What If You Could Screen 100,000 Molecules Before Lunch?
The computational chemistry bottleneck nobody’s solving... and the engine that just did.
Here’s something that should bother you.
In 1964, Pierre Hohenberg and Walter Kohn published a paper that would eventually win the Nobel Prize. The idea was beautiful: instead of solving the full Schrödinger equation for every electron in a molecule... a problem that scales catastrophically with system size... you could, in principle, get the same answer from the electron density alone. Three variables instead of 3N. Elegant. Powerful.
That was over sixty years ago.
And today, a single DFT calculation on a modest drug-like molecule still takes somewhere between two hours and two days, depending on the basis set, the functional, and how much coffee your cluster has had. The entire field of computational chemistry... one of the most cited branches of physics in history... is built on a method that scales as N³ to N⁷ with system size. Every time someone adds a few more atoms to the molecule, the compute bill doesn’t grow. It explodes.
Here’s the part that really gets me. The latest “breakthrough” in the field? GPU-accelerated DFT, which delivers... and I’m quoting actual papers here... 30 to 50 times faster. That’s the headline. That’s the state of the art in 2026.
Fifty times faster than catastrophically slow is still catastrophically slow.
The Screening Problem Nobody Talks About
If you work in drug discovery, you already know this pain. You have a promising scaffold. Chemistry has synthesized a few dozen variants. The medicinal chemist wants ADMET profiling... Absorption, Distribution, Metabolism, Excretion, Toxicity... the panel that tells you whether your molecule will survive contact with a human body.
Here’s the dirty secret: nine out of ten drug candidates fail in clinical trials. The average cost to bring a single drug to market? Somewhere between one and two billion dollars, over ten to fifteen years of development. And when you dig into why they fail, the numbers are sobering. About 40 to 50% fail for lack of efficacy. Another 30% fail because of toxicity. Unmanageable side effects discovered too late, after years of optimization and millions spent.
Thirty percent. Nearly a third of all clinical failures are toxicity... the kind of thing that a good ADMET screen could have flagged early. But here’s the catch: you can only screen what you can compute, and you can only compute what you can afford.
With traditional methods, running a comprehensive ADMET panel on even a modest library means weeks of compute time. So teams compromise. They screen 200 molecules instead of 10,000. They prioritize potency first, ADMET later. They optimize for the wrong thing for months, then discover at the finish line that their best compound is a hERG blocker or a CYP3A4 inhibitor.
It’s not that the information doesn’t exist. It’s that the cost of asking the question is too high.
What If Asking Was Free?
I’ve been building something for a while now. It’s called FluxMateria... a molecular and materials screening platform built on a new physics framework I developed. Not a faster DFT. Not a machine learning model trained on someone else’s simulations. A completely different computational engine that evaluates physical structure directly, from first principles, with zero fitted parameters.
One engine. Three domains... drug discovery, materials science, reaction chemistry. All from the same physics.
And it’s not ten times faster than DFT. Not a hundred times faster. Not even a thousand.
Three million, six hundred thousand times faster.
I’ll let that land for a second.
A full ADMET panel... solubility, permeability, CYP inhibition, hERG liability, hepatotoxicity, the whole thing... in 127 milliseconds. That’s per molecule. Single-threaded. No GPU.
A band gap prediction for a novel semiconductor? Under a millisecond. A complete reaction mechanism classification... SN1, SN2, E1, E2, E1cb... with activation barrier? Under a millisecond. A retrosynthetic route plan with FLUX-derived barriers across 29 reaction types? Under 50 milliseconds.
At ~350 molecules per second, a library of 100,000 compounds goes through a full ADMET panel in under five minutes. The same screen that would chew through a DFT cluster for weeks... just done. Before lunch. Before your second espresso, if we’re being honest.
And here’s what changes: when the cost of asking becomes nearly zero, you stop compromising. You don’t screen 200 molecules and hope. You screen everything. You run ADMET at the sketch stage... before a single flask is opened, before a single synthesis cycle burns through budget. You catch the hERG blockers on day one, not month nine.
“Fine. But Is It Accurate?”
Fair question. Speed without accuracy is just fast garbage.
So let me put some numbers on the table.
Mechanism prediction accuracy: 100%. Not 99.7%. Not “state of the art approaching human expert level.” One hundred percent on 336 out of 336 test cases covering every major organic reaction class. With an activation barrier MAE of 7.4 kJ/mol.
Bond length prediction: 0.079% mean error across 453 bonds and 64 elements. For hydrogen, FluxMateria predicts 74.13 pm; experiment measures 74.14 pm. That’s a deviation smaller than most instruments can reliably measure.
Solvation modeling: 0.06 logS MAE on nearly 10,000 compounds, validated leave-one-out... currently the top result in the field. IR spectroscopy: under 1% error. NMR: 0.3–0.5 ppm MAE. Band gaps across 1,048 materials... metals, semiconductors, perovskites, transition metal dichalcogenides... all from the same engine: 0.703 eV MAE.
1,361 molecular properties validated. Sub-percent mean error. Zero fitted parameters. ADMET screening validated across 178,000 compounds via leave-one-out, with three endpoints currently at #1 state of the art. Full benchmarks published here.
That last part matters. This isn’t a machine learning model trained on DFT outputs. There’s no training distribution to fall outside of. There’s no dataset to bias. The predictions come from a deterministic physics kernel... same input, same version, same output, every time. Novel chemistry works on day one because the engine runs on physics, not patterns.
The Comparison Table Nobody Expected
Here’s what the landscape looks like now:
I know what this looks like. I know how it reads. If I saw this table from someone else, I’d be skeptical too.
Good. You should be.
Every number on this page has a methodology behind it. Every benchmark has a validation set. Every claim is published on the FluxMateria benchmarks page with the test conditions. We publish our numbers not because we have to... but because, in computational chemistry, the only honest stance is “here’s what we got... check it yourself.”
The Obvious Question
If you’ve read this far, you’re probably asking: How?
How does a single engine, running on a single thread with no GPU, match or beat methods that took over sixty years of development and require supercomputer clusters? How do you get 3.6 million times faster without sacrificing accuracy?
The short answer: new physics.
Not a faster implementation of the old physics. Not a clever approximation scheme. Not a neural network that learned to mimic DFT outputs. A genuinely different computational substrate... a first-principles physics framework that evaluates molecular structure directly, rather than numerically simulating the Schrödinger equation at enormous cost.
Zero fitted parameters. No model training. One engine across drug discovery, materials science, and reaction chemistry.
I’m publishing the underlying physics separately... it’s a new theoretical framework, and it has its own story to tell. For now, what matters is what the engine does. And what it does is collapse the time constant of molecular computation from weeks to milliseconds.
The theory explains why. The platform shows what you can do with it.
Try It
FluxMateria is currently in research preview... and the demos are live.
ADMET Single-Molecule Demo... Paste a SMILES string, get the full ADMET panel with confidence indicators. No signup. Instant results.
Reaction Mechanism Demo... Enter reactant and product SMILES, get the dominant pathway, barrier estimates, and steering factors.
Band Gap Demo... Enter a material formula, get the first-principles band gap with gap-type classification.
If you want to run your own library or walk through a real use case with your data... request pilot access. Your molecules. Your questions.
Because the question isn’t whether 3.6 million times faster is possible.
The question is what your pipeline looks like when it is.
Next in this series: “The ADMET Problem Nobody Talks About”... why moving toxicity screening to day one changes everything in drug discovery.
About the Author
Roberto Campus is the creator of FLUX Theory and architect of FluxMateria. Born in Sardinia, raised in Rome, he’s spent over 35 years asking stubborn questions about how the universe works... and recently started getting answers. He’s either onto something profound or needs a better hobby. The benchmarks suggest the former.