[12x contests this; he primarily relies upon the premise that the primary endpoint has failed (Section 5.4); he is, of course, referring to ADL in this early AD trial. I will respond to this circular argument in a separate post. However, to the extent he also cites Section 4.6 for the proposition that there must be both biological plausibility and replication, he is wrong. I found a similar statement to his describing "Credibility" in Section 4.3, but there is no conflict, anyway; the more factors supporting a finding, the more credible it is. That doesn't mean that achieving credibility always requires replication if other factors, such as statistical likelihood and biological plausibility, are present in sufficient strength.]
Can the EMA consider Anavex's ABCLEAR3 analysis upon re-examination? Yes
Per the EMA, "the re-examination procedure may deal only with the points of the opinion initially identified by the applicant and may be based only on the scientific data available when the Committee adopted the initial opinion."
The Anavex September 2025 PR on the ABCLEAR2/3 findings provides the details of how many trial participants fell into each of the ABCLEAR 2 and ABCLEAR3 subgroups. This means that Anavex, pursuant to its trial design, had the data on individual COL24A1 WT carrier performance in hand when the trial concluded. And this data certainly were provided to CHMP as part of the MAA submission. Since the MAA already included the raw data, it is already available to the committee. That establishes the predicate to submit the ABCLEAR2/3 data analysis upon re-examination. (Data analysis is just an examination of existing data; it is not itself new data. It therefore is not barred from EMA consideration.)
What is the EMA standard for subgroup data to be statistically strong enough to be considered?
The EMA may consider a subgroup if there is "biological plausibility for an inconsistent effect in the direction expected. Credibility is particularly strong if evidence is replicated across multiple data sources. In submissions with only one trial in which the subgroup can be properly assessed, the fact that there is no opportunity to find replication within the phase III clinical program cannot be taken as an argument to ignore the signal. . . ." Id. at p. 12.
Can Anavex demonstrate biological plausibility for its post hoc ABCLEAR2/3 data? Likely, Yes
The multiplicity issue arises as a result of discovering a post-hoc subgroup in a GWAS population. In the prepress article, the Anavex-connected authors reported in the post hoc, 30 mg ABCLEAR3 population the following p values:
ADAS-Cog13 .0004 ADCS-ADL .0024 CDR-SB <.0001
To address multiplicity, these results need to be adjusted according to the number of genotypes present in the survey that identified the COL24A1 WT gene that underlies ABCLEAR2/3. In its PR and its prepress article, Anavex described the GWAS as limited to a "genetic variant population[] related to the mechanism" of blarcamesine. ChatGTP has estimated for me that this probably represents a range of only 150 to 350 genes. Using the Bonferroni Correction, we multiply the p values above by the range of 150 to 350 times to handle multiplicity for this post hoc subgroup, resulting in the following estimated p values for the 2b/3's co-primaries and its major secondary endpoint:
ADAS-Cog13 .06 to .14 (near miss) ADCS-ADL .36 to .84 (not stat sig) CDR-SB <.015 to <.035 (stat sig)
If the ABCLEAR3 subgroup were a truly false positive result, it would not be expected to appear across related measurements, such as ADAS-Cog13 and CDR-SB, or mirror the p value profile of the dosed ITT population -- with strong results for those two endpoints and a much weaker (but not totally random) result for ADCS-ADL. It would not achieve stat sig, as corrected, in CDR-SB.
However, as Hosai has posted, his ChatGTP engine noted the similarities don't stop there. The 30 mg ABCLEAR2 and ABCLEAR3 uncorrected p values were statistically significant for MMSE, CGI-I, NPI-Q, QoL-AD, and MRI whole brain and gray matter atrophy reduction. Results were in the same direction across multiple domains. The AI engine notes a monotonic, "stepwise biological gradient" with ABCLEAR3 as the best performing group, then ABCLEAR2, then ABCLEAR1, and, finally, the 30 mg ITT population "in a smooth, dose-like increase across three genetically distinct populations. This structured hierarchy strongly favours a real biological interaction." The pattern repeats with the MRI results, which are measured and evaluated by machines and not researchers --- which is "particularly compelling," according to the engine.
As Hosai's post also notes, the ABCLEAR2/3 OLE results also support non-randomness, because "[a] false positive subgroup usually collapses over time [and] does not maintain a 3-4 year divergence especially not on continuous measures like ADAS-Cog trajectories. Long term stability strengthens the biological interpretation. . . . COL24A1 is biologically plausible" for "its role in extracellular matrix (ECM) stability autophagy-linked neuronal pathways [and] hippocampal/cortical expression makes it mechanistically coherent with SIGMAR1-mediated autophagy restoration."
The final factor Hosai references is the size of the ABCLEAR2 subgroup, which includes about 72 percent of trial participants. "Large Ns reduce the volatility typical of spurious findings" and mitigates their likelihood. I will add that the ABCLEAR3 subgroup, the most impressive, relies equally upon the pre specified S1R WT population, boosting its statistical claims.
We are travelling in a regulatory gray zone, a decisional limbo. If the EMA had a clear rule that ADL must be met in this trial -- and it wasn't -- we would have already withdrawn the application or received a negative opinion. Much more likely, the MAA would not have been submitted. Similarly, the same would be true if the EMA requires two trials and will not accept just one. But the application is still alive, so neither of these absolute propositions can be true.
Instead, we appear to be within the EMA guidance scenario from the document linked above that states "The clinical data presented are statistically persuasive in the primary analysis population but with therapeutic benefit or risk-efficacy which is borderline or unconvincing." In that context, CHMP is willing to receive more information on the benefit/efficacy side, such as biomarker data. In other words, if ADL had been hit, we would have a positive opinion, but with ADL missing, the EMA needs more. I believe that they will get it during the re-examination.
Market Data 
Markets 
AI
