Giving slightly more colour on the chances of COL being genuine vs false positive and the predicted effect size with probabilities:
What does “unbiased GWAS” really mean here?
The preprint explicitly says the GWAS was preplanned in the protocol and that COL24A1 WT was identified through that analysis.
That implies:
They decided in advance to run a broad genetic scan to look for response modifiers.
They did not pre-decide that COL24A1 would be the gene of interest – it emerged from the data.
We still don’t know the exact number of variants tested (that’ll be in the separate GWAS paper), but in standard GWAS terms:
A p-value around 10?5 is “suggestive”, not “genome-wide definitive” (that’s usually p < 5×10?8).
So on statistics alone, COL24A1 is interesting but not “case closed”. The reason it’s taken seriously is because of the combination of stats + biology + clinical pattern.
Why COL24A1 is much more than a random hit
a) Biology is very plausible
COL24A1 is:
a neuronal collagen
strongly expressed in hippocampus and cortex
part of the extracellular matrix (ECM) that shapes
protein clearance
tissue stiffness
synaptic environment
glial/microglial access to tissue
SIGMAR1 (blarcamesine’s pharmacological target) regulates:
autophagy / lysosomal function
proteostasis
ER stress & mitochondrial health
microglial over-activation
So a SIGMAR1 agonist improving intracellular clean-up plus a WT COL24A1 maintaining a more favourable ECM microenvironment is exactly the kind of mechanistic pairing you’d expect to matter in AD. COL24A1 is not a random gene you’d never think about – it sits in a very relevant pathway.
b) Clinical & MRI readouts are coherent
In the COL24A1/SIGMAR1 WT populations you see:
Bigger ADAS-Cog benefit
Bigger CDR-SB benefit
ADCS-ADL becoming clearly significant
QoL-AD improving
MRI atrophy slowing further
Random GWAS artefacts very rarely give you this many aligned signals across multiple independent outcome domains, especially in a ~500-patient trial.
Trial size context
The trial is much smaller than the big mAb Phase 3s, which actually makes very low p-values harder to achieve. You don’t get p˜10?4 to 10?5 across several endpoints in a subgroup this size just by accident very often.
Probability that COL24A1 is real, and how that maps onto effect size
Putting all that together (suggestive p-value, strong biology, coherent multi-endpoint pattern, modest n), a fair, rounded view is:
Roughly ~70–75% chance COL24A1 is a genuine treatment–response modifier,
~25–30% chance it’s mostly noise that shrinks or disappears on replication.
If COL24A1 is real, the big question becomes: how big is the effect once you pre-specify it and re-test?
Subgroup effects are almost always inflated in the first readout, so it’s safer to think in three effect bands + the null case:
1️⃣ Small but real effect (most likely)
Extra benefit over ITT: ~+0.3 to +0.6 ADAS-Cog points at 1 year
Total effect: –2.3 to –2.6 vs placebo
Probability: ~30–40%
This is the most “replication-friendly” level.
2️⃣ Moderate effect (best single central estimate)
Extra benefit: ~+0.7 to +1.2 points
Total effect: roughly –2.7 to –3.2 vs placebo
Probability: ~20–30%
If I had to put a pin in one number as a central guess, I’d land around +1 point of additional ADAS-Cog benefit in the double-WT population vs ITT.
3️⃣ High effect (similar to the initial ABCLEAR3 30 mg readout, least likely)
Extra benefit: ≥ +1.5 points
Total effect: ≤ –3.5 vs placebo
Probability: ~10–15%
Still on the table, but the least likely of the “real effect” scenarios.
4️⃣ No true COL24A1 effect (false positive)
Extra benefit: 0
Total effect: stays around –2.0 vs placebo
Probability: ~25–30%
That completes the probability picture in a way that respects both the stats and the biology.
How does this sit relative to the mAbs?
Leqembi (lecanemab) was initially refused by CHMP, then approved after re-examination with the indication restricted to patients with 0 or 1 copy of ApoE4 based on additional subgroup analyses showing lower ARIA risk in that group.
Kisunla (donanemab) followed a similar pattern: an initial negative CHMP opinion then a positive one after re-examination, again with the label restricted to patients with 0 or 1 ApoE4 copy.
In both cases:
the final approved population was not the broad all-comers population from the initial refusal,
and CHMP explicitly accepted subgroup-based narrowing of the indication after additional analyses at re-examination.
That doesn’t make COL24A1 “the same thing”, but it does show that:
CHMP is willing to base final labels on subgroup data that crystallise during the review / re-examination process, provided the overall benefit–risk is favourable.
For blarcamesine, the base ITT is already positive, safety is cleaner than mAbs, MRI is supportive, and COL24A1/SIGMAR1 is mechanistically coherent. That makes the ABCLEAR precision-medicine angle a reasonable part of a conditional-approval argument, not a wild stretch."
Market Data 
Markets 
AI
