Anavex Life Sciences Corp (AVXL)

[Hosai]

Chat GPT did me a "for and against" ABCLEAR2/3 being a real effect rather than a false positive:

"1. Arguments Against ABCLEAR2/3 Being a Real Effect

A. COL24A1 was not pre-specified

The main legitimate concern:
SIGMAR1 (ABCLEAR1) was pre-specified.
COL24A1 (ABCLEAR2) was detected post-hoc via GWAS.
This raises:
multiple testing inflation concerns
risk of discovering a variant that correlates with treatment response by chance

B. Unknown number of variants and models tested

The companion GWAS manuscript is not yet public.
Therefore we cannot compute:
total SNPs tested
imputation models
interaction terms
LD pruning strategies
false discovery rate threshold
Until that appears, uncertainty remains.

C. Some endpoints appear stronger in 30 mg than 50 mg

This may reflect:
titration
tolerability
pharmacodynamic saturation
…but it also limits neat classical dose–response expectations and introduces noise.

2. Arguments Supporting ABCLEAR2/3 as a Real Biological Effect

(And these are unusually strong for a post-hoc discovery)

A. Multi-endpoint significance across all clinical domains
ABCLEAR2 and ABCLEAR3 were statistically significant on:
ADAS-Cog13 (primary cognition)
ADCS-ADL (primary function)
CDR-SB (secondary global clinical endpoint)
MMSE
CGI-I
NPI-Q
QoL-AD
MRI whole-brain and grey-matter atrophy

False positives rarely:
hit every endpoint,
with coherent directions,
and p-values often <0.01 or <0.001.
This breadth is atypical for noise.

B. Stepwise biological gradient: ITT > ABCLEAR1 > ABCLEAR2 > ABCLEAR3
The improvement is monotonic:
ITT: smallest effect
SIGMAR1 WT (pre-specified): larger
COL24A1 WT: larger still
Double WT (ABCLEAR3): largest
Noise does not typically produce a smooth, dose-like increase across three genetically distinct populations.
This structured hierarchy strongly favours a real biological interaction.

C. MRI results show the same genetic gradient
Imaging endpoints — whole brain and grey matter — also:
improve from ITT > ABCLEAR1 > ABCLEAR2 > ABCLEAR3
show 44–65% reductions in atrophy vs placebo
replicate the direction of clinical benefit
MRI is rater-independent, making this particularly compelling.

D. OLE durability supports non-randomness
ABCLEAR2/3 maintain or widen separation from natural history at 192 weeks.
A false-positive subgroup:
usually collapses over time
does not maintain a 3–4 year divergence
especially not on continuous measures like ADAS-Cog trajectories
Long-term stability strengthens the biological interpretation.

E. COL24A1 is biologically plausible
Its role in:
extracellular matrix (ECM) stability
autophagy-linked neuronal pathways
hippocampal/cortical expression
makes it mechanistically coherent with SIGMAR1-mediated autophagy restoration.
It is not a random gene irrelevant to neuronal biology.

F. Subgroup size is large, which supports stability
A large subgroup does not guarantee truth, but importantly:
false-positive subgroups tend to appear in small N
COL24A1 WT covers ~72% of participants
ABCLEAR3 still ~44%
Large Ns reduce the volatility typical of spurious findings.
This does not eliminate false positives — but does mitigate their likelihood.

G. PDD trial replication strengthens the pharmacogenomics
The Parkinson’s Disease Dementia trial showed:
significant dose-dependent improvement in episodic memory (p=0.003)
SIGMAR1 WT patients responded better
This replicates:
cognitive benefit with the same drug
pharmacogenomic response to SIGMAR1 WT
While COL24A1 wasn’t tested in PDD, SIGMAR1 replication alone strengthens the legitimacy of ABCLEAR3’s “double-WT” biology.

3. Probability Assessment — Is ABCLEAR2/3 Real?
Estimated probability ABCLEAR2/3 represents a real biological signal:
˜ 85–92%
Estimated probability the findings are a statistical false-positive:
˜ 8–15%
Reasoning:
non-pre-specification lowers confidence
but multi-endpoint consistency, MRI confirmation, OLE durability, large N, monotonic gradients, and biological plausibility push the probability strongly toward “real”"