Companies Applying AI to De-Risk Clinical Trials: 2026 Watchlist
Clinical development is still shaped by a stubborn asymmetry: only around one in ten drug candidates that enter human testing ever reaches approval, and Phase II/III failures remain the biggest drivers of wasted R&D spend.
Despite growing use of de-risking strategies such as genetics-anchored target selection, increasing use of biomarker-stratified and enrichment designs, and innovative designs such as adaptive enrichment and master protocols, late-stage failure rates are still high with only a small minority of Phase I assets ultimately reaching approval.
Against that backdrop, AI enters the picture as a way to leverage data that these methods leave on the table: integrating heterogeneous clinical and real-world data at scale, identifying complex patterns in patient response and site performance that are difficult to capture with low-dimensional statistical models, and updating risk estimates continuously as new data arrive.
De-risking clinical trials with AI plays out across three layers: how studies are designed, how control evidence is assembled, and how trials are run and analyzed operationally.
A few examples across these three layers would be:
- Design: Medidata’s Protocol Optimization and Intelligent Trials use predictive models trained on 36,000+ trials and 11 million patients to simulate how protocol choices affect burden, enrollment, site performance, and cost, and to flag enrollment and diversity risks using near-real-time benchmarks from ~8,000 active studies.
- Assemble evidence: Unlearn’s disease-specific “Digital Twin Generators” forecast each participant’s expected trajectory under standard of care and plug those forecasts into the analysis; retrospective work with AbbVie and J&J on Alzheimer’s Phase III datasets reports 19–33% reductions in required control-arm size, with similar ideas extended to Phase II designs.
- Run: ConcertAI reports using real-world oncology datasets plus AI to forecast site-level enrollment and diversity and to cut unplanned Phase II/III amendments, with internal benchmarks claiming roughly 50% fewer substantial amendments and several months saved per study when designs are iterated in silico first.
Another notable example of a company that sits squarely in the de-risking niche is London-based clinical-stage AI company, spun out of UCL and King’s College London, Hologen, co-founded by former Google CEO Eric Schmidt, sits squarely in this late-stage de-risking niche. It develops “Large Medicine Models”—large-scale, multi-modal generative models trained on real-world clinical and investigational data—to model individual disease trajectories and treatment response, with the stated aim of reducing patient enrollment, shortening timelines, and increasing statistical power in Phase II/III trials.
Hologen has made some recent news. Through a joint venture with MeiraGTx on the AAV-GAD gene therapy for Parkinson’s, Hologen has committed up to $430 million and is licensing these models to analyze prior trial and imaging datasets, refine the planned Phase 3 design, and, according to partner disclosures, increase the robustness and perceived probability of success of the program. Now, the company is seeking a $150 million Series A round at an approximate $850 million valuation to scale its Large Medicine Models as a general platform for de-risking late-stage clinical programs and associated diagnostics.
Here are some additional names worth tracking into 2026:
Design & Feasibility
Axiom Bio
Axiom Bio is a San Francisco–based biotech founded in late 2023 by Brandon White and Alex Beatson. The company focuses on predicting human drug toxicity—initially drug-induced liver injury (DILI), a leading cause of late-stage clinical failure and post-market withdrawal, with AI models trained on primary human liver biology. The explicit aim is to catch human-relevant toxicity before Phase II/III, reducing the fraction of trials and marketed drugs that fail or are pulled for liver safety reasons.
Axiom runs an automated wet lab that has generated what it describes as the world’s largest primary human liver dataset: >130,000 unique small molecules screened in primary hepatocytes, with high-content imaging, mechanistic readouts (mitochondrial toxicity, ER stress, ROS, cytotoxicity), and linked clinical exposure/toxicity annotations for 10,000+ molecules. Those data power structure-based ML models that estimate liver toxicity at clinically relevant exposures using structure and other modelled inputs. In April 2025, Axiom raised an early $15 million round to scale AI models for human liver toxicity prediction, with a focus on replacing parts of animal testing.
QuantHealth
Tel Aviv- and Cambridge-based QuantHealth was founded in 2020 by Orr Inbar and Arnon Horev. It focuses on simulating clinical trials in silico so sponsors can pressure-test trial designs and portfolio decisions before going into patients. Its QuantHealth Platform trains prediction models on hundreds of millions of longitudinal patient records plus biomedical knowledge graphs to forecast individual patient trajectories and virtual trial outcomes, then compares thousands of protocol variants to optimize inclusion criteria, endpoints, and sample size.
In 2025 the company reported having simulated more than 350 trials with up to ~90% predictive accuracy and about 31.4 million dollars in savings for a top-10 pharma client, with reported deployment across 23 therapeutic areas including oncology, immunology, cardiometabolic disease, and gastroenterology. Later that year Sanofi Ventures made a strategic investment and entered a collaboration to use QuantHealth’s digital-twin simulations to support Sanofi’s pipeline, bringing total funding to about 30 million dollars.
TriNetX
TriNetX, founded in 2013 and headquartered in Cambridge, Massachusetts, operates a federated global research network that connects EHR data from thousands of provider sites to life-science sponsors for feasibility, protocol design, recruitment, and real-world evidence (RWE). Its core product, the TriNetX LIVE platform, aggregates de-identified EHR, lab, and registry data from more than 8,000 provider locations and over 250 million patient records into a common data model, then applies AI and analytics so sponsors can run cohort queries, assess inclusion/exclusion impact, and explore site-level performance.
TriNetX markets the platform as reducing trial design cycles and protocol amendments by more than 20 percent, identifying over 50 percent more sites with eligible patients, and lowering the share of non-enrolling sites by roughly 20 percent, which directly targets feasibility and recruitment risk. In 2025 the company highlighted broad adoption of real-world data (RWD) and AI across biopharma in a survey and published content on using its network to de-risk clinical phases, and in early 2026 it rolled out a conversational AI interface and enhanced APIs that let teams pose natural-language questions and receive real-time feasibility and site-intelligence outputs inside existing workflows.
RWD-driven Evidence & Controls
Flatiron Health
Flatiron Health is a New York-based oncology data and software company founded in 2012 by Nat Turner and Zach Weinberg; it has been an independent affiliate of Roche since its acquisition in 2018. It curates longitudinal real-world datasets from oncology EHR systems and layers analytics, machine-learning, and workflow tools on top, with a focus on evidence generation and trial optimisation. The US database covers more than 5 million patient records from around 280 oncology practices and 800 sites of care, with structured and unstructured data abstracted into disease-specific datasets.
Technically, Flatiron’s infrastructure underpins a growing body of work on external control arms and trial emulation in oncology, where EHR-derived cohorts are used to mirror control arms of pivotal trials or build comparator groups for single-arm studies. In 2024–25 the company extended its network beyond the US by harmonising EHR-derived oncology datasets across the US, UK, Germany, and Japan and deploying a secure “trusted research environment” for cross-border patient-level analyses, enabling multi-country feasibility assessments and external-control work on a single platform. It also introduced six new hematology datasets covering over half a million blood-cancer patients, which increases the depth of data available for de-risking hematology trials via historical comparators and granular outcome benchmarks.
Owkin
Owkin is a French-American AI biotech founded in Paris in 2016 by oncologist Thomas Clozel and machine-learning researcher Gilles Wainrib, with offices today across Europe and the US. It focuses on using multimodal patient data from academic hospitals and research networks to identify new targets, optimize clinical trials, and develop AI diagnostics, underpinned by techniques such as federated learning so models are trained across sites without pooling raw patient data. Its platform suite includes K1.0, a proprietary operating system that integrates multimodal patient data and Owkin’s AI engines to build prognostic models, discover biomarkers, and power features such as AI-based external control arms and digital twins.
In May 2024 Owkin launched an AI-guided pipeline for cancer and immune diseases and licensed OKN4395 from Idorsia, a drug that blocks two inflammation-linked receptors and was already ready for human testing. In 2025 it dosed the first patient in a Phase I trial of OKN4395 in solid tumors, describing the study as AI-optimized, with its K1.0 platform guiding indication choice, building an AI-based external control framework, and defining biomarker-driven patient subgroups.
Owkin is also still embedded in big-pharma pipelines. With Sanofi, it is extending the use of its AI models from oncology into immunology, using multimodal patient data to choose which immune diseases and patient subgroups existing drugs are best suited for. A multi-year alliance with Bristol Myers Squibb applies Owkin’s platform to cardiovascular trials, where the partners are using machine learning on trial and real-world data to refine endpoints, identify high-risk patient subgroups, and tune study design.
Execution & Operations
AiCure
AiCure is a Houston-based company, founded in 2010 by Adam Hanina, Laura Shafner, and Gordon Kessler, originally backed by NIH and institutional investors. Rather than trial design or feasibility, it focuses on de-risking trials at the patient-behavior layer through AI-driven remote monitoring, medication-adherence measurement, and digital biomarker generation. Its H.Code platform uses smartphone-based AI computer vision to confirm dosing events, track engagement and extract behavioural signals, layered with predictive analytics and site dashboards so clinical teams can identify at-risk patients and intervene before adherence problems escalate.
In 2024, AiCure formally launched H.Code at DPHARM, published an overview fact sheet positioning it as a patient-engagement AI for trials, and featured the platform in industry events focused on using AI to improve adherence, compliance, and retention, all framed around making timelines and outcomes more predictable by stabilising patient-level data quality. Recent peer-reviewed studies and case work have used AiCure data to train machine-learning models that predict adherence in schizophrenia and attenuated psychosis trials, and report higher cumulative adherence than with standard monitoring alone, which directly addresses one of the silent failure modes in Phase II/III.
Tempus AI
Tempus, headquartered in Chicago and founded in 2015 by Eric Lefkofsky, builds a multimodal data and AI stack around oncology and other disease areas, spanning diagnostics, decision support, and clinical trial enablement. Its clinical development offering centres on the TIME program, which combines an AI-enabled trial matching platform (TApp), a curated trial portfolio, and rapid activation workflows that plug directly into provider EMRs. The matching engine uses natural language processing and rules over EMR-derived structured and unstructured clinical data, including information such as next-generation sequencing results, to identify trial-eligible patients at scale.
An ASCO-linked analysis reported that TIME network used algorithmic screening plus nurse review to scan more than 1.28 million patients across 94 trials in one year for trial eligibility, generating 573 consents in a single year. Tempus also went public on Nasdaq in June 2024, raising about 411 million dollars and reaching a fully diluted valuation above 6 billion dollars, which has funded further expansion of its trial-matching and data products. In 2025 it acquired Deep 6 AI, adding links to more than 30 million additional patients and over 750 provider locations, which strengthens its ability to de-risk feasibility and recruitment at the network level.
Currently, Tempus is also engaged in a multi-year collaboration with AstraZeneca and Pathos AI to build a large multimodal oncology foundation model trained on its de-identified cancer datasets.
For ongoing coverage of this space, subscribe to our weekly newsletter, Where Tech Meets Bio, that tracks standout developments at the intersection of life sciences and technology, with concise updates on new platforms, startups, partnerships, and investment moves in pharma, biotech, and healthcare.
Topic: AI in Bio
