AI Drug Discovery Platforms in 2026: The Honest State of the Pipeline

In 2026 the AI drug discovery story is no longer the breathless deck from 2019. A handful of companies have moved AI-designed molecules into late clinical stages, a handful have flamed out spectacularly, and a much larger contingent sits in the awkward middle where the chemistry works but the biology stays cruel. The honest read is more useful than either the techno-optimist or the cynic version.

Insilico Medicine has an AI-discovered, AI-designed fibrosis candidate (rentosertib, formerly INS018_055) reading out from Phase 2 trials — the first clinical molecule whose target and structure both came from an AI pipeline. Recursion’s pipeline has multiple Phase 2 readouts across rare-disease indications. Schrödinger continues to compound revenue on the software-plus-pipeline model. AlphaFold3 changed the protein-structure ground state for the whole industry. BenevolentAI restructured after disappointing early trial results. Exscientia merged with Recursion in 2024. The map is messier and more credible than the 2019 version, which is the honest mark of maturity.

The vendor and pipeline landscape#

Insilico Medicine is the clearest case study of end-to-end AI drug discovery — their Pharma.AI platform identifies targets via the PandaOmics module, designs candidate molecules via Chemistry42, and predicts clinical-trial outcomes via inClinico. Rentosertib (an AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis) entered Phase 2 in 2023 and remains the headline asset. Multiple additional INDs cleared since.

Recursion (now incorporating Exscientia) runs the highest-throughput phenotypic screening operation in the industry — millions of cell-imaging experiments per week, with neural networks learning the embedding space that maps perturbations to phenotypes. Multiple Phase 2 assets across rare diseases (CCM, FAP, NF2) with mixed but informative readouts. The Roche and Bayer collaborations remain active.

Schrödinger is the steady-compounder of the group — physics-based plus ML structure-based drug design as a software business, with an internal pipeline funded by software revenue. Their MM-GBSA, FEP+, and active-learning workflows are deeply embedded in big-pharma discovery groups. Several partnered programmes have progressed into clinical stages.

Atomwise ships virtual-screening services around their AtomNet structure-based prediction model, with a large global partner-program footprint.

BenevolentAI had a difficult 2023–2024 — a high-profile Phase 2 atopic-dermatitis failure (BEN-2293) triggered a restructuring. The cautionary tale of the cohort: target identification from literature graphs is not the same as having a winning molecule for that target. The company persists with a narrower focus.

Isomorphic Labs — Alphabet’s drug-discovery spinout from DeepMind — anchored multi-year deals with Novartis and Eli Lilly in 2024, leveraging AlphaFold-derived structure prediction and a proprietary generative chemistry stack. No clinical-stage assets disclosed as of 2026 but the platform partnerships are well-funded.

Generate Biomedicines, Absci, Cradle, EvolutionaryScale (ESM3) anchor the protein-design and antibody-engineering corner — generative biology for therapeutic protein design, with partnerships into Moderna, BMS, Genentech, and a long tail of biotech deals.

Owkin sits in the federated-learning-for-pharma niche, working primarily with academic medical centres on multi-modal patient data.

What the technical architecture really looks like#

There are five distinct technical workstreams, often confused in marketing decks.

Target identification. Graph neural networks over biological knowledge graphs (gene-protein-disease relationships), causal inference over multi-omics data, single-cell transcriptomics-driven target nomination. Outputs are ranked target hypotheses.

Structure prediction. AlphaFold2 changed the field in 2021; AlphaFold3 (2024) extended to protein-ligand and protein-nucleic-acid complexes with meaningful accuracy gains on previously hard targets. Open-source variants (OpenFold, ESMFold, RoseTTAFold All-Atom) sit alongside.

Generative chemistry. Transformer or diffusion-model architectures that propose novel small-molecule candidates against a target structure or activity profile. Reinforcement-learning loops with predicted properties (potency, ADMET, synthesizability) drive the optimization.

Phenotypic screening and embedding. Recursion’s approach — image cells under thousands of perturbations, learn an embedding, find similarities to known-effective compounds. Compatible with target-agnostic discovery.

Clinical-trial prediction. Modeling probability of trial success, patient response heterogeneity, optimal trial-design parameters. Less mature than the chemistry side; the data is sparser and noisier.

The unsexy reality is that physics-based methods (free-energy perturbation, molecular dynamics) and structure-based design (docking, pharmacophore search) still do enormous amounts of the heavy lifting. AI is augmenting, ranking, and accelerating; it has not replaced the underlying chemistry stack.

The clinical-pipeline reality in 2026#

Multiple AI-derived clinical-stage assets exist; few have completed Phase 3; commercial approvals of explicitly AI-discovered molecules remain pending. The two-track honest reading:

The optimistic track. Time-to-IND for AI-discovered programmes is meaningfully shorter than industry baselines, and several Phase 2 readouts have been positive enough to justify continued investment. The AlphaFold-class advances unlocked previously intractable structural biology.

The realistic track. Clinical-stage attrition rates for AI-discovered molecules are converging on industry baselines as the dataset matures. The promise of “ten-times-fewer failed trials” has not materialized. The biology remains the bottleneck — bad targets fail in the clinic regardless of how elegantly the molecule was designed.

Pharma partnerships and the BD model#

Almost every AI drug discovery company runs a dual model: internal pipeline plus pharma partnerships. Insilico-Sanofi, Recursion-Roche, Recursion-Bayer, Schrödinger-multiple, Isomorphic-Novartis-Lilly, Generate-Amgen, Atomwise-multiple. The partnerships fund the platform investment and validate the technology with the buyers that ultimately matter.

The structure of the deals has matured. Early-2020s milestones were heavy on platform-access payments and light on per-asset value; the 2025+ deals look more like traditional biotech licensing — meaningful upfront, larger per-asset milestones, royalties on commercialization, with the AI vendor often retaining co-development rights on selected programmes.

Where pdpspectra fits#

Discovery-platform builders need ML-Ops infrastructure that handles enormous data volumes, GPU-fleet orchestration, reproducible experiment tracking, and audit-grade lineage for regulatory submissions. We help discovery groups in pharma, biotech, and CRO settings build the operational backbone — data lakehouses for multi-omics and assay data, GPU scheduling, model registries, lineage from raw read to candidate molecule. See our ML and MLOps practice.

Related reading#

• AI in clinical trials in 2026
• Healthcare AI playbook: pilot to production
• Germany pharma: Bayer, Boehringer, Merck

If you are building or buying drug-discovery AI infrastructure and want a vendor-neutral read on the platform-versus-pipeline trade-offs, reach out.