The AI Drug Discovery Pipeline: From Target to IND in the Age of Machine Intelligence

"The drug that saves ten million lives was always knowable. The molecule existed. The target existed. The binding relationship existed. We simply lacked the computational intelligence to find it. That constraint has ended." — Zoirah Division Research Brief, Q2 2026

00. Transmission Header

CLASSIFICATION : Tresslers Group Intelligence // Zoirah Division
DOMAIN         : Drug Discovery / Computational Biology / AI in Medicine / Clinical AI
STATUS         : Active Intelligence — Clinical and Scientific
DATE           : 2026.05.10
KEY MILESTONE  : Rentosertib (Insilico Medicine): first end-to-end AI drug Phase IIa — 2025
                 AlphaFold 3 (Google DeepMind): ~50% improvement in protein-molecule interaction prediction — May 2024
                 No AI-discovered drug has received full FDA approval as of publication date
TRADITIONAL    : Drug discovery: 10–15 years average; $2.6B average cost (including failures)
ALERT LEVEL    : High — AI drug discovery is at clinical inflection; investment window active

The traditional pharmaceutical discovery process is one of the least efficient industrial processes in modern economies. Of every 10,000 compounds that enter pre-clinical testing, approximately 10 proceed to clinical trials — and of those 10, fewer than 1 is approved. The 10–15 year timeline and $2.6 billion average cost per approved drug (including the cost of failures) reflects an industry that has historically been forced to search a near-infinite chemical space using tools that were never designed for the scale of the problem.

AI is not incrementally improving this process. It is replacing its least efficient components with a fundamentally different approach: computational exploration of molecular space at scales previously impossible, combined with structural biology predictions of the precision previously requiring decades of X-ray crystallography.

The transition is happening faster than most of the pharmaceutical industry expected, and slower than most AI companies claimed. The 2025 Rentosertib Phase IIa results are the first rigorous clinical evidence that the end-to-end AI pipeline — from target identification through molecule design, through pre-clinical validation, through clinical proof-of-concept — can work. The industry has been watching. The investment has followed.

01. The Traditional Drug Discovery Pipeline — The Benchmark

Understanding AI's impact requires mapping the traditional process it is disrupting:

Rendering diagram...

The AI disruption points:

• ▸Stages 1–2 (Target ID/Validation): AI analysis of genomics, proteomics, and clinical outcome data to identify and validate disease targets — compressing 2–5 years to months
• ▸Stages 3–4 (Hit ID/Lead Optimization): Generative AI design of novel molecules with predicted binding affinity, selectivity, and drug-like properties — replacing years of high-throughput screening with computational generation
• ▸Stage 5 (Pre-Clinical): AI-predicted ADMET properties (absorption, distribution, metabolism, excretion, toxicity) — reducing animal testing requirements and identifying failures earlier

02. AlphaFold 3 — The Structural Biology Revolution

The protein folding problem and its solution:

For 50 years, structural biology's central challenge was predicting a protein's three-dimensional structure from its amino acid sequence. The structure determines function — and drug binding sites are structural features. X-ray crystallography and cryo-electron microscopy can determine protein structures, but at the cost of months to years per structure.

• ▸AlphaFold 2 (2020): solved the protein structure prediction problem for single proteins with near-experimental accuracy — Nobel Prize-recognized breakthrough
• ▸AlphaFold 3 (May 2024, Google DeepMind / Isomorphic Labs): extended prediction to interactions between proteins and other biomolecules — DNA, RNA, ligands (small molecule drugs), and ions

AlphaFold 3's drug discovery significance:

AlphaFold 3 improved the accuracy of predicting protein-ligand interactions — how a drug molecule binds to its protein target — by approximately 50% compared to previous methods (measured on PoseBusters benchmark for drug-like molecules).

Rendering diagram...

Isomorphic Labs — the drug discovery subsidiary of Google DeepMind — was founded explicitly to apply AlphaFold and successor technologies to drug discovery. It operates with direct access to the AlphaFold research team and has active pharmaceutical partnerships to discover drugs using these computational tools.

03. The First Clinical Proof-of-Concept — Rentosertib

Insilico Medicine and ISM001-055 (Rentosertib):

Insilico Medicine is the company that has come closest to validating the end-to-end AI drug discovery thesis. Its Pharma.AI platform:

• ▸Target identification: AI analysis of publicly available genomics and proteomics data to identify TNIK (TRAF2 and NCK-interacting kinase) as a target for IPF (idiopathic pulmonary fibrosis)
• ▸Molecule design: generative AI design of novel TNIK inhibitors with predicted drug-like properties
• ▸Candidate selection: ISM001-055 (Rentosertib) selected as the lead candidate
• ▸Pre-clinical validation: animal studies demonstrating efficacy in IPF models
• ▸IND filing and Phase I: safety established in healthy volunteers
• ▸Phase IIa (2025): RESULTS — favorable safety profile and dose-dependent efficacy signals in IPF patients

Why this matters: Rentosertib is the first drug where:

1. ▸The disease target was identified by AI
2. ▸The drug molecule was designed by AI
3. ▸The pre-clinical candidate selection was guided by AI
4. ▸It proceeded all the way to human clinical proof-of-concept

No other drug has completed this full end-to-end AI pathway through Phase II clinical trials. This is why 2025 is described as the "watershed year" for AI drug discovery — the clinical evidence, not just the computational promise, exists.

The honest caveat: Phase IIa completion does not mean approval. The drug must still complete Phase IIb and Phase III trials, which together may take 4–6 more years and may fail. The 2025 milestone is proof-of-concept, not proof-of-approval. The industry's projection for the first FDA-approved AI-discovered drug is 2026–2027.

04. The Key Companies — Mapping the AI Drug Discovery Landscape

Rendering diagram...

Recursion Pharmaceuticals — the scale approach: Recursion's strategy differs from Insilico's: rather than AI-designed molecules, Recursion uses massive biological imaging — running millions of cellular experiments, imaging the results, and training AI on these "phenotypic" datasets to identify drug candidates. Its REC-4881 (for familial adenomatous polyposis, FAP) has shown positive early efficacy signals leading to FDA registration pathway discussions.

The NVIDIA BioNemo platform: NVIDIA has entered drug discovery AI directly with BioNemo — a collection of foundation models pre-trained on molecular data (protein sequences, chemical structures, genomic sequences). BioNemo provides the ML infrastructure layer for drug discovery AI in the same way that PyTorch and CUDA provide the ML infrastructure layer for general AI. Drug discovery companies building on NVIDIA hardware now have foundation models they can fine-tune for specific discovery tasks.

05. The AI Drug Discovery Technology Stack

Rendering diagram...

The ADMET prediction layer — why it matters:

ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) properties determine whether a drug candidate is clinically viable regardless of its target binding affinity. A molecule that binds its target perfectly but is metabolized within minutes, or is toxic to the liver, or doesn't cross the blood-brain barrier when needed, fails in clinical development.

AI-predicted ADMET properties allow drug discovery teams to filter candidates computationally before synthesizing them physically — dramatically reducing the chemical synthesis workload and identifying failures earlier, when they are less expensive to discard.

06. The FDA Regulatory Engagement — Where the Interface Stands

The FDA has engaged actively with AI drug discovery as the programs have matured:

2025 FDA Draft Guidance: "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision Making for Drug and Biological Products" — provides the first formal FDA guidance on:

• ▸What AI-generated data can support regulatory submissions
• ▸Documentation requirements for AI models used in drug development (model version, training data, validation)
• ▸Transparency requirements for AI in clinical decision support
• ▸Criteria for when AI-generated evidence is sufficient vs. supplementary

The regulatory thesis: AI-generated evidence is being incorporated into regulatory submissions — but as supporting data alongside traditional experimental evidence, not as a replacement. An AI-predicted binding affinity still needs experimental confirmation. An AI-identified target biomarker still needs validation in patient samples. The FDA is integrating AI while maintaining the experimental validation requirements that protect patient safety.

The CDER AI Pilot Program: the FDA's Center for Drug Evaluation and Research has established an internal pilot program evaluating AI use in regulatory review itself — using AI to process and analyze the enormous data submissions that accompany NDA/BLA applications. This suggests the FDA is developing AI infrastructure on both sides of the drug development interface.

07. The Pipeline Status — Where the Evidence Is

Current status as of 2026 (verified):

The pipeline reality check: the table above shows what is in clinical trials — not what has succeeded. Clinical trial failure rates are 90%+ at Phase I-II. The AI drug discovery industry will produce many clinical failures before its first approval. The Rentosertib Phase IIa result is the most significant clinical validation, but it is not an approval.

08. The Timeline Compression Thesis — What AI Actually Delivers

The conventional pharmaceutical industry claims AI drug discovery can compress the 10–15 year timeline by 25–70%. What does this mean in practice?

The honest assessment: the compression is primarily in the pre-clinical stages — target ID through IND. Clinical biology has not changed. Phase I, II, and III trials still require the number of patients and the duration needed to observe clinical outcomes. AI can optimize trial design (patient selection, site identification, biomarker-stratified enrollment) but cannot change the fundamental biology of how long it takes to observe whether a drug works.

The 40–50% compression of pre-clinical stages translates to approximately 2–4 years off the total timeline — a significant improvement but not the "cure in months" narrative that has been overpromised by some AI drug discovery companies.

09. The Zoirah Intelligence Mandate

The Zoirah division monitors this field continuously across three intelligence products:

Drug discovery pipeline intelligence: continuous monitoring of AI drug discovery company pipeline updates, clinical trial registrations (ClinicalTrials.gov), FDA guidance updates, and scientific publications — synthesized into structured pipeline intelligence that updates as the field moves.

Target landscape mapping: knowledge graph of validated and emerging drug targets across therapeutic areas — identifying target validation status, competitive landscape, and AI-generated target hypotheses from literature.

Regulatory intelligence: tracking FDA's evolving stance on AI-generated evidence, country-by-country regulatory approvals of AI-assisted drug submissions, and the emerging international harmonization (or divergence) of AI drug discovery regulation.

10. The Tresslers Group Thesis

AI drug discovery is not a future technology. It is a current clinical reality — at one specific pipeline stage. The question is how fast the clinical evidence accumulates.

The Rentosertib Phase IIa result is not a finished product. It is the first clinical proof-of-concept — the moment at which the industry can no longer dismiss AI drug discovery as theoretical. The next milestone is the first full FDA approval, projected 2026–2027. After that, the precedent is established, and the floodgates of AI-designed drug submissions open.

The drug discovery industry is structured around knowledge of targets, mechanisms, and molecular structure-activity relationships. AI is converting that knowledge from implicit (stored in researchers' heads and laboratory notebooks) to explicit (encoded in models and databases). The organizations that build and maintain the richest AI drug discovery knowledge substrates — the training data, the validated models, the annotated outcomes — will be positioned to discover the drugs that save the next decade's patient population.

Zoirah exists at that intersection: intelligence infrastructure for the AI drug discovery era.

The molecule is findable. The intelligence finds it.

References & Source Intelligence

1. ▸Insilico Medicine. (2025). Rentosertib Phase IIa Trial Results: First End-to-End AI-Discovered Drug Clinical Proof-of-Concept. Insilico.com.
2. ▸Google DeepMind / Isomorphic Labs. (2024, May). AlphaFold 3: Predicting Protein-Ligand Interactions with ~50% Improvement. Nature / blog.google.
3. ▸FDA CDER. (2025). Draft Guidance: Considerations for the Use of Artificial Intelligence to Support Regulatory Decision Making for Drug and Biological Products.
4. ▸Recursion Pharmaceuticals. (2024–2025). REC-4881 Pipeline Update and FDA Registration Pathway Discussion.
5. ▸NVIDIA. (2024–2025). BioNemo: Foundation Models for Drug Discovery.
6. ▸Biospace. (2024). Insilico Medicine Nominates 20+ Development Candidates via Pharma.AI Platform.
7. ▸NIH / PubMed. AlphaFold 3 Benchmark Results: PoseBusters Dataset.
8. ▸Tresslers Group Intelligence. (2026). AI Diagnostics at Scale. [tresslersgroup.com/insights/ai-diagnostics-clinical-intelligence-2026]
9. ▸Tresslers Group Intelligence. (2026). Precision Medicine & Pharmacogenomics. [tresslersgroup.com/insights/precision-medicine-pharmacogenomics-2026]

Tresslers Group Intelligence — Zoirah Division Driven by Innovation. Defined by Impact. Clinical Intelligence at the Frontier of AI Medicine. © 2026 Tresslers Group. Transmission Complete.