Why Conformer Search Needs Machine Learning

Conformational search sits at the foundation of almost every downstream computational chemistry workflow — from predicting NMR spectra to estimating binding affinities. Yet the search itself remains one of the field's most stubborn bottlenecks.

The scaling problem

Classical approaches, whether systematic torsion scans or stochastic methods, all face the same enemy: combinatorial explosion. As the number of rotatable bonds grows, the conformational space grows exponentially, and exhaustive sampling quickly becomes intractable.

Meta-dynamics (MTD) approaches, like those implemented in CREST, address this by biasing molecular dynamics trajectories away from already-visited regions of conformational space. This works well, but the underlying force field still has to evaluate energies and gradients for every step of every trajectory — and the accuracy of the search is bounded by the accuracy of that force field.

Where machine learning fits in

This is where machine learning interatomic potentials (MLIPs) — and methods like AIQM — change the equation. Replacing a classical force field with a fast, near-quantum-accuracy surrogate lets the search explore more of the relevant chemical space without sacrificing the energetic fidelity needed to rank conformers correctly.

In our work implementing an AI-enhanced MTD algorithm within MLatom, the goal has been straightforward to state and hard to achieve: keep the breadth of search that meta-dynamics is good at, while raising the accuracy ceiling closer to what a full quantum chemistry calculation would give you.

What's next

Benchmarking against CREST across a diverse molecular test set has been the natural next step, and early results are encouraging. I'll share more details — including how the method performs on more flexible and larger systems — as the manuscript comes together.