Latrepirdine comes back: 3.36 ADAS-cog points on top of memantine, and nothing else moves

Latrepirdine is the molecule that taught the field how a promising Alzheimer's result can evaporate. A Russian antihistamine, repurposed, positive in a 183-patient trial published in the Lancet in 2008, then, as this paper's own introduction records, taken into a multinational study of 598 patients by Medivation and Pfizer and terminated midway for lack of effect. The commentary the authors cite for that episode is titled "Dimebon disappointment". Eighteen years after the first positive result, a Russian group has run it again in a new crystalline form, and reports a positive phase II. This is the issue's spine in a single object: a molecule with a spectacular hype cycle behind it, a real trial, and a gap between what the tables show and what the abstract says.

The rescue hypothesis, and what this trial does not test

The authors do not hide the failure. They state it, and then offer an explanation: none of the earlier studies took the polymorphism of latrepirdine into account, so the bioavailability of the material actually swallowed could have varied between trials. DMB-I is a new polymorphic modification with, in their previous work, higher availability in blood and brain and a reversal of scopolamine-induced impairment in rats. On that reasoning, the old failures tell you nothing about the new crystal.

It is a coherent hypothesis, and this trial does not test it. There is no arm receiving the older material, no head-to-head of the two polymorphic forms, and no pharmacokinetic comparison in these patients. The hypothesis also has to account for something it does not address: the 2008 Lancet result was positive using material that was equally uncharacterised, so uncontrolled polymorphism would have to explain both the earlier success and the later failure. Nothing here settles that. What we have is a fresh 26-week phase II, and it should be read on its own numbers rather than as a verdict on the history.

Note also what this trial cannot say about Alzheimer's specifically. The authors are candid: no cerebrospinal fluid biomarkers, no amyloid or tau imaging, diagnosis presumptive from cognitive assessment and MRI. By the 2024 revised criteria they cite themselves, this is probable Alzheimer's disease.

Exclusions were wide, and the authors name the cost themselves: the discussion lists cardiovascular disease including stroke, endocrine disorders such as diabetes, and moderate to severe depression among the study's limitations. The results section then describes exactly those conditions in the patients who were enrolled. Seventy-eight percent had vascular disease, with hypertension and atherosclerosis predominating; 44 percent had metabolic disorders, with dyslipidaemia, obesity and type 2 diabetes predominating; 41 percent had heart disease, with myocardial ischaemia and chronic heart failure predominating. This is the fourth place where the paper contradicts itself, and the other three, the abstract against the primary analysis, the abstract against Table 5, and the point estimates that disagree between text and tables, are set out below. Both accounts cannot hold for the same cohort. What follows from the discrepancy is narrower than it looks: the enrolled sample is not the comorbidity-free ideal the exclusion list implies, and which of the two descriptions is the accurate one cannot be established from the published text.

What the tables say

Baseline was well matched: ADAS-cog 33.98, 35.60 and 35.65 across the three arms (p = 0.6320), MMSE-2 about 18.8 in each. So these were mostly moderately impaired patients, and the ceiling for improvement was wide.

At 26 weeks the primary table gives a mean ADAS-cog change of -2.52 on 30 mg/day, -3.48 on 60 mg/day and -0.11 on placebo. Two things follow. First, the separation was produced by movement in the drug arms, not by decline in placebo: the placebo group finished where it started (p = 0.5740 for its own change). Second, only the 60 mg comparison cleared the pre-specified threshold. The 30 mg arm reached p = 0.0376 against an alpha of 0.0125, and its confidence interval crosses zero.

That distinction matters because the paper holds it unevenly. The abstract keeps it in one sentence, which reads that the efficacy of DMB-I at a dose of 60 mg/day was demonstrated by assessing the primary efficacy criterion, and loses it in the next, which reports significant improvement on ADAS-cog, CGI and IADL in the groups receiving both doses against placebo. The conclusion does not keep it at all: it states that treatment at 30 mg/day and 60 mg/day improved outcomes against placebo, without qualification. The wider claim rests on unadjusted secondary comparisons at 26 weeks, where the 30 mg arm reaches p = 0.0410 against placebo, rather than on the primary analysis, which explicitly did not reject the null for that dose. On the pre-specified test, one of the two doses failed.

The secondary measures are the more sobering half. MMSE-2 moved by 0.02, 0.63 and -0.41 points in the three arms and separated nowhere. QOL-AD is disposed of in a single sentence: neither the within-group nor the between-group analysis found significant differences, in absolute values or in change. No table is published for that scale, so no figures exist to check, and the strongest thing that can be said is that no significant difference was reported. IADL, the measure closest to what a family notices, showed a significant advantage for 60 mg over placebo at 12 weeks (p = 0.0472) and lost it by 26 weeks (p = 0.1190 between groups). The direction is worth stating exactly: at 26 weeks the two drug arms held level, +0.05 and +0.03 points on a scale running from 0 to 8, while placebo fell by 0.25, so what the drug shows on this measure is absence of decline, not a gain. The abstract nonetheless lists IADL among the scales significantly improved at both doses against placebo. Table 5 does not support that, and the table is the document I followed.

The clinician-rated global impression indices look like the one place something held, and they do not survive the standard applied above. The figures usually quoted from them, p = 0.0016 and p = 0.0006 for the CGI efficacy index and p = 0.0045 and p = 0.0005 for CGI improvement, are pairwise comparisons of absolute scores at week 26, unadjusted for multiplicity: precisely the class of comparison this piece has just refused to accept for the 30 mg arm on ADAS-cog. Applied to the change from baseline, which is what the primary endpoint measures, the between-group p for the CGI efficacy index is 0.1520, and the within-group change is significant only at 60 mg (p = 0.0425). CGI improvement does separate on change (between groups p = 0.0476), but pairwise against placebo only at 60 mg (p = 0.0468) and not at 30 mg (p = 0.1616). The clinician's rating of illness severity itself, CGI-S, did not move in any arm.

Safety was unremarkable and reported in reasonable detail: 145 adverse events in 61 of 133 patients in the safety population (45.86 percent), 135 of them mild, no severe or life-threatening events and no deaths. Overall incidence was 48.8 percent on 30 mg, 41.5 percent on 60 mg and 45.9 percent on placebo. Seventeen events in nine patients (6.77 percent) were dizziness, somnolence, lethargy and asthenia, which the authors classify as expected from the existing Dimebon label.

Two numbers in the paper disagree with each other and are worth naming rather than quietly dropping. The results text gives the 30 mg versus placebo point estimate as -2.52 with a confidence interval of -5.46 to 0.65, but Table 1 gives that difference as -2.41 with the same interval; -2.52 is the 30 mg arm's own change from baseline, and -2.52 minus -0.11 is -2.41, so the table is internally consistent and the sentence has copied the wrong figure into the wrong slot. Separately, Table 1 reports the 30 mg change as -2.52 in 46 patients while Table 2 reports the same quantity as -2.75 in 44. The 60 mg and placebo figures are identical across both tables, so the headline result is unaffected either way, and I have used the primary-analysis table throughout.

For your practice

Nothing here changes a prescription, and the trial does not ask it to: this is phase II, and the authors say a 450-patient, 52-week phase III at 12 Russian sites is what is meant to settle the question. The useful work is what you say when the question arrives in the room, because it will. Dimebon has a long memory in Russian families, it is cheap, and a positive trial in a Nature-portfolio journal will be quoted at you.

The honest summary is short. On top of memantine, at 60 mg/day, over six months, a cognitive test score separated from placebo by about three and a half points on a scale that runs from 0 to 82. Over the same six months, the patient's cognitive screening score did not change; no significant difference was reported on the quality-of-life questionnaire the patient answers about his own life, and no figures for it were published; and the ability to manage a telephone, medication and money separated from placebo at three months by holding level while placebo slipped, and no longer separated at six. A drug that improves a rating-scale total while daily function merely holds level and quality of life shows no reported change has not yet shown that it improves anyone's life, and this is not a criticism unique to latrepirdine: it is the standing problem with the whole symptomatic class the authors compare themselves to.

That gap is also where your work sits, and it is the second error this issue is about. While the argument runs over molecules, the things that did not move in this trial are exactly the things that are not pharmacological problems. Instrumental function decays through disuse and through families taking tasks over prematurely; quality of life in dementia responds to occupation, to social contact, to the caregiver's own state. None of that was the intervention here, and none of it is delivered by waiting for phase III. Read the trial as what it is, a modest positive signal that has to be confirmed, and treat the measures that stood still beside it, and the one whose numbers were never published, as an assignment rather than as bad luck.