Multiple Sclerosis Clinical Trial Landscape

Where the data comes from

Every heatmap is built from the public ClinicalTrials.gov registry, via its official API — interventional drug and biologic trials with a start date of 2008 or later. The master index holds over 145,000 trials and is refreshed weekly (see the “updated” date on this page). A disease landscape draws only from the active, Phase 1–3, industry-sponsored slice of that index.

• In scope: industry-sponsored trials in Phase 1, 2, or 3, with an active status (recruiting, active-not-recruiting, not-yet-recruiting, or enrolling by invitation). Phase 4 sits in the index but is left out of the landscapes.
• Filtered out: deeply stale programs (a primary completion date more than two years past with no update to completed or terminated); basket trials and incidental mentions (a trial counts toward a disease only when that disease is genuinely the subject of study — not a secondary cohort, an organ-of-origin overlap, or a passing mention); and healthy-volunteer studies.

We do not exclude trials by sponsor geography. Where a sponsor is based in China, the program is flagged on the page rather than hidden, so you can weigh it yourself. An automated test fails the weekly refresh if the underlying index is more than 14 days old, so a published grid is never built on a stale index.

How a trial is matched to a disease

Matching uses a structured medical ontology, not keyword guessing, and is designed so that no trial is ever silently dropped — every trial that clears the filters gets a classification, even if that is just “Other.” It runs as an ordered sequence of steps, stopping at the first that applies:

1. Healthy-volunteer studies are set aside as non-disease trials.
2. Ontology match — each tracked disease is linked to its official identifiers in the standard medical taxonomy (MeSH), so a trial can be matched even when its text uses a synonym.
3. Curated disease patterns — a hand-maintained library of over 150 disease-name patterns covers the more granular indications across oncology, hematology, infectious disease, cardiometabolic, immunology, and neuropsychiatry.
4. Basket guard — a trial matching four or more distinct diseases, or carrying explicit basket language (“tumor-agnostic,” “all solid tumors,” “pan-cancer”), is grouped into a single advanced-solid-tumor category rather than over-counted across every cancer it touches.
5. Therapeutic-area roll-up — a trial with no specific match, but which the taxonomy still places under a broad area, is assigned to that area (“Oncology — other,” “Immunology — other,” …), checking cancers first so a site-specific tumor isn’t filed under its anatomical system.

A “drop-if-parent-present” rule keeps a generic name from drowning out a subtype: a trial matching both lupus and lupus nephritis is reported only as lupus nephritis. Internal abbreviations are translated to the plain disease names used across the site (for example, “CRC” becomes “Colorectal Cancer”), and the same classifier is shared by every heatmap, so the same trial always maps to the same disease wherever it appears.

How a drug is matched to its mechanism

Mechanism of action is the hardest part to get right, so it is assigned in layers — leaning on curated and public data first, with AI as a last resort:

1. Curated rulebook (first). A rulebook we maintain — over 600 drug-to-mechanism rules — is checked first, matching on drug names, trial acronyms, sponsor trial identifiers, and intervention lists. First match wins, which stops a combination trial from being counted several times.
2. Public molecular-target data. Where no rule applies, each intervention’s target is looked up in a public target database, with verbose or gene-symbol labels normalized into consistent short forms so one target isn’t split across several columns.
3. Standard-of-care backbones. A small set of rules recognizes common combination scaffolds (checkpoint-inhibitor monotherapy, standard chemotherapy regimens, established standard-of-care agents) so they aren’t mistaken for the experimental arm.
4. AI as a last resort, then cross-checked. Only for genuinely opaque sponsor code-names that none of the first three steps can resolve do we ask an AI model to propose a mechanism — applied only above a fixed confidence bar, then automatically cross-checked against the sponsor’s own pipeline page. Where AI and the sponsor agree, the program is marked sponsor-verified. Where they contradict, the label is discarded entirely — not shown, not counted.

New mechanism rules are independently double-verified before they’re trusted — a second, adversarial pass set up to disprove the first — and each is checked so it can’t mislabel an unrelated trial. Drugs whose mechanism isn’t publicly disclosed are shown openly as “Emerging — not yet disclosed” rather than guessed at: for a tool meant to support real decisions, “we don’t yet know” is a more trustworthy answer than a confident guess.

Where AI is used — and where it isn’t

The disease and mechanism matching above is driven first by deterministic rules and public ontologies, not AI. AI plays three bounded, disclosed roles: (1) an optional extra check that a trial genuinely studies the disease, on top of the ontology match; (2) inferring a trial’s treatment setting on the competitive grids when the rules don’t cover it, only above a fixed confidence bar; and (3) the last-resort mechanism step above, always cross-checked against the sponsor’s disclosures. Wherever an AI label reaches a cell, the page marks it (⚙️ or ✅) — AI is never the silent, sole source of what you see.

What the on-page markers mean

• ✅ Sponsor-verified — AI proposed the mechanism and it matched the sponsor’s own pipeline page. High-trust.
• ⚙️ AI-classified — AI proposed it above the confidence bar but it has not yet been cross-checked against the sponsor. Useful; verify before citing. It never means a person reviewed it.
• ⚡ First-in-class — the mechanism hasn’t appeared in any other disease landscape we’ve built. This reflects the scope of landscapes published so far (the tooltip lists exactly which were scanned), not an absolute claim about the whole market.
• 🌱 First-in-indication — the only program competing on that mechanism within this disease.
• 🆕 NME candidate — the interventions match no drug in our approved-drug index, suggesting a new molecular entity. The index is incomplete — a signal, not a regulatory fact.
• 🔗 Combination · 👶 Pediatric · 🔥 Hot (readout within six months) · ⏳ Stale (completion date passed but still marked active — often a stalled program).

Sponsor names are resolved through a curated parent/subsidiary map; unrecognized sponsors appear under their raw registry name. The registry records the sponsor at a trial’s inception, so names are as originally filed and may not reflect later acquisitions. To keep large grids legible, mechanisms with a single program are listed separately rather than crowding the main grid, and very small players are listed below it — presentation choices only; nothing is removed from the underlying counts.

How we score programs — “what’s about to move”

Each program carries a 0–100 score that deliberately ranks imminence over raw stage — the most decision-relevant signal on a competitive grid. It is the sum of:

• Clinical phase — up to 40 points (Phase 3 = 40, Phase 2 = 25, Phase 1 = 10).
• Readout proximity — up to 60 points (next readout <6 months = 60, 6–12 months = 45, 1–2 years = 30, distant = 5).
• Stale penalty — the score is halved if a trial is past its expected readout but still listed as active.

Cell colour on the grid is driven by this score, so a Phase 2 program about to read out can — correctly — outrank a dormant Phase 3 one. It answers “what’s about to move,” not just “what’s furthest along.”

What each grid plots

• Indication landscape (this page) — one disease — companies (rows) × mechanism of action (columns): who is competing, and on what mechanism.
• Company portfolio — one company — diseases (rows) × mechanism (columns): where it is active, and what it is betting on.
• MOA platform — one mechanism family — drugs (rows) × diseases (columns): who is working on this class, and where.
• Competitive landscape — one disease — mechanism (rows) × clinical setting (columns), aggregated across companies; setting columns are tailored per disease (e.g. lines of therapy in oncology; biologic-naïve vs. biologic-experienced in IBD).

What we don’t claim

• First-in-class is editorial, not absolute — “not seen in the landscapes we’ve built,” not “novel across the industry.”
• NME candidate is a signal, not a filing — absent from our (incomplete) approved-drug index.
• Disease matching is automated and not exhaustively validated per disease — ontology and pattern matching can occasionally include or miss a trial.
• AI-classified mechanisms are machine-proposed — unconfirmed unless they also carry ✅.
• Sponsor names are as-filed and may lag current ownership.
• Grids are as fresh as their last rebuild from the weekly index — no faster continuous refresh is claimed.