June 1, 2026
Classical force fields gave us lecanemab. QSimulate is building the first-principles physics that might help us identify the knockout.
June is Alzheimer’s & Brain Awareness Month. Nearly seven million Americans are living with the disease. In April, a Cochrane Review of 17 clinical trials and more than 20,000 patients concluded that our best approved drugs provide virtually no clinically meaningful benefit.
You know the feeling if you’ve been in the room.
You’re presenting predictions to the project team. You’ve built a narrative — you have to, there’s a deadline, there’s a meeting, people need a direction. But only you know how wobbly the extra parameters were that you had to patch into the force field to make the system behave. Only you know that eight simulations ran and two failed in a way that is vaguely worrying, and you didn’t have time before the meeting to determine whether it’s noise or something systemic. So you focus on what’s scrutable. You show the results that cohere. You project confidence, because that’s what the room requires, and because the alternative — stopping everything to chase down the physics you can’t quite trust — isn’t really on offer.
I’ve been in that room. Most computational chemists have. What made those moments possible — and what made them inevitable — was the tool itself: classical force fields that approximated the physics we couldn’t yet afford to calculate properly.
A lot of Alzheimer’s drug discovery has been conducted on that kind of schedule, with that kind of model, by people who knew exactly what they were papering over and had no better option. That’s not a criticism — it’s a history. When I was involved in the early development of AMBER and Free Energy Perturbation methodology across the 1980s and 90s, the approximations weren’t failures of ambition. They were the state of the art, which is a different thing. The charge distributions were averaged. Polarization was ignored. The electrons, in any meaningful quantum mechanical sense, simply didn’t exist in our models. We built powerful, predictive tools on a foundation we knew was incomplete, and we made a promise to ourselves: when the hardware catches up, we come back and fix it.
The hardware has caught up. The promise is overdue. And for Alzheimer’s disease specifically, where three decades of clinical failures have left the field energized, humbled, and still searching, the cost of continuing with incomplete physics is no longer just methodological. It’s moral.
What Thirty Years Actually Proved
The honest state of the field is this: the Cochrane Review analyzed 17 clinical trials involving more than 20,000 patients and concluded that anti-amyloid drugs — including lecanemab and donanemab, the two that reached full regulatory approval of nine antibodies evaluated — provide no clinically meaningful benefit by the standards that matter to patients and families. Recent FDA approvals offered a genuine glimmer: lecanemab slowing decline by about 27% and donanemab by about 35% in its primary analysis population. That is proof the mechanism responds. The field pushed back on the Cochrane verdict, not without reason. Critics noted that the review pooled 15 failed or withdrawn drug trials alongside the two approved agents, diluting agent-specific efficacy, and that regulatory bodies in the US, UK, and EU continue to endorse both drugs for early-stage disease. Others argued that a 27% slowing of decline may be meaningful to individual patients even if it falls below population-level statistical benchmarks. But I think the verdict deserves to be sat with regardless, because it contains something more useful than a conclusion. It contains a diagnosis.
For decades, the story of Alzheimer’s was seductively simple: amyloid-beta plaques were the villain. Clear the plaques, stop the disease. Clean, linear, falsifiable. Which should have been our first warning, because biology is rarely linear.
The history of AD research is a long reckoning with that simplicity — trials that cleared plaques and didn’t improve cognition, a crisis of faith in the amyloid hypothesis, the gradual honest expansion to include Tau tangles, neuroinflammation, and the recognition that by the time a patient presents with memory loss, the molecular crime scene is already decades old. Both drugs require frequent intravenous infusion, are restricted to early-stage disease, and carry serious risks of cerebral edema and microhemorrhages in a non-trivial fraction of patients. The Cochrane verdict forces the honest question: is a 27% slowing of something we can measure the same as meaningful benefit to the person losing their way home?
The field’s answer, increasingly, is that we have proven the mechanism matters and demonstrated that our current molecules are too blunt an instrument to exploit it. The path to something better runs through understanding the physics of how these proteins misfold — at a precision we haven’t yet achieved. In 2026, the AD pipeline hosts 192 active clinical trials evaluating 158 distinct drugs, a third of them repurposed from other indications, which is itself a measure of how hard the target discovery problem remains. The field is not short of effort or intelligence.
It has been short of the right physics.
Where Classical Mechanics Hits a Wall
Before making that case, I want to be clear about what it is and isn’t. Classical simulation did contribute to the drugs we have. Lecanemab exists in part because of the computational methods our field has spent decades building. The failures of the amyloid hypothesis weren’t primarily computational — they were biological, clinical, and humbling in ways that better physics alone wouldn’t have prevented. The argument isn’t that the models caused thirty years of failure. It’s narrower and I think more defensible: that there are specific, identifiable points in the Alzheimer’s discovery process where classical approximations are systematically wrong in ways we can now fix, and where fixing them changes what gets into the clinic.
Classical force fields — the Newtonian view where atoms are balls and bonds are springs — are genuinely good at a great deal of drug discovery. For well-behaved, drug-like molecules binding to stable, folded proteins with clean active sites, classical FEP delivers reliable binding affinity predictions at throughput that quantum methods cannot match. If you’re optimizing a kinase inhibitor against a rigid ATP-binding pocket with no metal ions and no covalent chemistry, the classical toolkit will serve you well. I want to give that its due, because the argument that follows depends on the distinction being real rather than rhetorical.
Alzheimer’s targets are not those targets. Three failure modes define why — each with a quantum mechanical cause — and they escalate.
Intrinsically Disordered Proteins. Amyloid and Tau don’t have a single native shape. Classical force fields were parameterized on crystallographic data from stable, folded proteins, and they carry structural biases that fail badly when applied to proteins that are supposed to be disordered. A comprehensive benchmark of 17 atomistic force fields on Aβ aggregation kinetics found that the older AMBER parameterizations couldn’t predict β-sheet formation at all, while GROMOS iterations over-stabilized it at unphysical rates. To the field’s credit, subsequent reparameterizations substantially improved on this — and that progress is real. But reparameterization is not the same as solving the underlying physics. These models still operate on fixed point charges, still ignore electronic polarization, and still cannot adapt to a conformational landscape with no stable minimum. They are better Newtonian approximations. The electrons still don’t exist in them.
This matters more than it seems. The plaques visible on a PET scan are, in a cruel irony, probably not the primary killers. The real culprits are the transient, soluble oligomeric intermediates — disordered proto-aggregates that form before the plaques and appear far more toxic to neurons. They are nearly invisible to classical simulation, because they don’t sit still long enough to be characterized by methods that assume structural stability. To distinguish the toxic forms from the benign ones, you need to sample an enormous conformational ensemble with the energetic precision that only quantum mechanical accuracy provides. It is precisely this class of problem — disordered, high-dimensional, quantum-sensitive — where QM/MM methods operate closest to native reality.
If that were the only problem, it would be serious enough. But the same fixed-charge limitation that blinds us to disordered conformations also makes us categorically wrong about the metal ion chemistry that drives aggregation in the first place.
Metal Ion Coordination. Copper and zinc are both implicated in driving amyloid aggregation, and the chemistry of how they interact with Aβ is classical simulation’s hardest failure mode. Transition metals have partially filled d-orbitals that engage in directional coordinate bonding, significant charge transfer, and profound polarization of surrounding residues. Classical force fields represent them with fixed point charges and non-directional harmonic constraints — approximations that systematically misidentify coordination geometries and miscalculate binding stability.
The cost of getting this wrong is not theoretical. Clioquinol and its successor PBT2 — copper and zinc chelators — showed extraordinary promise in animal models, reducing brain amyloid by as much as 50% in mice. Both failed in human trials. Retrospective quantum-level spectroscopic analysis revealed the precise reason: PBT2 sequestered copper from only approximately 59% of Cu-Aβ species, leaving behind a tightly coordinated, solvent-inaccessible complex that continued to drive toxicity. The classical models guiding lead optimization lacked the electronic resolution to identify that certain coordination geometries were energetically impervious to the drug’s chelating reach.
The failure wasn’t in the metal hypothesis. It was in the model. And the model was wrong in exactly the way that anyone who has quietly worried about their force field parameters in a project room would recognize.
The QUELO platform applies a QM/MM approach that treats the ligand and critical active site residues — including coordinated metal ions — with explicit quantum mechanics, calculating electron density directly rather than approximating it with fixed charges. The directionality of d-orbital bonding, the charge transfer from metal to ligand, the precise geometry of coordination: these emerge from the physics rather than being assumed by a parameter set. When I was involved in the early development of AMBER, that level of rigor was computationally heroic — something you might attempt once, on a system you’d chosen specifically because it might survive the wait. The custom QM/MM MD engine we’ve developed changes that arithmetic. It doesn’t make the physics less demanding. It makes the throughput adequate to drug discovery’s actual scale.
Metal coordination is where the classical approximation has already cost us drugs that might have worked. The third failure mode is where it is costing us drugs we haven’t yet found.
The Covalent and the Electronic. The most powerful BACE1 inhibitors bind through a mechanism that classical models literally cannot represent. Classical docking parameterizes Arginine as a set of point charges, optimizing for hydrogen bond geometry and electrostatic distance. What it cannot see is the σ-π orbital overlap between an electron-rich aromatic inhibitor and the Arg235 residue that anchors the most potent BACE1 binders — a “flop-over” mechanism driven entirely by quantum orbital interaction. Compounds that rely on it receive artificially low interaction scores and are eliminated from the pipeline. Not because they don’t work. Because the model is blind to the physics that makes them work.
For covalent drug candidates — increasingly important for neuroinflammatory kinase targets — the gap is more fundamental still. Classical FEP cannot accurately simulate the changes resulting from bond forming. QUELO can handle covalent bonds where the warhead remains constant. But if you are optimizing the warhead or are focused on kinetics, the efficacy of an irreversible covalent inhibitor is governed by Kinact, the kinetic rate constant of bond formation, which is exponentially dependent on the activation energy of the transition state — a quantity that exists only in quantum mechanical terms. QuValent automates exactly this calculation: given a target protein, a defined nucleophile, and a library of candidate electrophilic warheads, it maps the transition state geometry and predicts Kinact values that can correlate with experimental readouts in ways classical methods repeatedly fail to match. I’ve watched covalent candidates get deprioritized because we couldn’t do this calculation at the point in the project when it actually mattered. The capability now exists at a stage where it can change decisions, not just rationalize them afterward.
The AI Question
The discourse around drug discovery has been saturated with AI for several years now, and not without reason — AlphaFold changed what’s possible in structural biology, and generative models are genuinely accelerating hit identification across large chemical spaces. But recent evaluations have shown that modern ML co-folding models consistently fail to correctly orient atoms for precise quantum interactions: the π-stacking, the hydrogen bonds, the orbital overlaps that determine whether a molecule actually binds.
In a 2026 prospective evaluation of AlphaFold3, Chai-1, and Boltz-2 across 557 SARS-CoV-2 Mac1 ligands plus three large docking hit-lists, AlphaFold3’s confidence metric showed almost no ability to distinguish true binders from false positives against the dopamine D4 receptor (AUC = 46.4 — slightly worse than random). The models assigned high confidence scores to molecules that were docking false positives, precisely because they looked convincing but had incorrect atomic orientations. Even when co-folding methods got important conformations of the binding site wrong, ligands were still often posed correctly — a non-physical aspect of ML methods that produces high-confidence wrong answers.
A model trained on classical approximations inherits classical errors. At some point, you are building a faster version of the project room — more confident, higher throughput, and carrying the same wobbly parameters forward at scale.
QuantumFP addresses this at the source: QSimulate’s molecular fingerprinting platform, generating molecular datasets built on quantum-mechanical ground truth rather than the classical approximations that current AI training sets largely inherit. The AI era doesn’t have to encode the mistakes we’re trying to fix — but only if someone builds the training data that doesn’t.
A Path Forward
The mechanism is real; our ability to exploit it with sufficient molecular precision is not yet adequate. We have been brave enough to question the amyloid hypothesis, smart enough to develop the first antibodies that move the needle, and honest enough to admit that 27% isn’t enough.
One thing we haven’t questioned hard enough is whether our fundamental computational tools are adequate to the task. For a disease of this molecular complexity — where the difference between a toxic oligomer and a harmless aggregate may come down to the position of a single electron — that question has to be taken seriously.
The researchers who spent careers on the amyloid hypothesis, who designed and ran those trials, who watched drug after drug fail in phase 3 — they were doing something that deserved to succeed. The field’s persistence in the face of that failure is its own kind of heroism.
The case for better physics isn’t abstract. Seven million Americans. Thirty years. This disease has stolen enough.
We have better methods now.
David Pearlman is a computational chemist with four decades of experience in molecular simulation and drug discovery, including foundational work on AMBER and Free Energy Perturbation methodology. He is VP of Product at QSimulate, where he leads product development of quantum mechanical simulation platforms for drug discovery.