Autonomous Drug Discovery & R&D Acceleration Agents

This module frames the whole course by mapping the discovery pipeline against where programmes actually fail. AI has compressed the earliest and cheapest stages while Phase II attrition — driven largely by the target being wrong — remains untouched. It sets out the public clinical record of AI-originated candidates honestly, and establishes target confidence as the metric that matters.

• 1.Where the Money Goes: Attrition, Predictive Validity and the Shape of the Pipeline15m
• 2.Reading the Clinical Record: What AI-Originated Candidates Have and Have Not Shown15m
• 3.Lab: Build a Programme Attrition Map and Target Confidence Dossier25m

Module 2 sets out what structure prediction and generative design actually deliver: what AlphaFold 2 and 3 predict and refuse to predict, how their licence terms constrain commercial teams, how the RFdiffusion family designs proteins, and how generative models navigate ligand space. Synthesisability is treated as the binding constraint, since a molecule no chemist can make is not a hit.

• 1.Reading AlphaFold Honestly: What AF2 and AF3 Predict, and Who Is Licensed to Use Them10m
• 2.From RFdiffusion Backbones to Ligand Space: Generation and the Synthesisability Constraint10m
• 3.Lab: Build a Synthesisability-Gated Design Triage Protocol for a Generative Campaign25m

Virtual screening agents are judged on enrichment, not accuracy: docking scores are weak predictors of binding affinity and ADMET models hold only inside narrow applicability domains. This module builds a screening cascade with numeric go/no-go criteria at every gate, and a register of the conclusions the agent is forbidden to draw.

• 1.Why Docking Scores Are Not Binding Affinities: Designing for Enrichment10m
• 2.ADMET Applicability Domains, PAINS Alerts and the hERG Selectivity Problem10m
• 3.Lab: Build a Screening Cascade and a Negative Authority Register20m

Most published multi-omics signatures do not replicate, and the causes are mundane: small n against enormous p, leakage through preprocessing, and site or platform effects mistaken for biology. This module sets the noise characteristics of genomics, transcriptomics, proteomics and metabolomics apart from one another, separates biomarker qualification from discovery, and produces the confounding audit the later modules build against.

• 1.Small n, Vast p: Why Multi-Omics Signatures Fail to Replicate15m
• 2.Four Omics Layers, Four Noise Profiles - and the Qualification Bar Beyond Discovery15m
• 3.Lab: Build a Confounding Audit and Context-of-Use Dossier for a Biomarker Agent25m

Swanson showed in 1986 that a discovery can sit unread across two literatures that never cite one another. This module builds the modern version of that idea: entity recognition, relation extraction and a provenance-carrying knowledge graph over the biomedical record. It is equally concerned with what the corpus is not, because retractions, preprints and publication bias make it evidence with a shape rather than neutral ground.

• 1.From Swanson's ABC Model to Agentic Search: What Counts as a Discovery15m
• 2.Entities, Relations and Provenance: Building a Corpus You Can Reason Over15m
• 3.Lab: Building a Hypothesis Dossier with a Corpus Bias Register25m

Physical experimentation is where an agent's decisions stop being reversible: reagent is consumed, instruments are occupied and samples are destroyed. This module covers closed-loop design-make-test-analyse architecture, Bayesian and active-learning experiment selection, LIMS and ELN integration, and the safety interlocks and ALCOA+ data-integrity controls that autonomy demands.

• 1.From Batch Screening to Closed Loop: DMTA Architecture and the Orchestration Layer10m
• 2.Choosing the Next Experiment: Bayesian Optimisation, Active Learning and Amplified Error15m
• 3.Lab: Author a Closed-Loop Campaign Control Document with Safety Interlocks and an ALCOA+ Audit Trail25m

Protocol design is where an agent's output meets a regulator who will read it closely. This module fixes the estimand as the object being optimised, sets adaptive and platform designs against what ICH E20 and the current FDA and EU texts actually say, and treats site selection as a prediction problem whose accuracy and fairness pull in opposite directions.

• 1.Estimands Before Endpoints: Writing a Protocol an ICH E9(R1) Reviewer Will Accept10m
• 2.Adaptive and Platform Designs Under ICH E20: Simulation, Sample Size and Blinding Firewalls10m
• 3.Lab: A Site Selection Scorecard with a Representativeness Constraint25m

Recruitment is the commonest cause of trial delay, and most of the failure is operational rather than analytical. This module separates what an agent may do with the record from what it may do to a patient, fixes the consent boundary under ICH E6(R3) and Regulation (EU) 536/2014, and treats dropout prediction and reachability bias as design problems rather than reporting ones.

• 1.Pre-Screening from the Record Versus Outreach: Cohort Discovery, Screen Failure and Reachability Bias15m
• 2.The Consent Boundary and Retention Agents: Predicting Dropout Without Coercing the Participant15m
• 3.Lab: Building a Recruitment Agent Action Register and Reachability Audit for an Adaptive Trial25m

Post-market safety data is the largest and least trustworthy corpus in pharma, and the temptation to automate it is proportionate to its volume. This module separates what spontaneous reporting can support from what it cannot, treats a disproportionality signal as a hypothesis rather than a finding, and shows why automating ICSR intake without automating quality control simply relocates the bottleneck onto the reviewer.

• 1.Spontaneous Reports as Evidence: Under-Reporting, Confounding by Indication and the Missing Denominator15m
• 2.Disproportionality, ICSR Case Processing and the Reviewer Bottleneck Nobody Budgets For15m
• 3.Lab: Build a Signal-to-Study Dossier Using Target Trial Emulation25m

A dossier is a legal representation, not a document set. This module treats the eCTD as a machine target with honest limits, sets out the risk-based credibility assessment that FDA and EMA both apply to AI-derived evidence, and is direct that establishing credibility for a defined context of use is the sponsor's burden and no one else's.

• 1.What the eCTD Actually Is: Lifecycle Operations, Granularity and the Limits of Assembly Automation15m
• 2.Credibility for a Context of Use: FDA's Seven Steps, the EMA Reflection Paper and the January 2026 Joint Principles15m
• 3.Lab: Building a Context-of-Use Credibility Dossier for an AI-Derived Endpoint25m

Repurposing inverts the usual difficulty: the science gets easier and the economics get worse, because a generic molecule carries no exclusivity to fund its own trial. This module covers knowledge graph construction and embedding-based link prediction, signature reversion through the Connectivity Map, and the discipline separating a plausible edge from a testable hypothesis.

• 1.Constructing a Repurposing Graph: Ontology Discipline, Metapaths and Embedding-Based Link Prediction15m
• 2.From Plausible Edge to Testable Hypothesis: Signature Reversion and the Lesson of the COVID Wave15m
• 3.Lab: A Repurposing Candidate Dossier with a Method-of-Use and Exclusivity Gate25m

Regulators rely on computational toxicology in a much narrower band than vendors imply, and ICH M7(R2) marks its edge: two complementary (Q)SAR methodologies can discharge a genotoxicity test, while nothing comparable exists for organ toxicity. This module builds the applicability-domain gate, the read-across argument and the evidence dossier that decide whether a prediction may be used at all.

• 1.ICH M7(R2): The One Place a Computational Prediction Discharges a Test15m
• 2.Applicability Domain, Read-Across and Adverse Outcome Pathways as Admissibility Tests15m
• 3.Lab: Build an In Silico Safety Assessment Dossier and Domain Gate Register25m

Manufacturing is where an agent's decisions acquire physical and commercial weight at the same moment: a forecast becomes a purchase order, a schedule change, an allocation between markets. This module covers perishable and intermittent demand forecasting, API concentration risk, serialisation obligations under DSCSA, the EU FMD, Tatmeen and SFDA RSD, and the batch release boundary an agent may support but never cross.

• 1.Forecasting Perishable Demand: Intermittent Series, Shelf Life and Asymmetric Cost10m
• 2.Single-Source APIs, Serialisation Data and Shortage as a Public-Health Signal15m
• 3.Lab: Author a Supply Chain Agent Authority Matrix with a GMP Batch Release Boundary25m

Chemical patent claims are Markush sets, not sentences, so the compound that matters is often inside a claim that never names it. This module builds claim-level landscaping with families, priority dates and honest coverage maps, and fixes the professional boundary: landscaping and prior-art surfacing are agent work, a freedom-to-operate conclusion is a regulated legal opinion.

• 1.Markush Structures and Claim Types: Why a Chemical Patent Does Not Say What It Covers15m
• 2.Freedom to Operate, Inventorship and the Line an Agent Must Not Cross15m
• 3.Lab: Building a Target Landscape Dossier with a Claim-Scope Register and an FTO Escalation Gate20m

The capstone integrates every artefact this course produced into one platform design, and the design question is authority rather than capability: which agents may act, which may only propose, where the human gates sit, and how a claim made at lead optimisation stays traceable at submission. It ends in a refusal register, because what the platform will not automate is the part worth defending.

• 1.The Authority Model: Which Agents May Act, Which May Only Propose15m
• 2.Evidence Chain of Custody: Keeping a Lead-Optimisation Claim Traceable at Submission15m
• 3.Lab: Build the Platform Authority Map, Evidence Chain and Refusal Register25m