Where FastHDX Fits in AI-Assisted Drug Discovery

Your docking software gives you several plausible ways a molecule could sit inside a protein.

Which one should guide the next experiment?

That is the part of AI-assisted drug discovery I find especially interesting. Generating a model is useful. Having experimental evidence to judge it is what makes the model worth building on.

As I learn more about hydrogen–deuterium exchange mass spectrometry, or HDX-MS, I see a useful opportunity for FastHDX: helping teams collect the solution-phase measurements that can inform those decisions.

What HDX adds to a binding model

A binding pose describes the position and orientation of a molecule bound to a protein.

HDX-MS looks at something different. It measures how certain hydrogen atoms in the protein backbone exchange with deuterium, a heavier form of hydrogen. The exchange reflects hydrogen bonding and the protein’s movements in solution.

Comparing the protein with and without a bound molecule can reveal regions whose exchange behavior changes. Those differences can help researchers investigate how binding affects the protein.

But a protected region is not automatically a contact site. Binding can also change the behavior of regions farther away. That distinction matters when interpreting an HDX experiment: the result helps constrain a structural explanation rather than directly drawing an atomic map.

How machine learning uses that information

The MInt-HDX preprint, posted July 16, 2026, explores this connection.

It uses XGBoost, a machine-learning method, to learn relationships between HDX changes and features of known protein–ligand complexes. Those predictions guide where physics-based docking searches and help rank the resulting poses.

The authors report encouraging results across three main evaluation systems. This remains preliminary: parts of the workflow were also tuned using that evaluation set.

What interests me is the purpose of the AI here. It helps translate experimental measurements into information that can guide a structural model.

Where FastHDX becomes useful

If experimental measurements are going to influence a model, collecting informative data needs to be practical.

FastHDX automates labeling and quenching, then transfers the sample into a liquid chromatography–mass spectrometry (LC–MS) workflow. Quenching sharply slows exchange before analysis. Automation reduces the manual handling involved in collecting these comparisons.

The FastHDX specifies a labeling range from 50 milliseconds to 24 hours. The value is having access to rapidly exchanging regions alongside slower ones, so the experiment can cover the behavior relevant to the question.

The FastHDX also specifies 8 µL of sample per time point. When protein is limited, that matters because different conditions, time points, and repeat measurements all draw from the same supply. It is a per-time-point figure, not the total sample requirement for a study.

For me, these capabilities make a strong case for considering FastHDX when building an experimental workflow around computational predictions. It handles the labeling and quenching; the LC–MS system and downstream analysis turn those samples into interpretable measurements.

Why the measurement window matters

Suppose a region relevant to your models has already exchanged extensively by the first measurement. You may have missed the part of its exchange curve that would tell you more about its behavior.

A 2023 study by Kish and colleagues demonstrated how automated millisecond HDX measurements could reveal exchange behavior in rapidly exchanging regions. The research instrument was called ms2min and helped lead to the creation of the FastHDX

Start with the disagreement between your models

My starting question would be: what do the competing models predict differently, and could HDX provide evidence relevant to that difference?

Then I would want to know whether the measured regions cover the area of interest and whether the observed changes are reproducible. Community recommendations for HDX-MS emphasize both reporting coverage and using repeat measurements to judge meaningful differences.

Give AI more of the experimental picture

This is where I see FastHDX becoming especially valuable for AI-assisted drug discovery.

Imagine that binding changes how quickly a flexible region exchanges, but both the bound and unbound protein are already heavily exchanged by your first seconds-scale measurement. The later measurements look similar, even though the exchange happened differently. Shorter labeling times could reveal that difference and give an HDX-guided model another experimental pattern to work with.

With FastHDX’s labeling times starting at 50 milliseconds, teams could capture more of that early exchange behavior alongside their longer-time measurements. For models that combine HDX data with docking, this could provide additional information for ranking competing poses, identifying predictions that fit the experimental behavior poorly, and choosing which binding hypotheses deserve further investigation.

What excites me is the possibility of building richer datasets for these models to learn from. Measurements across different protein–ligand pairs could capture binding-related changes that would be difficult to distinguish using later time points alone.

FastHDX could help make those differences part of the modeling process. That means giving computational teams more informative experimental evidence to work with—and a stronger basis for deciding which predictions to pursue.

Desmond Gilmour

Email: [email protected]

Phone: +1 519-546-9876

I work at Nicoya Lifesciences. The views expressed here are my own and do not necessarily reflect those of Nicoya. This is a personal publication, not an official company communication.