CACHE #1: can computers find drugs for a Parkinson's target?

CACHE is a series of open challenges that test how well computer methods find new drug starting points. Teams make predictions first. Then a lab tests the molecules they picked. Because the answers are not known in advance, nobody can tune a method to a known result. That makes CACHE a fair, prospective benchmark.

This paper reports the first challenge, CACHE #1. Julian M. Kleber is one of its 96 co-authors.

LRRK2 is a protein linked to Parkinson's disease. The challenge focused on one part of it, the WDR domain. At the start, no molecule was known to bind this domain, and the Protein Data Bank had only one structure of it without anything bound (an apo structure).

That is close to the worst case for computer methods. Many tools learn from known binders or rely on a structure with a drug-like molecule already in place. Here neither existed. The domain was also expected to be hard to drug at all.

Twenty-three teams each picked up to 100 molecules that can be bought from suppliers. In total 1,955 molecules were bought and tested.

The first lab test used surface plasmon resonance (SPR), which detects a molecule sticking to a protein. Seventy-three molecules bound with a dissociation constant (K_D) below 150 micromolar. A lower K_D means tighter binding, and 150 micromolar is weak by drug standards.

Those 73 hits moved to a second round. Each team could pick up to 50 related molecules, called analogs, to see if small changes improve binding.

Binding was confirmed in two independent lab tests for seven chemically different series of molecules. Their affinities ranged from 18 to 140 micromolar.

The seven successful teams used quite different approaches. Three used molecular dynamics to model how the binding pocket moves. Three included a fragment docking step. Three used generative design to invent molecules. Five used one or more deep learning steps.

Machine-learning accelerated methods got about the same results as brute-force docking, which checks huge numbers of molecules one by one.

The honest answer from the paper is sobering. First-in-class molecules confirmed in the lab were rare and weak. The authors conclude that recent advances are not yet enough to reliably crack hard targets like this one.

That is useful to know. It shows where computer methods stand today on a target with no prior data, and it gives every team a shared, fair baseline to improve on in later CACHE rounds.

Hit: a molecule that shows binding to the target in a lab test. Hits are starting points, far from finished drugs.

Docking: a computer method that tries to fit a molecule into a protein pocket and scores the fit. Brute-force docking does this for very large libraries.

Micromolar: a unit of concentration. A binding constant of 18 to 140 micromolar means weak binding. Approved drugs usually bind far more tightly, often in the nanomolar range.

Orthogonal assays: two lab tests based on different physical principles. Agreement between them makes a result more trustworthy.

Prospective benchmark: predictions are made before the answers exist, so methods cannot be tuned to known results.

First-in-class: the first molecule known to act on a target or binding site.

The results cover one protein domain and one round of predictions. The binders found are weak starting points, not drugs. Turning them into medicines would need many more rounds of chemistry and testing.

For a different approach to describing molecules for machine learning, see CARATE.