AI Drug Discovery: Every Step, Every Timeline, and Who Leads (October 2026)

AI is now inside almost every step of drug discovery, but it doesn't speed up every step equally. It cuts the discovery phase (target → drug candidate) from roughly 4–6 years to 1.5–3. The clinical trials still take 6–8 years, because human biology runs at its own pace. This report walks through each step and its timeline (US vs China, EU, Japan and Australia), lists the AI-designed drugs actually in human trials as of October 2026, and ranks the stocks that own the tools, the data and the drugs.

Facts are sourced (links at the bottom). Anything marked (inference) or (editorial) is BullishAgent's own judgement.

The short version

Key point: AI makes molecules that are safer and better behaved (that's why Phase 1 success is high), but it hasn't yet shown it picks targets that work better in patients (Phase 2 success is unchanged). Phase 2 is where most of the money is still lost.

Part 1 — The steps, one by one

Step 1 — Target identification & validation

What it is: Find the protein, gene or pathway that drives the disease, and prove that blocking it helps.

What AI does: Models read omics data (genes, proteins, single-cell), patient records, millions of papers and patents, and rank candidate targets. Insilico's PandaOmics picked TNIK as an anti-fibrosis target. Recursion photographs billions of cells under thousands of genetic and chemical changes ("phenomics") and maps which ones look like disease reversal.

Time: Traditional 1–3 years → AI-assisted 3–12 months. Validation in cells and animals still takes months.

Step 2 — Hit finding (first molecules that bind)

What it is: Find any molecule that sticks to the target.

What AI does: Structure prediction (AlphaFold 3, Boltz, Chai) gives you the target's 3D shape. Then virtual screening tests billions of virtual compounds, and generative models design new ones from scratch. Physics-based free-energy methods (Schrödinger's FEP+) predict how tightly each one binds before anyone makes it.

Time: Traditional 1–2 years of high-throughput screening → AI-assisted weeks to a few months.

Step 3 — Lead optimization (the design–make–test–analyze loop)

What it is: Turn a weak "hit" into a drug-like molecule that's potent, selective, orally absorbed, non-toxic and long-lasting. This is the most expensive part of discovery.

What AI does: Predicts ADMET (absorption, distribution, metabolism, excretion, toxicity) and suggests the next round of molecules. Robotic labs then make and test them, and the results retrain the model. Each cycle gets shorter and fewer compounds have to be made.

Time: Traditional 1.5–3 years and 2,500–5,000 compounds → AI-assisted 9–18 months and ~100–300 compounds. Recursion reached its CDK7 candidate REC-617 in under 11 months with 136 compounds.

Output: a preclinical candidate (PCC).

Step 4 — IND-enabling preclinical studies

What it is: The work regulators require before the first human dose: GLP toxicology in two species, safety pharmacology, a scaled-up manufacturing process (CMC), and a stable formulation.

What AI does: So far, the least of any step. It helps with toxicity prediction and process chemistry. This is the next frontier. In April 2025 the FDA announced a plan to phase out animal-testing requirements for monoclonal antibodies in favor of "new approach methods" (computer models, organoids). That opens the door for AI toxicology to replace part of this step. (inference): this is where the next real time saving will come from.

Time: 12–18 months in every country, with AI or without.

Step 5 — Permission to start human trials (IND / CTA)

This is where countries differ the most.

China matters for this theme specifically. Insilico runs its Phase 3 for rentosertib in China, and Chinese AI biotechs (Insilico, XtalPi) use fast domestic trials to create data they then license to Western pharma.

Step 6 — Phase 1 (safety, dosing; ~20–100 people)

What AI does: It improves the molecules that enter, so they fail less here (80–90% success in the BCG analysis). AI also helps with patient selection and digital "twins" of control arms.

Time: 1–1.5 years (shorter in Australia/China because sites start faster).

Step 7 — Phase 2 (does it work? ~100–300 patients)

What AI does: Biomarker-based patient selection, picking the indication, adaptive trial design. Tempus-type clinical-genomic data is used to enroll the patients most likely to respond.

Time: 2–3 years. Success remains ~40%, the same as conventional drugs. This is the step AI hasn't solved yet.

Step 8 — Phase 3 (confirm at scale; 300–3,000+ patients)

What AI does: Site selection, recruitment, operations, and faster data cleaning (IQVIA, Veeva, Medidata). It saves months, not years.

Time: 2–4 years.

Step 9 — Marketing application & review

The FDA's AI-specific rulebook is still the January 2025 draft guidance (a 7-step, risk-based "credibility framework" for any AI model used to support a regulatory decision). As of this writing it has not been finalized.

Step 10 — Post-market

Real-world evidence, safety signal detection (pharmacovigilance), and label expansion. AI is used heavily here for data mining, but it rarely moves stocks.

Total timeline: two real examples

• Rentosertib (Insilico, TNIK inhibitor, IPF): AI-picked target and AI-generated molecule. ~18 months from program start to candidate. Phase 2a data published in Nature Medicine (June 2025). Phase 3 (GENESIS-IPF-3, 320 patients, 47 sites in China) began in 2026. The lead investigator estimates 3–4 years from Phase 3 start to possible approval, so roughly 2029–2030, about 10 years after the target was found.
• Zasocitinib (TAK-279, TYK2 inhibitor, psoriasis): designed by Nimbus with Schrödinger's physics-based FEP platform starting in 2016. Takeda bought it in 2023. Phase 3 was positive in Dec 2025. The FDA accepted the NDA under priority review in Sept 2026, with a decision expected Q1 2027. If approved, it is likely the first approved drug that was mainly designed by computer. (inference): that's about 11 years from design start, so the clinic, not the computer, still sets the clock.

Part 2 — AI-designed drugs in the pipeline (Oct 2026)

Industry trackers count ~170+ AI-derived programs in clinical trials (~94 in Phase 1, ~56 in Phase 2, ~15 in Phase 3). These are the ones that matter for stocks:

What to watch for traders (editorial):
1. Zasocitinib's FDA decision (Q1 2027) is the sector's first "an AI/computational drug got approved" headline. It's a sentiment catalyst for SDGR and the whole group, even though SDGR's direct economics are small (it gets distributions from Nimbus sales milestones, not a big royalty).
2. Rentosertib Phase 3 is the first test of whether an AI-picked target holds up at scale.
3. Isomorphic's first patient dosed is a GOOGL headline, but it moves the small-cap AI-bio names on sympathy.

Part 3 — Who owns the tools and the data

The value chain has four layers. Whoever owns a scarce layer earns the money.

1. Compute & foundation models (picks and shovels)
- NVIDIA: BioNeMo / Clara are the default model frameworks. A $1B, 5-year co-innovation lab with Eli Lilly was announced Jan 2026 in South San Francisco, built on Vera Rubin. Nearly every AI biotech trains on NVIDIA hardware.
- Alphabet: AlphaFold (DeepMind) is the most-used structure tool in the field, and AlphaFold 3 powers Isomorphic Labs (raised $2.1B in 2026, led by Thrive). Partnerships with Lilly and Novartis.
- Microsoft / Amazon: Cloud and model hosting (Azure AI Foundry, AWS). These are enablers, not AI-bio pure plays.

2. Proprietary data (the real moat)
Models are becoming commodities. Data no one else has is the scarce input.
- Eli Lilly: Decades of proprietary assay data opened to biotechs through TuneLab. It's the biggest pharma spender on AI infrastructure.
- Tempus AI: Probably the largest linked clinical + genomic (multimodal) oncology dataset. Data licensing grew +36% YoY in Q2 2026, with 70+ pharma data customers (AZ, GSK, BMS, Pfizer, Novartis, Merck). It built an oncology foundation model for AstraZeneca. 2026 revenue guide is ~$1.6B.
- Recursion: One of the largest proprietary phenomics (cell-image) datasets, plus its own supercomputer (BioHive).
- IQVIA: The largest real-world / prescription / claims dataset plus clinical trial operations.
- Veeva: Owns the software where pharma's clinical and commercial data lives (Vault CRM, clinical platform).
- Illumina / 10x Genomics: They make the sequencers and single-cell instruments that generate the omics data the models eat.

3. Design software (the "CAD for molecules")
- Schrödinger: Physics + ML (FEP+), used by most large pharma. It's moving to a hosted/SaaS model (hosted = 47% of software revenue). An agentic platform called "Bunsen" is due late 2026.
- Certara, Simulations Plus: Biosimulation (PK/PD modeling) used in nearly every regulatory filing.
- Dassault Systèmes (BIOVIA, Medidata): trades as DASTY OTC / DSY.PA.

4. AI-native drug makers (own the molecules)
Insilico, Generate, Recursion, Relay, Absci, AbCellera, XtalPi, plus private ones: Isomorphic, Xaira, Chai Discovery, Iambic.

Part 4 — Ticker list, ranked by importance to the theme (editorial)

The ranking weighs how central each company is to AI drug discovery (owns the scarce layer, leads the clinic) and its pure-play exposure. It is not a buy list. Big tech is ranked high for influence, but AI-bio is a tiny share of its revenue.

Private leaders to watch for IPOs: Isomorphic Labs (Alphabet), Xaira Therapeutics, Chai Discovery, Iambic Therapeutics.

Part 5 — Why diagnostics trade like tools (and the tools scorecard)

Based on a BullishAgent study of 54 AI-bio, tools and diagnostics names, measuring the run from Jul 1 to Oct 5, 2026.

Why diagnostics behave like tools

They share the same economics and the same investors:

1. Both sell a product with repeat revenue. Tools companies sell instruments plus consumables. Diagnostics companies sell tests. Both are valued on revenue multiples, not on drug-approval odds. In the study they landed in the same "real revenue" group, and their medians are close: diagnostics +41%, tools +37%.
2. Both produce the patient data AI models need. Many diagnostics tests run on sequencing (Illumina's machines). Each GRAL, GH or NTRA test produces paired genomic and clinical data. GH and NTRA already license that data to pharma, the same business Tempus has.
3. Same 2021 history. Both groups were pandemic and genomics-bubble darlings in 2021, crashed 50–95%, and now trade with the same analysts and funds. When money rotates back into "genomics", it buys both.

The difference: diagnostics have extra catalysts that tools don't, such as FDA decisions and Medicare coverage. That's why the two biggest diagnostics runs were their own stories: GRAL had a clean FDA advisory-panel briefing on Galleri, and CDNA had revenue +52% with raised guidance. VCYT (−19%) and GH (+25%) show diagnostics don't all move together. The diagnostics sample is also small (6 names), so treat the +41% median loosely.

Tools tickers scorecard (run from Jul 1 to Oct 5, 2026)

1. Data generators (sequencing, single-cell, spatial, proteomics, DNA synthesis). These feed AI models directly. Median +40%.

2. Bioprocess (equipment and materials for making drugs, not data). Median +38%.

3. Diversified large caps. Median +29%, the slowest.

What the tools group shows

• The closer to AI data, the harder the run. Data generators beat bioprocess, and bioprocess beat the diversified large caps.
• Mid-caps ($2–12B) with real revenue led: TXG, TWST, AZTA, ADPT.
• Late-innings signal: the near-zero-revenue micro-caps are ripping in the last month, much faster than their 3-month numbers: PACB +121%, NAUT +103%, QSI +68%. Recursion's +16% day on 3× volume (Oct 5) is the same pattern among the AI drug developers. Money is now reaching the weakest names. (inference) That usually marks the late stage of a theme.
• Catch-up candidate in tools: BRKR. It's a spatial-biology player in TXG's lane. It has been basing in a rising range since June (higher lows: $48.63 Aug 5 → $51.54 Sep 14 → $59.17 Sep 24). A close above the Sep 30 high of $66.98 would break it out. TECH is not a candidate: Merck KGaA is buying it for $73 cash. The acquisition itself shows the same thesis, though: a strategic buyer paid a 36% premium for a spatial-biology and proteins franchise.

Bottom line

1. AI compresses discovery, not trials. The 4–6 year discovery phase is becoming 1.5–3 years. The 6–8 year clinic is barely changed. Any company promising an "18-month drug" means the time to a candidate, not to approval.
2. The scorecard is Phase 2/3, not Phase 1. High Phase 1 success is real but expected. The proof comes from rentosertib's Phase 3, REC-4881's registrational path, and GB-0895's Phase 3.
3. The 2027 catalyst: the zasocitinib FDA decision (Q1 2027) will likely be the sector's first approval headline.
4. The money in data and tools is steadier than the money in molecules (editorial). NVDA, GOOGL, LLY, TEM and SDGR earn whether any single AI drug works. The pure-play drug makers (RXRX, ABSI, GENB, RLAY) are binary biotech bets wearing an AI label, so value them with the biotech guide, not as tech stocks.

Sources
- Insilico: Phase 3 GENESIS-IPF-3 first patient dosed (EurekAlert) · Drug Target Review · Insilico 1H26 results / HK listing
- Takeda: FDA accepts zasocitinib NDA, priority review (Business Wire via Morningstar) · Fierce Biotech
- Recursion Q1 2026 update · Recursion Q2 2026
- Generate Biomedicines IPO pricing · Renaissance Capital
- Absci ABS-201 interim Phase 1
- Relay 2026 milestones
- Isomorphic Labs trial timeline (Reuters via Investing.com) · $2.1B raise (R&D World)
- NVIDIA–Lilly co-innovation lab
- Tempus Q2 2026 results
- Schrödinger Q2 2026
- XtalPi–DoveTree payments (HKEX)
- AI-discovered drugs in trials 2026 (IntuitionLabs)
- FDA draft AI guidance & credibility framework
- China NMPA 30-day IND pathway (Covington) · NMPA
- GRAIL FDA briefing docs (TIKR) · CareDx guidance raise (Tiger) · Twist / Lilly TuneLab (StocksToTrade) · 10x Genomics AI spatial collaborations (Simply Wall St) · Merck KGaA to acquire Bio-Techne (RTTNews)