction on dynamic structures.Most binding-site predictors see one static crystal structure. We predict and score pockets across simulation trajectories, so transient and cryptic sites that only open during motion are detected and ranked, with an explicit druggability assessment rather than a geometry score alone.
- Structure- and pocket-aware molecule generation.Generative models conditioned on the actual three-dimensional binding site, including ensembles of pocket conformations taken from dynamics rather than a single fixed snapshot, so proposed molecules are shaped by the pocket as it really behaves.
Built to run unattended
The pipeline is not only an interactive desktop environment. The same components run as long, checkpointed batch jobs: multi-million-compound library intake and filtering, GPU docking queues, and iterative generative design loops that run for hundreds of rounds. Jobs are supervised, resume cleanly after interruption or power loss, refuse duplicate processes, and write logs to dedicated storage. Scale in production runs is millions of compounds screened and hundreds of thousands docked per campaign.
Deployment
The whole stack installs as a portable bundle on a Linux workstation, with per-application environments and optional secure remote access, so a collaborating group can run the same pipeline on their own hardware.
What we offer
- Target assessment and binding-site identification, including cryptic and dynamics-dependent sites
- De novo and scaffold-based ligand design against a defined pocket
- Large-scale virtual screening with consensus property filtering
- 3D-QSAR model building on client compound series
- Molecular dynamics, cosolvent mapping and free-energy calculations
- Custom pipeline development and deployment on client infrastructure