ArcaScience, operating from offices in Paris and Boston, has introduced Flow, a specialized AI system designed to help pharmaceutical organizations evaluate the benefit-risk balance of drug candidates. Alongside the product launch, the company has released benchmark data suggesting its tailored approach identifies substantially more adverse-reaction evidence than general-purpose frontier models while requiring significantly lower computational costs.

The startup, which Europa Now has tracked since its seed funding round last year, is now taking its most significant product step forward. The launch tests a hypothesis gaining attention across the AI industry: in regulated sectors with high stakes, a specialized system designed for a specific task may outperform the largest general-purpose models.

The importance of benefit-risk assessment in drug development

The journey from laboratory to pharmacy shelf typically spans more than a decade. Roughly nine out of ten drug candidates never reach approval, and the financial investment required to bring a single drug to market averages approximately $2.3 billion. Romain Clément, the company's founder, has articulated the case for identifying these challenges at earlier stages. Regulatory bodies including the FDA and EMA mandate that pharmaceutical companies demonstrate a new drug's advantages exceed its drawbacks relative to currently available treatments.

The necessary evidence exists scattered across clinical trials, scientific literature and adverse event databases, making it difficult to weigh different scenarios or identify where a candidate shows vulnerability. ArcaScience contends that numerous development failures stem from suboptimal positioning: a molecule with genuine potential directed toward the wrong patient population, or safety concerns emerging too late in the development timeline.

How Flow operates

Flow integrates 24 specialized AI models with access to what the company describes as 100 billion biomedical data points. Safety and clinical teams can model various benefit-risk scenarios, identify missing evidence and determine what additional research to prioritize. The system refreshes its analysis as fresh trial data becomes available.

The platform represents an evolution from ArcaScience's previous approach of generating individual query reports, now offering interactive software capabilities. Beyond analysis, it can generate regulatory submissions, including PSUR/PBRER documents, Risk Management Plans and the benefit-risk components of eCTD Module 2.5. Each finding traces back to its original source material, and the company emphasizes that the same query consistently produces the same auditable results. This reproducibility addresses a recognized limitation of general-purpose LLMs and holds particular weight with regulatory authorities.

A promising molecule deserves the strongest possible development strategy

Romain Clément, CEO of ArcaScience

Clément's commitment stems from personal circumstances; the company emerged from his own journey as a cancer patient, as he shared in a conversation earlier this year.

Benchmark results: specialized versus general-purpose AI

Using its BRB-C v3.7 benchmark, ArcaScience measured its core engine against five general-purpose frontier LLM APIs. The company reports the following findings:

  • Recall: Flow's engine identified 104 out of 108 adverse-reaction concepts, achieving 96.3% accuracy. The highest-performing general model, Gemini 3.6 Flash, identified 70 concepts, or 64.8%.
  • Cost: For a scenario encompassing 30 million articles, ArcaScience estimates its cost at $0.0162 per article, totaling between $486K and $1.78M. General LLM APIs running the same dataset would cost between $1.35M and $18.53M, with Claude Opus 5 representing the upper end.

ArcaScience designed and executed the study independently. The consultancy inExtenso conducted a qualitative review. The complete methodology is available on ArcaScience's website for independent verification.

Client base and recent funding

ArcaScience reports working with over 20 pharmaceutical clients, an increase from 10 at the time of its $7 million seed round (approximately €6 million) one year prior. Its client roster includes Sanofi, AstraZeneca, GSK, Takeda and ICON, alongside the Paris Brain Institute (ICM). The French government engaged the company during the COVID-19 pandemic to organize scientific literature on the virus.

The September 2025 funding round was anchored by The Moon Venture, with participation from Pléiade Venture, Plug and Play Ventures, Bpifrance and AKKA Technologies. Flow is now accessible to pharmaceutical companies, biotech firms and regulatory organizations.

Why this matters

The technology press typically concentrates on increasingly large general-purpose models. ArcaScience is pursuing an alternative direction. Should its performance claims prove accurate, the qualities of determinism, traceability and cost efficiency at scale may ultimately carry greater weight with pharmaceutical buyers than raw model capacity.

Source: Startup Reporter