How Compound Management Reshaped Drug Discovery: 3 Axes of AI × Robotics-Driven Asset Reuse | Vol.2

Key Takeaways

• Major pharmaceutical companies hold libraries of millions to tens of millions of compounds, accumulated through high-throughput screening (HTS) over 20-30 years. The vast majority of these are “sleeping assets” — never advanced to development. AI is changing this by re-evaluating these dormant assets and discovering new indications and optimal combinations from existing drugs — a “compound recycling” revolution.
• Major applications of AI-driven compound management: (1) drug repurposing (discovering new indications for existing drugs), (2) matching with target identification (screening existing compounds against new targets), (3) ADMET property prediction (pre-evaluation of absorption/distribution/metabolism/excretion/toxicity), (4) multi-objective optimization (balancing activity, selectivity, properties).
• Concrete example: BenevolentAI’s strategic partnership with AstraZeneca predicted Baricitinib (existing JAK inhibitor) as a COVID-19 indication, leading to emergency use authorization (2020). This established a new narrative of AI-driven “repurposing existing drugs for different diseases.”
• Commercial impact: AI-driven drug repurposing achieves $300-500M / 3-5 years efficiency, vs $2.6B/15 years for new drug development — orders of magnitude better. Failure risk also lower (existing drugs have established safety/PK). Becoming a key axis of pharma R&D strategy.

Introduction — Technology That Awakens “Sleeping Assets”

“Pharmaceutical R&D” typically conjures the image of “designing entirely new molecules from scratch.” While creating new chemical entities (NCE) is pharma’s prestige work, in its shadow lies a vast overlooked asset: compounds tested at some point but not adopted.

Major pharma compound libraries typically:

• Pfizer: millions of compounds
• Merck KGaA / MSD: 5+ million
• Roche / Genentech: 3-5 million
• Takeda: 1-2 million (including Millennium, Arena acquisitions)
• Daiichi Sankyo: 1+ million

These are HTS assets accumulated over 20-30 years. Adopted compounds advance to drug development; the majority are dormant for reasons like “didn’t work for the target,” “poor ADMET,” “patent strategy issues.”

AI is changing this — re-evaluating “dormant assets.” This article covers AI-driven compound management’s major applications, commercial impact, industry initiatives, and remaining challenges.

Main Body

1. Drug Repurposing — “New Indications for Existing Drugs”

Drug repurposing (drug repositioning) is the strategy of re-applying approved drugs to different diseases. AI has become decisive in this area over the past 5 years.

Representative cases:

• Baricitinib: JAK1/2 inhibitor approved for rheumatoid arthritis. BenevolentAI’s AI predicted COVID-19 severity prevention indication in February 2020. Confirmed in NIH ACTT-2 trial, FDA Emergency Use Authorization in May 2020.
• Sirolimus: Immunosuppressant. Insilico’s AI analysis discovered tuberous sclerosis additional indication.
• Metformin: Diabetes drug. AI analysis is driving validation of anti-aging and cancer prevention effects (TAME Trial, NIH-led).
• Disulfiram: Alcohol dependence drug. AI analysis discovered antitumor effect in metastatic melanoma; Phase 2 ongoing.

2. AI-Driven Drug Repurposing — Technical Approaches

Four AI methods to “wake up” sleeping drugs:

(1) Knowledge graph-driven inference: BenevolentAI, Recursion etc. Integrate papers, databases, and patents into a massive knowledge graph; machine learning discovers indirect drug-disease relations.

(2) Molecular similarity search: From the molecular fingerprint of a drug effective against target A, predict compounds likely effective against target B. Atomwise, BenchSci core technology.

(3) Pharmacophore / 3D structure matching: Based on target protein structure, narrow down binding candidates from existing drug libraries. Schrödinger’s core.

(4) Cell phenotype + machine learning: Recursion’s flagship technology. Treat cancer cells, fibroblasts, neurons etc. with each drug, machine-learn the imaging changes (cell morphology, signaling markers, mitochondrial function) into “fingerprints.” Drugs producing the same phenotype likely act on the same target/pathway.

3. Multi-Objective Optimization — “ADMET vs Activity Tradeoffs”

New compound development requires multi-dimensional optimization beyond just activity (target binding, pharmacological effect): ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) and properties (solubility, stability, manufacturability).

AI excels in multi-objective optimization:

• Schrödinger LiveDesign: predicts activity, ADMET, properties simultaneously during molecular design, visualizes tradeoffs
• Genentech’s internal platform: re-evaluates entire internal compound library with ADMET prediction models, extracts “reactivation candidates”
• Pfizer’s AI-driven SAR: predicts structure-activity relationships with AI, halves optimization round count

These multi-objective optimization technologies also apply to re-evaluating “dormant compounds.” Compounds previously rejected for “poor ADMET” can re-emerge as repurposing candidates with modern AI models targeting different objectives.

4. Internal Compound Library “Re-Inventory”

Major pharma’s strategic initiative: full AI re-evaluation (re-inventory) of internal libraries active in 2024-26.

Typical re-inventory project:

1. Re-organize 30 years of HTS data with standardized molecular representations (SMILES, InChI, molecular graphs)
2. Re-calculate molecular descriptors (properties, structural features, ADMET predictions) with AI models
3. Run virtual screening of entire library against new targets (thousands identified in past 20 years)
4. Promising candidates re-evaluated experimentally, advanced to development

Through this approach, Takeda, Roche, Pfizer etc. discover hundreds-thousands of new development candidates per year from previously dormant compounds. Most are excluded at preclinical stage; a few percent advance to Phase 1, several reach Phase 2/3.

5. Linkage with Target Identification

Compound management closely links with target identification. When AI-driven target identification discovers “a key new target for a disease,” a parallel mechanism searches whether “existing compounds in library hit that target”.

Concrete examples:

• Recursion + Bayer partnership: integrates Recursion’s cell phenotype AI with Bayer’s compound library, advancing matching of fibrotic disease new targets + existing drug candidates
• Roche’s AI-driven BD/BD (Business Development & Drug Discovery): internal AI team continuously re-evaluates new target priorities, optimizes compound resource allocation

6. Limitations and Challenges

AI-driven compound management limitations:

First, data quality. 30 years of HTS data is often not standardized: compound ID duplications/missing, assay condition variations. “Garbage in, garbage out” risk.

Second, biological validity. High AI prediction scores don’t guarantee actual clinical translation. The cell assay → mouse model → human clinical translatability wall persists.

Third, patent strategy. New indications for existing drugs raise the issue of remaining patent life. Without remaining base patent protection, market exclusivity is hard to secure; competition with generic manufacturers.

Fourth, regulatory approval pathway. Drug repurposing still requires Phase 2/3 trials; “existing drug = easy” doesn’t apply.

7. Commercial Impact

Economic value of AI-driven compound management:

While unlikely to produce $10B+ blockbusters like new drug discovery, capturing $1-3B markets with $300-500M investment is an efficient approach. AI-driven repurposing has become important in pharma R&D portfolios.

8. Limitations Complemented — “AI Cannot Solve Everything”

An important limitation: “AI cannot predict what’s not in data.” Truly new targets, completely new chemical scaffolds, completely new disease mechanisms — for these, AI infers only from “past similarities,” so true innovation still depends on human scientist creativity.

However, AI’s strength is “discovering patterns humans cannot find” — extracting maximum value from existing assets. The “maximum utilization of existing assets” narrative will repeat in Volumes 3-5 (preclinical-to-clinical, trial operations, pharmacovigilance).

Summary

• Major pharma’s millions-to-tens-of-millions compound libraries are 20-30 year HTS assets. AI is rapidly becoming critical for re-evaluating these “dormant assets.”
• Major applications: drug repurposing, target matching, ADMET prediction, multi-objective optimization.
• Representative cases: BenevolentAI’s Baricitinib COVID-19 indication prediction, Insilico’s Sirolimus tuberous sclerosis discovery, ongoing metformin anti-aging validation.
• Technical approaches: knowledge graph inference, molecular similarity search, pharmacophore/3D structure matching, cell phenotype + machine learning.
• Commercial value: 3-5 year development, $300-500M cost, 30-40% Phase 1 → approval (3× new drug development).
• Limitations: data quality, biological validity, patent strategy, regulatory pathway, fundamental “cannot predict what’s not in data” constraint.

My Thoughts and Outlook

The article’s core insight: “AI drug discovery’s implementation value lies not just in new molecule design, but in re-utilization of past assets.” Compound management — an unglamorous theme — has become a key axis substantively lifting pharma R&D profit structure.
Three structural implications for the global ecosystem. First, proprietary compound libraries become increasingly strategic assets. The pharma majors (Pfizer, Merck, Roche, Novartis) with the largest historical HTS data have the most material to mine via AI re-inventory. Smaller biotechs increasingly seek library access partnerships with majors. Second, drug repurposing economics changes pharma R&D portfolio thinking. A $300-500M, 3-5 year repurposing program competing alongside $2.6B, 10-15 year new drug programs reshapes capital allocation. CFOs at major pharma are increasingly receptive to repurposing portfolios as risk-balanced complements to flagship NCE programs. Third, regulatory pathways for AI-discovered repurposings are evolving. The FDA’s 505(b)(2) pathway and accelerated approval mechanisms now accommodate AI-driven repurposing data. The first AI-discovered drug to reach approval via repurposing pathway will be a defining moment.
2026 is the year AI is rapidly commoditizing knowledge work. Compound management AI is precisely augmentation territory — AI surfaces patterns that humans cannot, but humans (medicinal chemists, biologists, clinicians) ultimately validate and develop. Volume 3 will explore the AI of preclinical-to-clinical translation.

Coming Next

Volume 3 covers “Preclinical-to-Clinical Speed — Mechanistic Modeling × AI.” PBPK (physiologically-based pharmacokinetic modeling), QSP (quantitative systems pharmacology), digital twins — how the fusion of these computational models with AI is transforming preclinical → Phase 1 transition prediction accuracy, trial design optimization, automated dose-regimen decisions, and adverse event prediction.

Edited by the Morningglorysciences team.

Related Articles

• AI Drug Discovery: Front and Back, Vol. 1: Headlines vs Real Profit Structure
• AI Drug Discovery: Front and Back, Vol. 3: Preclinical-to-Clinical Speed (forthcoming)