Publisher Sage Licenses 400 Journals to Causaly's AI Drug-Discovery Platform

Academic publisher Sage has struck a licensing deal giving AI drug-discovery platform Causaly access to the full text of roughly 400 of its peer-reviewed journals — a notable step in how artificial intelligence is being wired into pharmaceutical research.

What the deal covers

Announced July 29, the agreement gives Causaly full-text access — methods, results, tables and supplemental data — from about 400 titles in Sage’s 1,500-plus portfolio, including biomedical titles from Mary Ann Liebert, which Sage acquired in 2025. It is Causaly’s first partnership with a major scientific-publishing house.

The inclusion of methods, tables and supplemental data is the substantive part. Abstracts are freely available and searchable; what they omit is exactly what determines whether a finding is usable — the sample size, the actual measurements, the conditions, the negative results reported in a supplementary file and mentioned nowhere else.

Why retrieval rather than training

This is a RAG (retrieval-augmented generation) deal: the content builds a searchable knowledge base the AI can cite, but it does not train Causaly’s underlying models.

The technical distinction has direct legal and practical consequences. In training, text is absorbed into model weights and cannot be removed or attributed afterwards. In retrieval, the content sits in a separate index; the model searches it, retrieves relevant passages and generates an answer grounded in them.

That means the licensed material remains identifiable, can be withdrawn if the agreement ends, and every claim can point to its source.

It is also what makes the deal possible. Publishers have litigated over unauthorised use of their content for model training, and a licence permitting retrieval while explicitly excluding training is the arrangement both sides can accept.

Why citation matters more here than elsewhere

Grounded, citable output is not a general nicety in this application — it is the requirement.

A researcher deciding whether to pursue a target must be able to check the underlying evidence. A generated assertion with no traceable source is not usable for that purpose regardless of whether it happens to be correct, because the researcher cannot evaluate the study design, sample or context.

Fabricated references are also a well-documented failure mode of generative models, and in scientific work a plausible-looking citation to a paper that does not exist is worse than no answer.

Retrieval addresses both: the model can only cite what is in the index.

How it is used

Causaly’s AI agents surface relevant biomedical evidence inside researchers’ workflows — a tool already used by a dozen of the top 20 global pharma companies, including Novo Nordisk, Novartis, Takeda and J&J.

The platform offers a “single platform that brings together evidence and governed scientific reasoning,” said Causaly COO Marco Costa.

The problem it addresses

Drug discovery runs on synthesising vast, scattered literature.

The volume is the obstacle. Biomedical publishing produces well over a million papers annually, and no researcher can read the relevant fraction. Evidence bearing on a single target is distributed across decades, disciplines and journals that a specialist in one field may never encounter.

The consequence is that relevant work is routinely missed — a protein studied in one context has implications for another that nobody connects, and programmes are pursued that earlier published evidence would have discouraged.

Retrieval across full text is a direct response to that, and it is a search problem more than a reasoning one.

What it signals

Licensing deals like this indicate a shift: high-quality, licensed scientific content is becoming a strategic input for AI tools, with publishers turning archives into a new revenue stream and drugmakers betting AI can compress research timelines.

The coverage problem this does not solve

Four hundred journals is substantial and it is a fraction of the biomedical literature, which bears on how such a tool should be used.

Scientific publishing is fragmented across many publishers, and the journals most central to a given field may belong to any of them — commercial houses, learned societies, or independent operations. A licence with one publisher covers what that publisher owns.

The practical consequence is that a comprehensive answer requires either many separate agreements or reliance on open-access material and abstracts for everything else. A researcher using the tool cannot easily tell which sources were available to it and which were not, and an evidence summary drawn from an incomplete corpus looks identical to one drawn from a complete one.

Open access changes the arithmetic over time. A growing share of new research is published under licences permitting reuse, and funder mandates continue to expand that share — so the gap narrows for recent literature while remaining wide for the decades of work published under subscription.

Which makes deals of this kind most valuable precisely for the archive, and least necessary for what is being published now.

For publishers the model has appeal beyond the payment. It monetises the archive without undermining subscriptions, and it establishes licensing as the expected route — which is worth more strategically than any single agreement. Business news, not investment advice.