[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7APTDZKBN44WNT4KP25BFS4":3,"skill-same:01M7APTDZKBN44WNT4KP25BFS4":48},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":29,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":32,"updatedAt":33,"install":34,"resources":47},"01M7APTDZKBN44WNT4KP25BFS4","drugdev-reference",null,"Use when correcting a drug discovery software misconception, looking up an attrition rate, dataset size, accuracy or timeline figure, finding the tools and resources, or needing a quick-reference picker — plus the current state of AI drug discovery clinical evidence and the FDA and EMA regulatory framework for AI. Companion to the other drug discovery software skills.","the-vibey-project\u002Fvibey","the-vibey-project","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F329298758?v=4","01M7AJHGZ9722CTJ4JNQZVRH0Q","the vibey project","01M7AJHGZVMG2K7TGXQ6XFHVG3",5,"MIT","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fmedicine-design-software-development\u002Fskills\u002Fdrugdev-reference\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: drugdev-reference\ndescription: \"Use when correcting a drug discovery software misconception, looking up an attrition rate, dataset size, accuracy or timeline figure, finding the tools and resources, or needing a quick-reference picker — plus the current state of AI drug discovery clinical evidence and the FDA and EMA regulatory framework for AI. Companion to the other drug discovery software skills.\"\n---\n\n# Drug Discovery Software: What's Live, Misconceptions, Numbers, and Tools\n\n> **Part 6 of 6** of the *Medicine Design Software Development* reference (plugin `medicine-design-software-development`), covering §24–§29. Sibling skills: `drugdev-pipeline-targets-and-drug-likeness` (§0–§3), `drugdev-representation-cheminformatics-data-quality-and-leakage` (§4–§7), `drugdev-protein-structure-docking-and-molecular-dynamics` (§8–§12), `drugdev-qsar-admet-generative-models-and-validation` (§13–§18), `drugdev-pipeline-engineering-compute-and-regulated-software` (§19–§23). Section numbers are shared across the set; a reference written as §N → `skill` points into that sibling skill.\n>\n> **Currency:** The chemistry and physics are stable. Two areas moved. See §24 for the AI drug discovery clinical evidence, and the FDA and EMA regulatory framework for AI.\n\n> **⚠️ Written for software engineers entering computational drug discovery — a domain\n> where the code is easy and the EVALUATION is brutally hard.** ⚠️ **You can build a\n> molecular property predictor in an afternoon and get a beautiful R². It will very\n> probably be measuring leakage, not chemistry** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`).\n>\n> **Complements an AI\u002FML reference (model architectures), a healthcare\u002Fclinical reference\n> (the trial end), and a scientific-computing reference (HPC and reproducibility).**\n>\n> **⚠️ GOTCHA** boxes mark the evaluation traps and the places where a correct-looking\n> result is meaningless.\n>\n> **The three ideas that organize this document:**\n> 1. **⚠️ The bottleneck is BIOLOGY, not compute** (§1 → `drugdev-pipeline-targets-and-drug-likeness`, §24.1). **Roughly 90% of candidates\n>    entering clinical trials fail, mostly for efficacy and toxicity reasons that appear\n>    only in humans — and no amount of upstream modelling has yet moved that number.**\n> 2. **⚠️ Random train\u002Ftest splits are almost always WRONG here** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`). **Chemical data is\n>    clustered into analogue series, so random splitting tests interpolation within a\n>    series and reports it as generalization. This is the field's defining methodological\n>    failure.**\n> 3. **⚠️ Prediction is cheap; VALIDATION is the product** (§18 → `drugdev-qsar-admet-generative-models-and-validation`, §24.2). **In a regulated\n>    context, an unvalidated model is not evidence — the FDA framework is explicitly about\n>    credibility for a stated context of use, not about accuracy in the abstract.**\n\n---\n\n## §24. What's Live — verified August 2026\n\n### 24.1 ⚠️ AI drug discovery meets clinical reality\n**⚠️ 2026 is the year the evidence starts arriving, and the honest position is \"promising\nearly, unproven where it counts.\"**\n\n- **⚠️ The scale of the bet**: **reported ~$60B invested in the sector since 2019 (with\n  broader life-science AI investment cited above $100B), ⚠️ around 173–175 AI-originated\n  drug programmes have entered human trials, and ⚠️ NONE has yet received FDA approval.**\n  ⚠️ **Industry estimates suggest 15–20 of those programmes could enter pivotal Phase III\n  trials during 2026.**\n- **⚠️ The encouraging number**: **published analyses report ⚠️ 80–90% Phase 1 success for\n  AI-discovered molecules, against a historical industry average variously cited at\n  ~40–65% or ~52%.** ⚠️ **Phase 2 success is reported around 40% — \"based on a limited\n  sample and broadly comparable to historical industry performance.\"**\n- **⚠️ Timelines do appear compressed**: **reported candidates reaching human trials in\n  under 18 months, versus a typical multi-year discovery phase.**\n\n> **⚠️ GOTCHA — the selection-bias objection is strong and should be stated before the\n> optimistic reading.** ⚠️ **The sample sizes are small and likely skewed, and early AI\n> programmes may simply have selected EASIER TARGETS** — **a pattern one analysis notes\n> is familiar from early monoclonal antibody development, \"when initial success rates\n> reflected careful target selection rather than technological superiority.\"**\n> ⚠️ **The open question is whether the early-stage advantage COMPOUNDS or merely\n> FRONT-LOADS the same attrition.** **⚠️ Phase 1 tests safety in healthy volunteers;\n> good molecular design plausibly helps there. Phase 2 tests whether the biology was\n> right, and design quality helps far less.**\n> **⚠️ As one summary put it: biology does not care how cleverly you found your molecule.**\n\n**⚠️ The current consensus position, stated carefully**: ⚠️ **\"the evidence does not yet\nshow that AI materially changes the probability of surviving Phase 2 and Phase 3 testing.\"**\n⚠️ **Clinical attrition remains around 90%, and Derek Lowe's observation is quoted across\nsources — most failures still result from efficacy or safety issues that emerge only in\nhumans** (§1 → `drugdev-pipeline-targets-and-drug-likeness`). ⚠️ **A plausible outcome flagged by commentators is \"accelerated timelines\nwithout improved efficacy — commercially valuable but scientifically underwhelming.\"**\n**⚠️ Sourcing caution: much of the enthusiastic coverage comes from vendors, consultancies\nand AI-in-pharma trade outlets; the peer-reviewed narrative reviews are notably more\nmeasured, and I've weighted toward those.**\n\n### 24.2 ⚠️ The regulatory framework for AI arrived\n**⚠️ This is the section that changes how you should BUILD, not just what you should\nbelieve.**\n\n- **⚠️ FDA published draft guidance on 6 January 2025** — ***Considerations for the Use of\n  Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological\n  Products*** — **establishing ⚠️ a risk-based CREDIBILITY ASSESSMENT FRAMEWORK.**\n  ⚠️ **The comment period closed 7 April 2025 (reported extended), and final guidance is\n  reported as expected in Q2 2026.**\n- **⚠️ The central concept is CONTEXT OF USE (COU)**: ⚠️ **credibility is defined as\n  \"trust in the performance of an AI model for a particular context of use.\"** **⚠️ There\n  is no such thing as a validated model in the abstract — only a model credible for a\n  stated role in a stated decision.**\n- **⚠️ The 7-step framework**, **as reported across sources:**\n```\n⚠️ 1. Define the QUESTION OF INTEREST — the specific regulatory question\n⚠️ 2. Define the CONTEXT OF USE — the model's precise role and how its\n      outputs influence decisions\n⚠️ 3. Assess MODEL RISK — ⚠️ from MODEL INFLUENCE × DECISION CONSEQUENCE\n      (a risk matrix; risk rises as either increases)\n   4. Develop a CREDIBILITY ASSESSMENT PLAN proportionate to that risk\n   5. Execute the plan\n   6. Document results in a CREDIBILITY ASSESSMENT REPORT\n⚠️ 7. Determine ADEQUACY for the context of use\n```\n- **⚠️ The critical scoping distinction for discovery teams**: ⚠️ **if AI is used to\n  DISCOVER a drug but traditional validation confirms safety and efficacy, extensive AI\n  documentation may not be required** — **the guidance targets AI that DIRECTLY SUPPORTS\n  REGULATORY DECISION-MAKING.** **⚠️ So a generative model that proposed a molecule later\n  validated conventionally is in a very different position from a model whose output\n  substitutes for an experiment.**\n- **⚠️ International alignment moved**: ⚠️ **on 14 January 2026 the FDA and EMA jointly\n  published *Guiding Principles of Good Machine Learning Practice for Drug Development* —\n  reported as ten high-level principles complementing the credibility framework.**\n  **EMA's September 2024 Reflection Paper on AI in medicines runs in parallel.**\n  ⚠️ **FDA also announced an AI-Enabled Optimization of Early-Phase Clinical Trials pilot,\n  with an RFI published 29 April 2026.**\n\n> **⚠️ GOTCHA — the engineering implications are concrete and they land on §19 → `drugdev-pipeline-engineering-compute-and-regulated-software`, §21 → `drugdev-pipeline-engineering-compute-and-regulated-software` and\n> §22 → `drugdev-pipeline-engineering-compute-and-regulated-software`.** ⚠️ **A credibility assessment plan must describe model design, DATA STRATEGY,\n> training methodology, performance metrics and evaluation methodology — and be\n> submitted early, within a submission or on request during inspection.**\n> **⚠️ That means: data provenance and versioning are regulatory artefacts, not\n> engineering hygiene; ⚠️ the split strategy (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) is a documented decision you must\n> defend; ⚠️ model versioning and the ability to reproduce a specific historical\n> prediction are requirements; and ⚠️ \"we retrained it and it's better now\" is a change\n> control event.**\n> **⚠️ Sponsors are also expected to define a threshold or risk matrix for when a\n> technology counts as AI under FDA's definition versus a complex decision tree —\n> which is a governance question most teams have not answered.**\n\n**⚠️ The honest caveat on all of this**: ⚠️ **the draft is a draft, a critical peer-reviewed\nreview of it exists in the literature, and the US policy environment around AI shifted\nwith the January 2025 executive order — so implementation detail is genuinely in motion.**\n**⚠️ The DIRECTION, however, is stable and internationally converging: risk-proportionate,\ncontext-of-use-anchored, documentation-heavy.** **⚠️ Build for that.**\n\n---\n\n## §25. Misconceptions\n\n| Misconception | Correction |\n|---|---|\n| AI will collapse drug development timelines | ⚠️ **Early-stage yes; clinical attrition unchanged** (§1 → `drugdev-pipeline-targets-and-drug-likeness`, §24.1) |\n| The bottleneck is compute | ⚠️ **It's biology and target validation** (§1 → `drugdev-pipeline-targets-and-drug-likeness`) |\n| Random train\u002Ftest split is standard practice | ⚠️ **It leaks badly here. Scaffold or time split** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| High test R² means the model works | ⚠️ **Check the split, then the noise floor** (§6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`, §7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| The same molecule has one SMILES | ⚠️ **Canonicalize, or you'll duplicate everything** (§4 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| Lipinski's rules define drug-likeness | ⚠️ **A retrospective observation, not a filter** (§3 → `drugdev-pipeline-targets-and-drug-likeness`) |\n| Public bioactivity data is precise | ⚠️ **~1 order of magnitude between labs is common** (§6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| AlphaFold solved structure for drug design | ⚠️ **Apo-like; docking into it underperforms** (§9 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| pLDDT is an accuracy guarantee | ⚠️ **A confidence score; low regions may be disordered** (§9 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| Docking scores predict affinity | ⚠️ **Poses reasonable, scoring poor. Triage only** (§10 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| Better docking software fixes scoring | ⚠️ **Protein prep matters more than program choice** (§10 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| MD tells you what the protein does | ⚠️ **Microseconds vs milliseconds of real biology** (§11 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| MM\u002FGBSA is a cheap FEP | ⚠️ **Much less reliable and widely over-trusted** (§12 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| Deep learning beats classical QSAR | ⚠️ **RF on ECFP is often competitive. Run it** (§13 → `drugdev-qsar-admet-generative-models-and-validation`, §17 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Similar structure means similar activity | ⚠️ **Activity cliffs, and they're the interesting cases** (§13 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Generating novel molecules is the hard part | ⚠️ **Scoring is. Generators hack the objective** (§15 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Novelty metrics show a generative model works | ⚠️ **Nearly meaningless. Synthesizability matters** (§15 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Benchmark SOTA transfers to practice | ⚠️ **Documented leakage in the standard benchmarks** (§17 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Test-set accuracy is the goal | ⚠️ **Fewer DMTA cycles is the goal** (§16 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Uncertainty estimation is optional | ⚠️ **It's what drives compound selection** (§16 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Validation means a good validation set | ⚠️ **In GxP it means documented qualification** (§22 → `drugdev-pipeline-engineering-compute-and-regulated-software`) |\n| Research code can be made GxP later | ⚠️ **Retrofitting traceability is far more expensive** (§22 → `drugdev-pipeline-engineering-compute-and-regulated-software`) |\n| A model is validated or not | ⚠️ **Credible for a CONTEXT OF USE, or not** (§24.2) |\n| AI-discovered drugs have been approved | ⚠️ **~173 in trials, zero approvals as of 2026** (§24.1) |\n| 90% Phase 1 success proves AI works | ⚠️ **Small, likely skewed sample; easier targets** (§24.1) |\n\n---\n\n## §26. Numbers\n\n```\n⚠️ Timeline \u002F cost      reported 10–15 yrs · >$1–2B per approved drug\n⚠️ Clinical attrition   ⚠️ ~90% of entrants never approved\n⚠️ Failure causes       reported ~40–50% efficacy · ~30% toxicity\n⚠️ Funnel               >10,000 screened → 20–100 in Phase I → ~1 approved\n⚠️ Experimental noise   ⚠️ ~0.5 log units on pIC50; ~1 order of magnitude\n                        between labs — ⚠️ this is your accuracy ceiling\n⚠️ FEP accuracy         reported ~1 kcal\u002Fmol (⚠️ ≈ 5× in affinity)\n⚠️ AlphaFold            pLDDT >90 for well-determined regions\n⚠️ AI programmes in trials  ⚠️ ~173–175 · ⚠️ ZERO approvals\n⚠️ AI Phase 1 success   reported 80–90% (vs ~40–65% historical)\n⚠️ AI Phase 2 success   ⚠️ reported ~40% — comparable to historical\n⚠️ Sector investment    reported ~$60B since 2019\n⚠️ FDA draft guidance   6 Jan 2025 · comments closed 7 Apr 2025 ·\n                        ⚠️ final expected Q2 2026\n⚠️ FDA–EMA principles   14 Jan 2026, reported 10 principles\n```\n\n---\n\n## §27. Tools and Resources\n\n| Tool \u002F Source | Why |\n|---|---|\n| **RDKit** | ⚠️ **The foundation. Learn it first** (§5 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| **ChEMBL \u002F PubChem \u002F BindingDB** | §6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`'s data, with §6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`'s caveats |\n| **Therapeutics Data Commons (TDC)** | ⚠️ **Benchmarks — read §17 → `drugdev-qsar-admet-generative-models-and-validation` before trusting them** |\n| **AutoDock Vina \u002F Smina** | Free docking (§10 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| **OpenMM** | ⚠️ **The most programmable MD engine** (§11 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| **AlphaFold DB \u002F ESMFold** | §9 → `drugdev-protein-structure-docking-and-molecular-dynamics` |\n| **DeepChem \u002F chemprop** | ⚠️ **chemprop is the strong GNN baseline** (§13 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| **scikit-learn + ECFP** | ⚠️ **THE baseline you must beat** (§17 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| **Nextflow \u002F Snakemake** | §19 → `drugdev-pipeline-engineering-compute-and-regulated-software` |\n| **Leach, *Molecular Modelling*** | The standard textbook |\n| **Gasteiger & Engel, *Chemoinformatics*** | §4–§5 → `drugdev-representation-cheminformatics-data-quality-and-leakage` foundations |\n| **Derek Lowe, *In the Pipeline*** | ⚠️ **The best sceptical running commentary on this field** |\n| **Pat Walters, *Practical Cheminformatics*** | ⚠️ **Excellent on §7 → `drugdev-representation-cheminformatics-data-quality-and-leakage` and §17 → `drugdev-qsar-admet-generative-models-and-validation`'s methodology traps** |\n| **FDA draft guidance (Jan 2025)** | ⚠️ **§24.2, primary source. Read it directly** |\n| **GAMP 5 (2nd ed.) \u002F FDA CSA guidance** | §22 → `drugdev-pipeline-engineering-compute-and-regulated-software` |\n| **IEC 62304 \u002F ISO 14971** | §23 → `drugdev-pipeline-engineering-compute-and-regulated-software` |\n\n---\n\n## §28. Quick Reference\n\n### 28.1 Picker\n| Question | Where |\n|---|---|\n| My model has great metrics — is it real? | ⚠️ **Check the split first** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| What baseline should I beat? | ⚠️ **Random forest on ECFP4** (§13 → `drugdev-qsar-admet-generative-models-and-validation`, §17 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| How accurate can this model possibly be? | ⚠️ **The experimental noise floor** (§6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| Should I use docking scores to rank? | ⚠️ **No. Triage only** (§10 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| I need affinity predictions for a series | ⚠️ **FEP, with a good structure** (§12 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| Can I dock into an AlphaFold model? | ⚠️ **You can; expect worse performance** (§9 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| Generative model outputs look great | ⚠️ **Check synthesizability and objective hacking** (§15 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Which compounds should we make next? | ⚠️ **Active learning with uncertainty** (§16 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| How do I join two chemical datasets? | ⚠️ **Standardize, then InChIKey** (§4 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| Model works in-house, fails on new chemistry | ⚠️ **Applicability domain** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| Will this need to be validated? | ⚠️ **Ask now, not later** (§22 → `drugdev-pipeline-engineering-compute-and-regulated-software`) |\n| Does the FDA guidance apply to us? | ⚠️ **Depends on context of use** (§24.2) |\n\n### 28.2 Before you trust a model\n- [ ] ⚠️ **Split is scaffold-based or time-based, and justified** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`)\n- [ ] ⚠️ **Duplicates removed after canonicalization** (§4 → `drugdev-representation-cheminformatics-data-quality-and-leakage`)\n- [ ] Standardization done AFTER splitting, or identically to both (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`)\n- [ ] ⚠️ **Simple fingerprint baseline run and reported** (§17 → `drugdev-qsar-admet-generative-models-and-validation`)\n- [ ] Performance compared against the experimental noise floor (§6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`)\n- [ ] ⚠️ **Applicability domain defined and enforced at inference** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`)\n- [ ] Uncertainty estimates present and calibrated (§16 → `drugdev-qsar-admet-generative-models-and-validation`)\n- [ ] ⚠️ **Activity cliff performance reported separately if relevant** (§13 → `drugdev-qsar-admet-generative-models-and-validation`)\n- [ ] ⚠️ **Prospective test planned, and failures will be recorded** (§18 → `drugdev-qsar-admet-generative-models-and-validation`)\n- [ ] Model version, data version and environment pinned (§21 → `drugdev-pipeline-engineering-compute-and-regulated-software`)\n\n---\n\n## §29. Method\n\n**§1–§23 → `drugdev-pipeline-targets-and-drug-likeness`, `drugdev-representation-cheminformatics-data-quality-and-leakage`, `drugdev-protein-structure-docking-and-molecular-dynamics`, `drugdev-qsar-admet-generative-models-and-validation`, `drugdev-pipeline-engineering-compute-and-regulated-software` rests on established cheminformatics, structural biology, computational chemistry\nand regulated-software practice** — **molecular representation, force fields and free\nenergy theory, QSAR methodology, and the GxP\u002FGAMP\u002F21 CFR Part 11 framework.** ⚠️ **The\nsplit-leakage problem and the docking scoring limitation have both been documented\nconsistently for many years and needed no verification.**\n\n**Two searches were run in August 2026**, on **AI drug discovery clinical evidence** and\n**the FDA regulatory framework** — ⚠️ **the first because the field's central empirical\nclaim is now being tested, the second because it changes how these systems must be built.**\n\n**Confidence.** **High** in §7 → `drugdev-representation-cheminformatics-data-quality-and-leakage` and §17 → `drugdev-qsar-admet-generative-models-and-validation`, which are the sections I'd most want read.\n⚠️ **Random splitting on clustered chemical data is the field's defining methodological\nfailure, and the repeated independent finding that a random forest on ECFP fingerprints\nmatches elaborate architectures once splits are fixed is the single most useful sanity\ncheck available.** **§6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`'s noise-floor point is the companion: a model reporting error below\nthe experimental measurement error is fitting something other than chemistry.**\n\n**High** in §10 → `drugdev-protein-structure-docking-and-molecular-dynamics`'s docking assessment and §12 → `drugdev-protein-structure-docking-and-molecular-dynamics`'s FEP framing — ⚠️ **both are long-standing,\nwell-characterized limitations that vendors consistently understate.**\n\n**Moderate-to-high** on §24.1. ⚠️ **The structural facts — roughly 173–175 AI-originated\nprogrammes in trials, zero approvals, ~90% overall clinical attrition — are consistent\nacross every source including the peer-reviewed reviews.** ⚠️ **The 80–90% Phase 1 figure\nappears widely but rests on a small sample (one source traces it to 24 molecules through\nDecember 2023), and the comparison baseline varies between sources (~40–65% vs ~52%),\nwhich alone should induce caution.** **⚠️ I've given the selection-bias objection prominence\nbecause it is the strongest argument in the debate and comes from the more sceptical\nsources, and because the enthusiastic coverage is dominated by vendors and trade outlets\nwith obvious interests.** **⚠️ My own read is that the Phase 1 advantage is plausible on\nmechanism — molecular design quality should help with safety in healthy volunteers — and\nthat Phase 2 is where the claim actually gets tested.**\n\n**High** on §24.2's framework, which traces to FDA's own published guidance and is\ncorroborated by multiple independent law-firm and industry analyses. ⚠️ **The seven steps,\nthe context-of-use concept, the model-influence × decision-consequence risk matrix, and\nthe credibility assessment plan and report are all from primary or near-primary sources.**\n⚠️ **The Q2 2026 finalization date and the January 2026 FDA–EMA joint principles are\nreported rather than confirmed by me against a primary document, and the guidance remains\na DRAFT — so I'd verify current status before relying on specifics.** **⚠️ The engineering\nimplications I've drawn in the gotcha box are my inference from the documentation\nrequirements, not language quoted from the guidance.**\n","# Drug Discovery Software: What's Live, Misconceptions, Numbers, and Tools\n\n> **Part 6 of 6** of the *Medicine Design Software Development* reference (plugin `medicine-design-software-development`), covering §24–§29. Sibling skills: `drugdev-pipeline-targets-and-drug-likeness` (§0–§3), `drugdev-representation-cheminformatics-data-quality-and-leakage` (§4–§7), `drugdev-protein-structure-docking-and-molecular-dynamics` (§8–§12), `drugdev-qsar-admet-generative-models-and-validation` (§13–§18), `drugdev-pipeline-engineering-compute-and-regulated-software` (§19–§23). Section numbers are shared across the set; a reference written as §N → `skill` points into that sibling skill.\n>\n> **Currency:** The chemistry and physics are stable. Two areas moved. See §24 for the AI drug discovery clinical evidence, and the FDA and EMA regulatory framework for AI.\n\n> **⚠️ Written for software engineers entering computational drug discovery — a domain\n> where the code is easy and the EVALUATION is brutally hard.** ⚠️ **You can build a\n> molecular property predictor in an afternoon and get a beautiful R². It will very\n> probably be measuring leakage, not chemistry** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`).\n>\n> **Complements an AI\u002FML reference (model architectures), a healthcare\u002Fclinical reference\n> (the trial end), and a scientific-computing reference (HPC and reproducibility).**\n>\n> **⚠️ GOTCHA** boxes mark the evaluation traps and the places where a correct-looking\n> result is meaningless.\n>\n> **The three ideas that organize this document:**\n> 1. **⚠️ The bottleneck is BIOLOGY, not compute** (§1 → `drugdev-pipeline-targets-and-drug-likeness`, §24.1). **Roughly 90% of candidates\n>    entering clinical trials fail, mostly for efficacy and toxicity reasons that appear\n>    only in humans — and no amount of upstream modelling has yet moved that number.**\n> 2. **⚠️ Random train\u002Ftest splits are almost always WRONG here** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`). **Chemical data is\n>    clustered into analogue series, so random splitting tests interpolation within a\n>    series and reports it as generalization. This is the field's defining methodological\n>    failure.**\n> 3. **⚠️ Prediction is cheap; VALIDATION is the product** (§18 → `drugdev-qsar-admet-generative-models-and-validation`, §24.2). **In a regulated\n>    context, an unvalidated model is not evidence — the FDA framework is explicitly about\n>    credibility for a stated context of use, not about accuracy in the abstract.**\n\n---\n\n## §24. What's Live — verified August 2026\n\n### 24.1 ⚠️ AI drug discovery meets clinical reality\n**⚠️ 2026 is the year the evidence starts arriving, and the honest position is \"promising\nearly, unproven where it counts.\"**\n\n- **⚠️ The scale of the bet**: **reported ~$60B invested in the sector since 2019 (with\n  broader life-science AI investment cited above $100B), ⚠️ around 173–175 AI-originated\n  drug programmes have entered human trials, and ⚠️ NONE has yet received FDA approval.**\n  ⚠️ **Industry estimates suggest 15–20 of those programmes could enter pivotal Phase III\n  trials during 2026.**\n- **⚠️ The encouraging number**: **published analyses report ⚠️ 80–90% Phase 1 success for\n  AI-discovered molecules, against a historical industry average variously cited at\n  ~40–65% or ~52%.** ⚠️ **Phase 2 success is reported around 40% — \"based on a limited\n  sample and broadly comparable to historical industry performance.\"**\n- **⚠️ Timelines do appear compressed**: **reported candidates reaching human trials in\n  under 18 months, versus a typical multi-year discovery phase.**\n\n> **⚠️ GOTCHA — the selection-bias objection is strong and should be stated before the\n> optimistic reading.** ⚠️ **The sample sizes are small and likely skewed, and early AI\n> programmes may simply have selected EASIER TARGETS** — **a pattern one analysis notes\n> is familiar from early monoclonal antibody development, \"when initial success rates\n> reflected careful target selection rather than technological superiority.\"**\n> ⚠️ **The open question is whether the early-stage advantage COMPOUNDS or merely\n> FRONT-LOADS the same attrition.** **⚠️ Phase 1 tests safety in healthy volunteers;\n> good molecular design plausibly helps there. Phase 2 tests whether the biology was\n> right, and design quality helps far less.**\n> **⚠️ As one summary put it: biology does not care how cleverly you found your molecule.**\n\n**⚠️ The current consensus position, stated carefully**: ⚠️ **\"the evidence does not yet\nshow that AI materially changes the probability of surviving Phase 2 and Phase 3 testing.\"**\n⚠️ **Clinical attrition remains around 90%, and Derek Lowe's observation is quoted across\nsources — most failures still result from efficacy or safety issues that emerge only in\nhumans** (§1 → `drugdev-pipeline-targets-and-drug-likeness`). ⚠️ **A plausible outcome flagged by commentators is \"accelerated timelines\nwithout improved efficacy — commercially valuable but scientifically underwhelming.\"**\n**⚠️ Sourcing caution: much of the enthusiastic coverage comes from vendors, consultancies\nand AI-in-pharma trade outlets; the peer-reviewed narrative reviews are notably more\nmeasured, and I've weighted toward those.**\n\n### 24.2 ⚠️ The regulatory framework for AI arrived\n**⚠️ This is the section that changes how you should BUILD, not just what you should\nbelieve.**\n\n- **⚠️ FDA published draft guidance on 6 January 2025** — ***Considerations for the Use of\n  Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological\n  Products*** — **establishing ⚠️ a risk-based CREDIBILITY ASSESSMENT FRAMEWORK.**\n  ⚠️ **The comment period closed 7 April 2025 (reported extended), and final guidance is\n  reported as expected in Q2 2026.**\n- **⚠️ The central concept is CONTEXT OF USE (COU)**: ⚠️ **credibility is defined as\n  \"trust in the performance of an AI model for a particular context of use.\"** **⚠️ There\n  is no such thing as a validated model in the abstract — only a model credible for a\n  stated role in a stated decision.**\n- **⚠️ The 7-step framework**, **as reported across sources:**\n```\n⚠️ 1. Define the QUESTION OF INTEREST — the specific regulatory question\n⚠️ 2. Define the CONTEXT OF USE — the model's precise role and how its\n      outputs influence decisions\n⚠️ 3. Assess MODEL RISK — ⚠️ from MODEL INFLUENCE × DECISION CONSEQUENCE\n      (a risk matrix; risk rises as either increases)\n   4. Develop a CREDIBILITY ASSESSMENT PLAN proportionate to that risk\n   5. Execute the plan\n   6. Document results in a CREDIBILITY ASSESSMENT REPORT\n⚠️ 7. Determine ADEQUACY for the context of use\n```\n- **⚠️ The critical scoping distinction for discovery teams**: ⚠️ **if AI is used to\n  DISCOVER a drug but traditional validation confirms safety and efficacy, extensive AI\n  documentation may not be required** — **the guidance targets AI that DIRECTLY SUPPORTS\n  REGULATORY DECISION-MAKING.** **⚠️ So a generative model that proposed a molecule later\n  validated conventionally is in a very different position from a model whose output\n  substitutes for an experiment.**\n- **⚠️ International alignment moved**: ⚠️ **on 14 January 2026 the FDA and EMA jointly\n  published *Guiding Principles of Good Machine Learning Practice for Drug Development* —\n  reported as ten high-level principles complementing the credibility framework.**\n  **EMA's September 2024 Reflection Paper on AI in medicines runs in parallel.**\n  ⚠️ **FDA also announced an AI-Enabled Optimization of Early-Phase Clinical Trials pilot,\n  with an RFI published 29 April 2026.**\n\n> **⚠️ GOTCHA — the engineering implications are concrete and they land on §19 → `drugdev-pipeline-engineering-compute-and-regulated-software`, §21 → `drugdev-pipeline-engineering-compute-and-regulated-software` and\n> §22 → `drugdev-pipeline-engineering-compute-and-regulated-software`.** ⚠️ **A credibility assessment plan must describe model design, DATA STRATEGY,\n> training methodology, performance metrics and evaluation methodology — and be\n> submitted early, within a submission or on request during inspection.**\n> **⚠️ That means: data provenance and versioning are regulatory artefacts, not\n> engineering hygiene; ⚠️ the split strategy (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) is a documented decision you must\n> defend; ⚠️ model versioning and the ability to reproduce a specific historical\n> prediction are requirements; and ⚠️ \"we retrained it and it's better now\" is a change\n> control event.**\n> **⚠️ Sponsors are also expected to define a threshold or risk matrix for when a\n> technology counts as AI under FDA's definition versus a complex decision tree —\n> which is a governance question most teams have not answered.**\n\n**⚠️ The honest caveat on all of this**: ⚠️ **the draft is a draft, a critical peer-reviewed\nreview of it exists in the literature, and the US policy environment around AI shifted\nwith the January 2025 executive order — so implementation detail is genuinely in motion.**\n**⚠️ The DIRECTION, however, is stable and internationally converging: risk-proportionate,\ncontext-of-use-anchored, documentation-heavy.** **⚠️ Build for that.**\n\n---\n\n## §25. Misconceptions\n\n| Misconception | Correction |\n|---|---|\n| AI will collapse drug development timelines | ⚠️ **Early-stage yes; clinical attrition unchanged** (§1 → `drugdev-pipeline-targets-and-drug-likeness`, §24.1) |\n| The bottleneck is compute | ⚠️ **It's biology and target validation** (§1 → `drugdev-pipeline-targets-and-drug-likeness`) |\n| Random train\u002Ftest split is standard practice | ⚠️ **It leaks badly here. Scaffold or time split** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| High test R² means the model works | ⚠️ **Check the split, then the noise floor** (§6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`, §7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| The same molecule has one SMILES | ⚠️ **Canonicalize, or you'll duplicate everything** (§4 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| Lipinski's rules define drug-likeness | ⚠️ **A retrospective observation, not a filter** (§3 → `drugdev-pipeline-targets-and-drug-likeness`) |\n| Public bioactivity data is precise | ⚠️ **~1 order of magnitude between labs is common** (§6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| AlphaFold solved structure for drug design | ⚠️ **Apo-like; docking into it underperforms** (§9 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| pLDDT is an accuracy guarantee | ⚠️ **A confidence score; low regions may be disordered** (§9 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| Docking scores predict affinity | ⚠️ **Poses reasonable, scoring poor. Triage only** (§10 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| Better docking software fixes scoring | ⚠️ **Protein prep matters more than program choice** (§10 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| MD tells you what the protein does | ⚠️ **Microseconds vs milliseconds of real biology** (§11 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| MM\u002FGBSA is a cheap FEP | ⚠️ **Much less reliable and widely over-trusted** (§12 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| Deep learning beats classical QSAR | ⚠️ **RF on ECFP is often competitive. Run it** (§13 → `drugdev-qsar-admet-generative-models-and-validation`, §17 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Similar structure means similar activity | ⚠️ **Activity cliffs, and they're the interesting cases** (§13 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Generating novel molecules is the hard part | ⚠️ **Scoring is. Generators hack the objective** (§15 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Novelty metrics show a generative model works | ⚠️ **Nearly meaningless. Synthesizability matters** (§15 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Benchmark SOTA transfers to practice | ⚠️ **Documented leakage in the standard benchmarks** (§17 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Test-set accuracy is the goal | ⚠️ **Fewer DMTA cycles is the goal** (§16 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Uncertainty estimation is optional | ⚠️ **It's what drives compound selection** (§16 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Validation means a good validation set | ⚠️ **In GxP it means documented qualification** (§22 → `drugdev-pipeline-engineering-compute-and-regulated-software`) |\n| Research code can be made GxP later | ⚠️ **Retrofitting traceability is far more expensive** (§22 → `drugdev-pipeline-engineering-compute-and-regulated-software`) |\n| A model is validated or not | ⚠️ **Credible for a CONTEXT OF USE, or not** (§24.2) |\n| AI-discovered drugs have been approved | ⚠️ **~173 in trials, zero approvals as of 2026** (§24.1) |\n| 90% Phase 1 success proves AI works | ⚠️ **Small, likely skewed sample; easier targets** (§24.1) |\n\n---\n\n## §26. Numbers\n\n```\n⚠️ Timeline \u002F cost      reported 10–15 yrs · >$1–2B per approved drug\n⚠️ Clinical attrition   ⚠️ ~90% of entrants never approved\n⚠️ Failure causes       reported ~40–50% efficacy · ~30% toxicity\n⚠️ Funnel               >10,000 screened → 20–100 in Phase I → ~1 approved\n⚠️ Experimental noise   ⚠️ ~0.5 log units on pIC50; ~1 order of magnitude\n                        between labs — ⚠️ this is your accuracy ceiling\n⚠️ FEP accuracy         reported ~1 kcal\u002Fmol (⚠️ ≈ 5× in affinity)\n⚠️ AlphaFold            pLDDT >90 for well-determined regions\n⚠️ AI programmes in trials  ⚠️ ~173–175 · ⚠️ ZERO approvals\n⚠️ AI Phase 1 success   reported 80–90% (vs ~40–65% historical)\n⚠️ AI Phase 2 success   ⚠️ reported ~40% — comparable to historical\n⚠️ Sector investment    reported ~$60B since 2019\n⚠️ FDA draft guidance   6 Jan 2025 · comments closed 7 Apr 2025 ·\n                        ⚠️ final expected Q2 2026\n⚠️ FDA–EMA principles   14 Jan 2026, reported 10 principles\n```\n\n---\n\n## §27. Tools and Resources\n\n| Tool \u002F Source | Why |\n|---|---|\n| **RDKit** | ⚠️ **The foundation. Learn it first** (§5 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| **ChEMBL \u002F PubChem \u002F BindingDB** | §6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`'s data, with §6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`'s caveats |\n| **Therapeutics Data Commons (TDC)** | ⚠️ **Benchmarks — read §17 → `drugdev-qsar-admet-generative-models-and-validation` before trusting them** |\n| **AutoDock Vina \u002F Smina** | Free docking (§10 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| **OpenMM** | ⚠️ **The most programmable MD engine** (§11 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| **AlphaFold DB \u002F ESMFold** | §9 → `drugdev-protein-structure-docking-and-molecular-dynamics` |\n| **DeepChem \u002F chemprop** | ⚠️ **chemprop is the strong GNN baseline** (§13 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| **scikit-learn + ECFP** | ⚠️ **THE baseline you must beat** (§17 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| **Nextflow \u002F Snakemake** | §19 → `drugdev-pipeline-engineering-compute-and-regulated-software` |\n| **Leach, *Molecular Modelling*** | The standard textbook |\n| **Gasteiger & Engel, *Chemoinformatics*** | §4–§5 → `drugdev-representation-cheminformatics-data-quality-and-leakage` foundations |\n| **Derek Lowe, *In the Pipeline*** | ⚠️ **The best sceptical running commentary on this field** |\n| **Pat Walters, *Practical Cheminformatics*** | ⚠️ **Excellent on §7 → `drugdev-representation-cheminformatics-data-quality-and-leakage` and §17 → `drugdev-qsar-admet-generative-models-and-validation`'s methodology traps** |\n| **FDA draft guidance (Jan 2025)** | ⚠️ **§24.2, primary source. Read it directly** |\n| **GAMP 5 (2nd ed.) \u002F FDA CSA guidance** | §22 → `drugdev-pipeline-engineering-compute-and-regulated-software` |\n| **IEC 62304 \u002F ISO 14971** | §23 → `drugdev-pipeline-engineering-compute-and-regulated-software` |\n\n---\n\n## §28. Quick Reference\n\n### 28.1 Picker\n| Question | Where |\n|---|---|\n| My model has great metrics — is it real? | ⚠️ **Check the split first** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| What baseline should I beat? | ⚠️ **Random forest on ECFP4** (§13 → `drugdev-qsar-admet-generative-models-and-validation`, §17 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| How accurate can this model possibly be? | ⚠️ **The experimental noise floor** (§6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| Should I use docking scores to rank? | ⚠️ **No. Triage only** (§10 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| I need affinity predictions for a series | ⚠️ **FEP, with a good structure** (§12 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| Can I dock into an AlphaFold model? | ⚠️ **You can; expect worse performance** (§9 → `drugdev-protein-structure-docking-and-molecular-dynamics`) |\n| Generative model outputs look great | ⚠️ **Check synthesizability and objective hacking** (§15 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| Which compounds should we make next? | ⚠️ **Active learning with uncertainty** (§16 → `drugdev-qsar-admet-generative-models-and-validation`) |\n| How do I join two chemical datasets? | ⚠️ **Standardize, then InChIKey** (§4 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| Model works in-house, fails on new chemistry | ⚠️ **Applicability domain** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`) |\n| Will this need to be validated? | ⚠️ **Ask now, not later** (§22 → `drugdev-pipeline-engineering-compute-and-regulated-software`) |\n| Does the FDA guidance apply to us? | ⚠️ **Depends on context of use** (§24.2) |\n\n### 28.2 Before you trust a model\n- [ ] ⚠️ **Split is scaffold-based or time-based, and justified** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`)\n- [ ] ⚠️ **Duplicates removed after canonicalization** (§4 → `drugdev-representation-cheminformatics-data-quality-and-leakage`)\n- [ ] Standardization done AFTER splitting, or identically to both (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`)\n- [ ] ⚠️ **Simple fingerprint baseline run and reported** (§17 → `drugdev-qsar-admet-generative-models-and-validation`)\n- [ ] Performance compared against the experimental noise floor (§6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`)\n- [ ] ⚠️ **Applicability domain defined and enforced at inference** (§7 → `drugdev-representation-cheminformatics-data-quality-and-leakage`)\n- [ ] Uncertainty estimates present and calibrated (§16 → `drugdev-qsar-admet-generative-models-and-validation`)\n- [ ] ⚠️ **Activity cliff performance reported separately if relevant** (§13 → `drugdev-qsar-admet-generative-models-and-validation`)\n- [ ] ⚠️ **Prospective test planned, and failures will be recorded** (§18 → `drugdev-qsar-admet-generative-models-and-validation`)\n- [ ] Model version, data version and environment pinned (§21 → `drugdev-pipeline-engineering-compute-and-regulated-software`)\n\n---\n\n## §29. Method\n\n**§1–§23 → `drugdev-pipeline-targets-and-drug-likeness`, `drugdev-representation-cheminformatics-data-quality-and-leakage`, `drugdev-protein-structure-docking-and-molecular-dynamics`, `drugdev-qsar-admet-generative-models-and-validation`, `drugdev-pipeline-engineering-compute-and-regulated-software` rests on established cheminformatics, structural biology, computational chemistry\nand regulated-software practice** — **molecular representation, force fields and free\nenergy theory, QSAR methodology, and the GxP\u002FGAMP\u002F21 CFR Part 11 framework.** ⚠️ **The\nsplit-leakage problem and the docking scoring limitation have both been documented\nconsistently for many years and needed no verification.**\n\n**Two searches were run in August 2026**, on **AI drug discovery clinical evidence** and\n**the FDA regulatory framework** — ⚠️ **the first because the field's central empirical\nclaim is now being tested, the second because it changes how these systems must be built.**\n\n**Confidence.** **High** in §7 → `drugdev-representation-cheminformatics-data-quality-and-leakage` and §17 → `drugdev-qsar-admet-generative-models-and-validation`, which are the sections I'd most want read.\n⚠️ **Random splitting on clustered chemical data is the field's defining methodological\nfailure, and the repeated independent finding that a random forest on ECFP fingerprints\nmatches elaborate architectures once splits are fixed is the single most useful sanity\ncheck available.** **§6 → `drugdev-representation-cheminformatics-data-quality-and-leakage`'s noise-floor point is the companion: a model reporting error below\nthe experimental measurement error is fitting something other than chemistry.**\n\n**High** in §10 → `drugdev-protein-structure-docking-and-molecular-dynamics`'s docking assessment and §12 → `drugdev-protein-structure-docking-and-molecular-dynamics`'s FEP framing — ⚠️ **both are long-standing,\nwell-characterized limitations that vendors consistently understate.**\n\n**Moderate-to-high** on §24.1. ⚠️ **The structural facts — roughly 173–175 AI-originated\nprogrammes in trials, zero approvals, ~90% overall clinical attrition — are consistent\nacross every source including the peer-reviewed reviews.** ⚠️ **The 80–90% Phase 1 figure\nappears widely but rests on a small sample (one source traces it to 24 molecules through\nDecember 2023), and the comparison baseline varies between sources (~40–65% vs ~52%),\nwhich alone should induce caution.** **⚠️ I've given the selection-bias objection prominence\nbecause it is the strongest argument in the debate and comes from the more sceptical\nsources, and because the enthusiastic coverage is dominated by vendors and trade outlets\nwith obvious interests.** **⚠️ My own read is that the Phase 1 advantage is plausible on\nmechanism — molecular design quality should help with safety in healthy volunteers — and\nthat Phase 2 is where the claim actually gets tested.**\n\n**High** on §24.2's framework, which traces to FDA's own published guidance and is\ncorroborated by multiple independent law-firm and industry analyses. ⚠️ **The seven steps,\nthe context-of-use concept, the model-influence × decision-consequence risk matrix, and\nthe credibility assessment plan and report are all from primary or near-primary sources.**\n⚠️ **The Q2 2026 finalization date and the January 2026 FDA–EMA joint principles are\nreported rather than confirmed by me against a primary document, and the guidance remains\na DRAFT — so I'd verify current status before relying on specifics.** **⚠️ The engineering\nimplications I've drawn in the gotcha box are my inference from the documentation\nrequirements, not language quoted from the guidance.**\n",{"name":5,"description":7},"develop","ae32736a1aabddabca5d2c96d665aa0ebc7d8700","src\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fmedicine-design-software-development\u002Fskills\u002Fdrugdev-reference",3,"https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey","https:\u002F\u002Fthe-vibey-project.github.io\u002Fvibey\u002F","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project","https:\u002F\u002Fvibewithadam.matthewsteinberger.com","freedom for software engineers","2026-11-15 05:47:42",{"source":8,"skillName":5,"commands":35,"prompt":45,"zipAvailable":46},[36,39,42],{"tool":37,"command":38},"npx","npx skills add the-vibey-project\u002Fvibey --skill drugdev-reference",{"tool":40,"command":41},"bunx","bunx skills add the-vibey-project\u002Fvibey --skill drugdev-reference",{"tool":43,"command":44},"pnpm","pnpm dlx skills add the-vibey-project\u002Fvibey --skill drugdev-reference","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fmedicine-design-software-development\u002Fskills\u002Fdrugdev-reference\u002FSKILL.md",true,[],[49,56,63,70,77,84],{"skillNo":50,"name":51,"nameZh":6,"description":52,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":53,"categories":54,"keywords":55,"publishedAt":19},"01M7AT1FQJAE1FQYAMDTF490Y4","white-paper-writing","Use when writing white papers, technical reports, policy documents, thought leadership pieces, or any long-form authoritative document. Triggers on requests to write, structure, outline, or review white papers, technical reports, position papers, policy briefs, or evidence-driven arguments. Also use for B2B content marketing strategy, technical documentation requiring authority, academic-style reports, or any document that must educate while persuading. Use when someone asks \"how should I structure this\" for a long-form document.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fwhite-paper-writing\u002FSKILL.md",[],[],{"skillNo":57,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M7AT16BJRJJKK2MASXRZS76G","prose-technical-mechanics","Use when editing or writing prose at a technical\u002Fmechanical level: choosing between Anglo-Saxon and Latinate diction, placing information in topic vs stress positions, writing periodic vs cumulative sentences, handling sentence rhythm, punctuation as logic, transitions, paragraph structure, or phonesthetics. Use when someone wants to understand WHY prose works or doesn't, not just surface-level style advice. Triggers on word choice, sentence structure, emphasis, rhythm, diction, or any deep writing mechanics question.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fprose-technical-mechanics\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":66,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":19},"01M7AT0W9HT0ZZ9Y287NW99MR3","prose-mechanics","Use when writing, editing, or analyzing prose style; when trying to avoid generic AI-sounding writing; when applying specific prose styles such as journalistic, academic, narrative, KJV biblical, Hemingway plain, Latinate\u002FJamesian, Orwellian, Augustan, Victorian, Early Modern (Bacon, Browne, Milton), or free indirect discourse; or when coaching someone on writing voice and mechanics. Triggers on any request to improve, analyze, or write in a specific prose style, to fix \"AI-sounding\" writing, to make prose more vivid or concrete, to vary sentence rhythm, or to understand how a particular writer achieves their effects.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fprose-mechanics\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M7AT0KR635JX66X2X6F3E2CP","narrative-structure","Use when structuring stories, pitches, presentations, white papers, or any persuasive narrative; analyzing the hero's journey in fiction or non-fiction; applying StoryBrand for business communication; designing narrative arcs for thought leadership; or understanding Vogler, Snyder, Propp, or Campbell frameworks. Triggers on any narrative structure, storytelling, pitch structure, StoryBrand, hero's journey, three-act structure, Freytag, Save the Cat, narrative arc, or \"how should I structure this story\u002Fpitch\u002Fpresentation\" request.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fnarrative-structure\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M7AT0AKGW09NKKZDYBT8VRDE","legal-brief-writing","Use when writing legal briefs, motions, or any court filing; when analyzing judicial cognition and persuasion; when applying elite brief-writing techniques from Roberts, Clement, Kagan, Garner; or when structuring legal arguments for maximum persuasion. Also use for autistic\u002Fneurodivergent advocates who need to understand how their natural communication style maps to elite legal advocacy. Triggers on any legal writing, brief drafting, motion preparation, or legal argument structuring request.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Flegal-brief-writing\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M7AT00VWPS4Q8B5VEMXX9NQT","journal-publishability-criteria","Use whenever a research paper, article or preprint is being written, revised, reviewed or published, and specifically before every publication of docs\u002Fpaper.md. Covers what legitimate journals actually gate on, in order of gating — the editor's desk screen (scope, novelty, package problems), what reviewers judge (methodological soundness, claims the evidence can carry, statistical and qualitative reporting), EQUATOR reporting guidelines, the ICMJE authorship criteria, COPE's view of text recycling, the near-universal generative-AI policy (no AI author, disclosure, author responsibility), TOP 2025 transparency standards, how to verify a journal is legitimate, submitting without an institution, and a pre-submission checklist. Triggers on requests to evaluate, review, prepare or submit a paper, to judge publishability, to write a declarations or data-availability section, to disclose AI use, or to check a journal.","https:\u002F\u002Fgithub.com\u002Fthe-vibey-project\u002Fvibey\u002Fblob\u002Fdevelop\u002Fsrc\u002Fvibey_tools\u002Fskills\u002Fplugins\u002Fwriting-craft\u002Fskills\u002Fjournal-publishability-criteria\u002FSKILL.md",[],[]]