Ref F2
Computational Service
Confidence Reported
Scoped After Feasibility Review
Service Overview
Reinforcement learning suits the optimisation problem that follows hit finding: a starting molecule exists, several properties must improve at once, and the trade-offs between them are real. We define a multi-objective scoring function with you — potency proxy, selectivity, ADMET endpoints, synthetic accessibility, structural constraints — then run iterative RL-guided optimisation that explores chemistry around your series. The critical work is designing the reward so the agent optimises the molecule you want rather than the metric you happened to write down.
Key Features & Advantages
Reward Design As The Core Task
Scoring components, weights and hard constraints are designed and pressure-tested with you, because a careless reward produces high-scoring molecules no chemist would make.
Multi-Objective Trade-Offs Made Visible
Results are delivered as a Pareto front rather than a single ranking, so the potency-versus-property compromises are explicit and yours to choose between.
Chemistry Constrained By Construction
Reaction-aware and scaffold-constrained modes keep proposals within accessible chemistry and, where required, within a fixed core.
Reward Hacking Actively Policed
Output is inspected for the pathological structures RL is known to find when exploiting a scoring function, and the reward is revised when such behaviour appears.
Where This Service Is Used
1Optimising a hit series against several properties simultaneously
2Rescuing a potent series that fails on solubility, permeability or metabolic stability
3Generating focused libraries around a confirmed scaffold
4Exploring substitution patterns systematically before committing synthesis resource
5Improving predicted selectivity against a specific off-target while retaining on-target activity
Methods and Techniques Applied
1Scoring function construction from predictive models, docking, property predictors and hard filters
2Prior model selection or fine-tuning on your chemical series where sufficient data exists
3Policy optimisation with diversity and similarity regularisation to prevent mode collapse
4Scaffold-constrained and reaction-aware generation modes where the chemistry must be preserved
5Pareto analysis across objectives with trade-off visualisation
6Independent confirmation of selected molecules by docking, rescoring and property prediction
What We Need From You
1Starting molecules or scaffold, with measured data where available
2The properties to optimise and, importantly, their relative priority
3Hard constraints: substructures to retain or exclude, property limits, accessible chemistry
4Any internal predictive models you wish incorporated into the reward function
What You Receive
1Optimised molecule set with predicted profiles across every objective
2Pareto front visualisation showing the trade-off structure of the design space
3Scoring function documentation with components, weights and validation
4Confirmation results from docking and property prediction for the selected molecules
5Optimisation trajectory analysis showing how the chemistry evolved and where it saturated
What We Compute and Report
1Multi-objective score per molecule with the contribution of each objective shown separately
2Pareto front across your objectives, making the trade-offs explicit rather than hidden in a total
3Optimisation trajectory statistics showing where progress plateaued
4Chemical diversity and similarity-to-parent metrics across the optimised set
5Independent confirmation scores for the selected molecules
Questions We Are Asked Before Scoping
What data do we need to start?
▸A defined target or predictive model and a starting scaffold are enough. Measured series data improves the reward function substantially, and without it the optimisation leans more heavily on generic predictors.
Will the molecules be synthesisable?
▸Reaction-aware generation and accessibility scoring keep proposals within reachable chemistry, and the final set is route-assessed. A chemist will still reject some, which is expected.
Can RL optimise against experimental results?
▸Yes, and that is its strongest use. Where you can return assay data between rounds, the model is retrained on it, which is considerably more reliable than optimising against predictions alone.
Interpretation and Limitations
Computational predictions prioritise hypotheses; they do not independently establish binding, biological activity, therapeutic efficacy, safety or clinical validity. Performance depends on input quality, target class, the structural evidence available and the applicability domain of the selected methods. Scores used for ranking are comparative devices and should not be read as measured quantities. All deliverables are for research use only and are not intended for diagnostic or therapeutic use.
Is This the Right Service for You?
Choose this when
A hit series must improve on several properties simultaneously
Single-property fixes keep costing activity and the trade-offs need making explicit
Assay data can be returned between rounds so the model can be retrained on it
Consider another service when
No chemical matter exists yet for the pocket
Generative Molecular Design The optimisation question is really about metabolism or permeability
ADMET & Drug-Likeness Prediction What Goes Wrong, and How We Handle It
These are the failure modes that produce confident-looking results which do not survive the next experiment. Each one is designed against rather than discovered afterwards.
The agent exploiting the scoring function rather than solving the problem
Output is inspected for the pathological structures this failure produces, and the reward is revised when they appear.
A single weighted score hiding the trade-offs
Results are delivered as a Pareto front so the compromise between potency and properties is chosen by you, not by a weighting.
Optimisation converging on one narrow chemotype
Diversity regularisation is applied and chemical diversity is reported across the optimised set.
A Worked Example
Rescuing potency-driven compounds that failed on solubility
The situation
The series had strong activity, negligible aqueous solubility and a history of losing potency whenever polarity was added.
What the work involved
A reward combining potency proxy, solubility, permeability and accessibility was constructed and pressure-tested, then optimisation was run under scaffold constraints for twelve rounds.
What came back
A Pareto front showing where solubility gains started costing predicted potency, with 40 molecules on the front and the point of diminishing return identified explicitly.
Illustrative of the shape this work takes. Details are generalised, no client or target is identified, and scope and outcome vary with the system in question.
What Published Work Reports
The results below were published by independent groups and are cited to show what these computational approaches have achieved in the literature. They are not ICDMO project results, and outcomes vary substantially between targets, datasets and programmes. We do not promise comparable results for any particular project.
A generative-AI-designed inhibitor reported positive Phase 2a lung function data
A programme in which both the target and the molecule were identified using AI methods reached a randomised Phase 2a trial in idiopathic pulmonary fibrosis. Across 71 patients treated for 12 weeks, the 60 mg once-daily arm showed a mean forced vital capacity change of +98.4 mL against −20.3 mL in the placebo arm, with a manageable safety profile. It remains an early-phase result in a small trial, but it is the first published clinical proof of concept for an AI-discovered target and AI-designed compound.
Nature Medicine, 2025 ↗Service Workflow
1
Scientific Consultation
We discuss the objective, the decision it supports and the data you hold. Nothing is quoted before we understand what the result has to answer.
2
Feasibility Review
We check that the data, target class and available structural evidence can support the proposed approach, and say plainly where they cannot.
3
Technical Proposal
Scope, methods, deliverables, timeline and price are confirmed in writing, together with the limitations we expect to report.
4
Computation and QC
The analysis runs on pinned tool versions in a containerised environment, with quality checkpoints and method comparison where that is informative.
5
Report and Follow-Up
You receive the ranked results, the documented methods, the limitations and a proposed validation plan, plus support in interpreting them.
Service Process
1
Online Consultation
›
2
Solution Matching
›
3
Service Contract
›
4
Project Execution
›
5
Data Delivery
Specifications & Turnaround
Note: ICDMO delivers computational (in silico) services. Results prioritise hypotheses for testing; they do not by themselves establish binding, biological activity, efficacy or safety. All deliverables are for research use only and are not intended for diagnostic or therapeutic use. Downstream experimental work is scoped separately and executed through qualified partners.
Get a Custom Quote
Our scientific team responds within 24 hours with a detailed technical proposal and pricing tailored to your project.
Contact Us Online ConsultationStandard Deliverables
✓Sequence-verified final product
✓Full QC documentation
✓Certificate of Analysis (CoA)
✓Electronic data files
✓Project completion report
Why ICDMO?
Computational CDMO Focus
Reproducible, Versioned Pipelines
Confidence & Limitations Reported
Experiment-Aware Recommendations
Response Within One Business Day