Every value, traced to its source.

Three trial registries make up a dataset

Atlas datasets are built from ClinicalTrials.gov, the EU Clinical Trials Information System (CTIS) and the EU Clinical Trials Register. Each row shows what the registry record states, such as phase, recruitment status, key dates, enrollment, sponsors, conditions and interventions, and links back to that record. Company SEC filings add no study rows; they can add program names to a search, and PhaseFolio can group a program's names from them into one row, as set out below.

How records from different registries link

One trial often appears in more than one registry. Atlas groups those records as one study when a registry record states the other's trial number, such as an NCT, EU CT or EudraCT number, or when the link has been reviewed. Some ClinicalTrials.gov records state an EU CT number in a general identifier field, leave off its last two digits, or label it as a EudraCT number. Atlas reads those by fixed rules and links the record only when a single CTIS record matches. Sponsor protocol codes and similar titles are not used to link records.

When registries disagree

Atlas does the matching and keeps the disagreement in view. If one registry gives a trial as completed and another as ongoing, the study shows both values, each with its source, and says they differ. Records are never merged. No registry's value is overwritten by another's.

No result is not proof of absence

If no study matches, no record in the searched registries matched as of the data-as-of date. A trial can still be registered in a source Atlas has not connected yet, such as China's ChiCTR or the CDE drug trial registry, or be registered after that date.

Pinned, with earlier versions kept

A newer data release never changes a saved dataset. Updating it to newer data makes a new version, and earlier versions stay as they were, so a result in a memo can be reopened as it was.

Attribution

ClinicalTrials.gov data comes from the US National Library of Medicine, which does not warrant it, and no proprietary right is claimed in it. EU records come from the European Medicines Agency's public registers, and the agency does not endorse Atlas exports. Program names, agreements and press releases from company filings come from the US Securities and Exchange Commission's EDGAR system, and each is cited to its filing. Drug pages show records from Drugs@FDA, published by the US Food and Drug Administration, and from the European Medicines Agency's medicines catalogue; PhaseFolio captured and processed both and matches their products to a drug by the ingredient each record states, and neither agency endorses PhaseFolio. Every export carries each source's terms, its record links and the data-as-of date.

What a gap means

A value Atlas cannot show is labelled with the reason, never left blank, filled in or guessed.

Not collected yet

No source in the release supplies this field yet.

Not mapped

The registry gives a value Atlas has no category for. The value is kept, with the reason.

Hidden by source rights

The registry's terms do not allow Atlas to show the value.

Not searched

The data release does not read that field, so Atlas says nothing either way.

Withheld

The registry excludes the value, or PhaseFolio withholds a sponsor name at the request of the person named. The record's other sponsor names are still shown.

Programs a filing ties to a target

When a dataset searches interventions by a drug target PhaseFolio has reviewed, such as TL1A, Atlas adds the program names that company filings tie to that target. They arrive as one listed change, separate from the chat's answer, and you can undo it. Each name shows the company, the form, the filing date, one sentence quoted exactly as filed and a link to the filing on sec.gov. The chat model does not read filings.

Other names for the same program

When you open an intervention name in a dataset, and in CSV and XLSX exports, Atlas shows other names a company's SEC filing gives for the same program, with the sentence that states both and how many filings do. Atlas has not merged them: each name keeps its own studies, and this list combines no rows or counts. It lists only a name's own ties, never a tie of a tie, and leaves out names that several companies' filings use. Atlas groups names from filings only as program rows, set out below.

Program rows grouped from filings

On a dataset that searches a reviewed drug target, the board can show one row per program instead of one row per study. An export's Programs sheet carries the same rows, and a drug page lists its program's other names. PhaseFolio groups two names only when a reviewed tie's sentence states that both name one program: one name in brackets after the other, or with "formerly", "previously", "also known as" or "also referred to as" between them. A slash pair, a table row or two names only mentioned in the same sentence never group two names by themselves. Ties chain: if one filing ties A to B and another ties B to C, all three share a row, and each tie shows its own sentence. Where ties would put two names the review found are not one program into one row, that row's ties are held back for review and its names are not grouped. A brand name that Drugs@FDA or the EMA lists, or that the sentence tying it writes with a trademark sign straight after the name, is shown on its row but never joins two rows. A code the review found naming different programs at two companies never joins, and names that four or more companies' filings use are left out. Only companies that file with the SEC, and only reviewed targets, are covered. A program can still appear as more than one row where PhaseFolio has not grouped its names, including names a filing ties only in a form that does not group. A row counts studies by registered phase, not a program's stage.

How each tie is read and reviewed

PhaseFolio extracts each tie from the filing text. A language model proposes the tie and the sentence that states it, and code then checks that the sentence appears word for word in the filing; a tie whose sentence does not is set aside. Every remaining tie is read against its sentence, and against the filing where the sentence is unclear, then accepted or rejected with a recorded reason. This review is done by an AI reviewer working to a written guide. Atlas shows accepted ties only.

A filer's statement, as of its filing date

Each tie is what the company's filing stated on its filing date, not a fact PhaseFolio has verified, and a later filing can say otherwise. Only companies that file with the SEC are covered, so a target or drug with no names listed may still have programs or other names elsewhere. For program names, Atlas keeps no development stage and no filing text beyond the quoted sentence, and it keeps no personal contact details from any filing.

Deals on company and drug pages

A company page's Deals tab lists the companies that its own Forms 8-K and 6-K name as the other party to an agreement. Each row shows that company's name as written, as text and not a link, the kind of agreement beside the filing's own words for it, the drug or program and the territory as written, the filer's role, how many filings state it and the latest filing date. The kind of agreement and the filer's role are labels the automated reading gives. A row opens to its sentences, each quoted whole and exactly as filed, with the filer, form, exhibit, filing date and a link to the filing on sec.gov. A deal is listed on the filer's company page, not on the other company's. A drug page's Deals tab lists only statements whose program as written is one of the drug's names and whose sentence names both that program and the other company.

How agreements are read and accepted

Language models read the filings whose text mentions programs or deals several times, and propose statements that the filer has an agreement with another company, each with the text that states it. Code, not the model, then checks that the text appears word for word in the filing, and a statement whose text does not is set aside. Statements are not reviewed one by one. The reading was measured instead, on 157 filing documents that separate AI reviewers had labelled to a written guide without seeing its results. Where the reading and the labels disagreed, other AI reviewers settled it from the filing text alone, using the same model that reads the longest filings. About 94 in 100 of the statements it found named the right other company and kind of agreement. That is the lowest of the models' scores, and Atlas shows these statements only because each model scored at least 90 in 100, on at least 30 of its statements. The same documents were also used to choose the models and their settings, so the figure is not a fresh test. The drug or program, the territory and the filer's role were not part of that measure. On the same documents each model found only between about a third and about half of the agreements the reviewers found, so a company missing from a Deals tab may still be a partner, and a blank tab says nothing about whether a company has agreements.

How a deal row is grouped, and what is held back

Within one filer's filings, spellings of a company that differ only in case, punctuation, words such as Inc. or Ltd., or trailing words such as Pharmaceuticals or Therapeutics share a row. Trailing words are dropped only when a distinctive word of at least four letters remains. The row shows the most common spelling. Other spellings of one company, such as an abbreviation or a different regional name, get rows of their own, and two companies whose names differ only in those words would share one. Nothing is merged across filers, the other company is never matched to an Atlas company, and rows are listed newest filing first, which is not a ranking. A statement is shown only with a sentence that names the other company and can be shown whole and as filed. One whose sentence the filing's text breaks or that runs past 600 characters is held back and counted, and so is one whose sentence sits within quotation marks in the filing, speaks in the first person singular or comes from a call transcript, attributes a quote to someone with said, added, commented or stated, names a person with an honorific such as Dr. or an M.D. or Ph.D. set off by a comma, or uses a title such as chief executive officer. Atlas reads no deal amounts into fields: a sentence that states one is quoted whole, and nothing is summed, sorted or filtered by it. Atlas does not value or rank deals.

News from press releases filed with the SEC

A company page's News tab lists the press releases the company attached to its Forms 8-K and 6-K as Exhibit 99 since late 2021, newest first. Each shows its headline as written, its filing date, its opening sentence where a whole one is found, and a link to the exhibit on sec.gov. A release filed more than once is listed once: a company page shows it with that company's own filing, and a drug page with every filing. A drug page's News tab lists releases whose headline or opening sentence names one of the drug's names as whole words. Releases a company publishes only elsewhere, such as on its website or through a newswire, are not listed, nor are releases filed as the body of a Form 6-K rather than as an exhibit. Financial statements, presentations and other exhibits are left out.

How a press release is recognized

Code, not a model, decides which exhibits are press releases: one whose opening text holds a dateline with a title line before it, or a dateline after header lines only with a title line after it. The headline is that title line, joined only where the filing wraps it, and the opening sentence is the dateline's first sentence, kept whole or not at all. A separate AI reviewer that never saw the rule read the opening of 100 sampled exhibits from these companies. Every exhibit the rule called a press release was one, with its headline and the release's own date exact. From that sample, the rule is estimated to find about 86 in 100 of the press releases these companies filed as exhibits. A release is not listed when its headline uses words Atlas reads as about people, such as appointed, named, joins, succeeds, Dr. or a board or executive title. The rule goes by the words alone, so it can also leave out a release that is not about people. An opening sentence that uses those words, or attributes a quote to someone with said, added, commented or stated, is left out, and contact details are never shown. Headlines are the company's own words: Atlas does not judge the results they describe, and PhaseFolio never repeats them as its own statement.

Dated, covered and kept inside Atlas

Each tab shows its own date: for Deals, the latest filing date the reading reached, and for News, when filings were last checked. Only reviewed companies in Atlas with a US stock listing can be covered, and Atlas does not read every such company's filings yet; their pages say so. PhaseFolio matches each company to its SEC filer record through that listing. A separate AI reviewer, not told the matches, named the company behind each filer record, and every one agreed. A company with more than one SEC filer keeps each filer's rows under the name the SEC lists for it. Deals and News appear on signed-in company and drug pages only. They are not in exports, downloads, share pages, examples or signed-out pages, and the chat does not read them.