Design in silico

Target triage, pocket detection, structure prediction and generative design produce a prioritised, synthesisable set – not a raw enumeration.

Make at scale

Route feasibility is scored before a compound is selected, then handed to synthetic chemists in the same organisation for parallel synthesis.

Test in the lab

Biochemical and cell assays, DMPK and early tox run against the predictions that generated the molecules – no re-validation of someone else’s assay.

Learn and retrain

Every result – including the failures – becomes program-specific training signal. The model gets sharper on your chemical series, not on a benchmark.

Target Discovery

Multi-omics, literature and patent evidence assembled into a ranked target hypothesis with a druggability call and a route to validation. You see why a target scored where it did – pathway context, expression, genetic support and competitive crowding – before biology budget is committed.

Protein Structure Prediction

Folding, complex prediction and pocket detection for targets with thin or absent experimental structure – including conformational states and cryptic pockets. Predicted models are stress-tested with molecular dynamics before anything downstream is designed against them.

Small Molecule Design

Ultra-large virtual screening, structure-based and generative de novo design, scaffold hopping and multi-parameter lead optimisation. Synthetic accessibility and IP whitespace are scoring terms, not afterthoughts – so what comes out is a series a medicinal chemist would actually run.

ADMET & Druggability

Absorption, metabolism, clearance, permeability, hERG and hepatotoxicity predicted at the design stage, then confirmed against in-house DMPK. Prediction and measurement live under one roof, which is the only way model calibration ever improves.

Biologics & Antibody Design

Epitope prediction, CDR design, humanisation, affinity maturation in silico and developability triage for mAbs, bispecifics and ADC scaffolds. Sequence liabilities – aggregation, immunogenicity, poor expression – are flagged before a construct is ordered.

Oligonucleotide Therapeutics

siRNA and ASO sequence design with on-target potency modelling, off-target and seed-mismatch screening, chemical modification patterning and delivery-conjugate selection. The fastest-moving modality in the industry, and the one where sequence space is far too large to search by hand.

CRO Service Integration

The module that makes the other six mean something. Every prediction routes into synthesis, assay development, biology, DMPK, tox, CMC and GMP manufacturing under one contract – with results flowing back as training data instead of a PDF report.

Predictions get tested

A design proposed on Monday can be in synthesis the same week. No procurement cycle between the model and the bench.

Models learn your series

Proprietary assay data from your own program retrains the scoring, including the negative results that public datasets never contain.

Chemists in the loop

Medicinal chemists review what the model proposes. Undesirable and unmakeable chemotypes are removed by people who have made them before.

One chain to IND

Discovery, DMPK, tox, CMC and manufacturing sit in the same organisation – so the data package is built continuously, not reassembled.

Data that reflects real chemistry

Public bioactivity databases are skewed towards what worked and towards targets that were already tractable. Models trained only on that inherit the bias – they are confident about chemotypes the field has already exhausted and quiet about everything else.

XDrug.AI trains on curated public structure and activity data, and then on program data generated in our own laboratories: potency, selectivity, stability, permeability and the compounds that failed. Failure data is the scarcest and most valuable input in the field, and a platform without a lab attached to it does not have any.

Models chosen per task, not per fashion

Structure prediction, generative design, property prediction and sequence design are different problems and are not well served by one architecture. The platform combines deep learning models with physics-based methods – docking, free energy estimation, molecular dynamics – because the physics catches failure modes that a learned scoring function will confidently miss.

Where a prediction is uncertain, the platform says so. An unqualified score is worse than no score when it drives a synthesis decision.

Judgement at the output

Every module’s output passes a scientist before it becomes a plan. Computational chemists check pose plausibility and pocket occupancy; medicinal chemists check makeability, stability and whether a chemotype carries known liabilities; biologists check that the proposed assay reads out on the mechanism in question.

That review step is why the shortlists are short. A hundred designs a chemist will not make is not an output – it is a queue.

A biology hypothesis, no target

Start at target discovery. You get a ranked set of candidate targets with the evidence behind each, plus the validation experiments that would confirm or kill them.

A target with no structure

Start at structure prediction. Modelled conformations and pocket maps make structure-based design possible years before a crystal arrives – if one ever does.

Hits that will not optimise

Start at small molecule design plus ADMET. Scaffold hopping and multi-parameter optimisation attack potency and property cliffs at the same time instead of trading one for the other.

A new modality, no infrastructure

Start at biologics or oligonucleotide design. Sequence design plus the CRO chain gives you a modality capability without building one internally.

Can we use one module without the rest?

Yes. Each module runs as a standalone engagement – a structure prediction package, a virtual screening campaign, an siRNA design set. Most programs then extend into synthesis and assays once the first output lands.

Who owns the IP and the data?

You do. Designs, sequences, assay data and any resulting IP belong to your program. Program-specific models trained on your data are not applied to other clients’ work.

How is this different from licensing AI software?

A licence gives you predictions and leaves execution to you. XDrug.AI includes the execution – synthesis, assays and DMPK – so the loop closes and the models improve on your chemistry rather than on benchmarks.

Which modalities are supported?

Small molecules, biologics including mAbs, bispecifics and ADC scaffolds, and oligonucleotide therapeutics – siRNA and ASO. Emerging modalities are scoped case by case.

What do you need from us to start?

A target or a series, whatever structural and assay data exists, and the decision you are trying to reach. Where a target is undisclosed, we scope under CDA first.

How do engagements run commercially?

FTE, fee-for-service or blended team models, matched to program stage – the same contracting structures ChemPartner uses for integrated discovery programs.