← Gallery

RealQM vs DFT — Feasibility Across Problem Classes

Accuracy is not the only axis. Where DFT is more accurate it is also small and vastly more expensive to have built; where the real problem is large and dynamical, DFT cannot follow at all.

An honest comparison. This is deliberately even-handed: it concedes where DFT is better, prices what that capability cost, and marks where RealQM can go and DFT cannot — separating feasibility (can it run at all?) from validated accuracy (is it right?).

Two axes, not one

The usual comparison collapses everything to accuracy, on which DFT wins for small molecules. But two axes matter:

Accuracy — DFT wins on small systems (bought with 60 years and enormous resources).
Feasibility — can the method run the problem at all? DFT's O(N³) cost and femtosecond time steps put large, long-timescale dynamics out of reach; RealQM's cost is set by a fixed grid, independent of electron count, so it runs where DFT cannot.

Class A — small reaction mechanisms (DFT's home turf)

The four Medicine examples (omeprazole, β-lactam, kinase warhead, GAD65) were run as ~15–20-atom active-site cores. On these, DFT/QM-MM is the stronger tool: quantitative barriers, correct transition-state geometry, decades of validation. RealQM gave only qualitative sign/trend, and its rigid perpendicular scan even failed two substitution mechanisms (aspirin, epoxide). So on Class A, DFT wins. Three honest qualifications, though:

The size was chosen for tractability — and it is exactly DFT's comfort zone. Reducing a mechanism to 15 atoms is picking the ground where the incumbent is strongest.
That accuracy was bought with orders of magnitude more investment (below).
Real drug design is not a 15-atom core. The reduction flatters DFT and misrepresents the problem (next section).

What DFT's Class-A accuracy cost to build

	DFT / computational quantum chemistry	RealQM (this work)
Time	~60–70 years (Hohenberg–Kohn 1964 onward)	months to ~2 years
People	tens of thousands; multiple Nobel Prizes (Kohn & Pople 1998; Karplus/Levitt/Warshel 2013)	one researcher + an AI assistant
Software	Gaussian, ORCA, VASP, Q-Chem… — thousands of person-years	one browser code (molecule.js)
Compute	millions of core-years cumulatively; ~thousands of core-hours per QM/MM study	a laptop GPU, minutes per run
Funding	cumulatively billions of dollars	~zero

So “DFT is more accurate on the small examples” is true — and it is the product of a 10⁴–10⁶× larger investment. That RealQM recovers the right qualitative physics with a laptop and one person is a statement about leverage, not (yet) about competitive accuracy.

Class B — large-scale dynamics (where the real problem lives)

Real drug action is not Class A. The covalent reaction, its selectivity, the approach geometry, and binding all happen inside a full protein pocket + solvent, in motion — thousands of atoms, on nanosecond–microsecond timescales. That is Class B, and here the comparison inverts.

DFT cannot follow. Ab-initio DFT-MD is femtosecond-locked, each step a full SCF (~minutes for ~100 atoms). Reaching a microsecond folding event is ~10⁸–10⁹ steps — centuries to millennia of compute. In practice DFT-MD tops out at ~10–100 picoseconds: orders of magnitude short of the event.

RealQM runs it — on a laptop. Its fixed-grid GPU dynamics costs the same for 85 atoms as for 190, because the grid, not the electron count, sets the cost:

system	size	fold time	DFT-MD	RealQM
Chignolin (10-residue β-hairpin)	85 atoms (dry reduced; ~138 with H)	~0.6 µs	~ps ceiling (~10³ yr to fold)	200³ grid, 200k steps, laptop GPU — runs
Trp-cage (20-residue mini-protein)	190 atoms	~4 µs	infeasible	same 200³ grid — runs
crambin / GB1 / coiled coil / α-helix	~hundreds–thousands	—	far out of reach	trajectories run (fixed grid)

Same fixed grid, 85 vs 190 atoms, same cost — that is the whole feasibility story in two numbers. This is a place DFT categorically cannot go, and RealQM does.

Why the large step — and why hydrogen need not be dropped

The reason ordinary molecular dynamics is chained to femtosecond steps is that explicit Newtonian integration goes unstable above the fastest oscillation — the X–H bond stretches (~10 fs period). Those vibrations are thermal jitter, irrelevant to the slow fold, yet they dictate the step for the whole system.

RealQM moves the nuclei by overdamped Brownian dynamics — no inertia, hence no fast oscillations to resolve; the step is limited only by the slow, collective motion one actually wants. This removes the fs–oscillation lock regardless of whether hydrogens are present. So hydrogen need not be dropped to reach long timescales: the earlier reduced runs left out H (and water) only to halve the atom count and simplify, not out of necessity. The hydrogens — and explicit water — can be carried, restoring the physics that actually drives folding (backbone N–H···O=C hydrogen bonds and the water-driven hydrophobic effect), with the same overdamped dynamics still taking large diffusive steps. This is exactly what standard Newtonian MD cannot do — keep the hydrogens and still step past the fs wall.

The price is honest and unchanged: overdamped dynamics captures the slow pathway and thermodynamics, not the real-time kinetics (the “time” is diffusive/model time, not a calibrated clock).

The honesty guard — feasibility is not validated accuracy

Running is not the same as being right, and this must not be conflated:

RealQM's Class-B runs are feasible but unvalidated. The trajectories run; whether they reach the correct fold blind is open — blind folding needs biases (electron repulsion dominates at distance).
Binding affinity still needs dispersion, which RealQM lacks. Reaching Class B does not by itself solve the drug problem.

So the claim is precisely: RealQM can operate where DFT cannot — and must still earn accuracy there.

Bottom line

Small (Class A): DFT is more accurate — on its home turf, bought with 60 years and billions; RealQM gets the right physics with a laptop and one person.
Large & dynamical (Class B): DFT cannot run it at all; RealQM can, on a laptop — the regime where the real drug problem, and most of biology, actually lives.
The trap to avoid: reducing drug design to a 15-atom core to make it a Class-A problem. That flatters the incumbent and hides that the real question is Class B — where feasibility, not accuracy on a toy, is the first hurdle.

Drug design is not a simple thing. The honest comparison is not “who is more accurate on a 15-atom model,” but “who can even run the large, dynamical system the drug actually acts in” — and there, only one of the two can start.

← Back to the Gallery · chignolin dynamics · companion: RealQM vs DFT for folding