Beyond Static Conservation: A Hands-On Guide to Dynamics-Aware Evolutionary Profiling with ADEPT
- Feb 2
- 5 min read
Updated: 4 days ago
If you've ever used a conservation score to interpret a missense variant, you've quietly accepted a tradeoff: traditional conservation metrics treat every residue the same way, regardless of whether it sits in a rigid hydrophobic core or a flexible loop that moves through nanometers of space during catalysis.
That's a problem. Because those two kinds of residues fail for completely different biophysical reasons when you mutate them, and modern pathogenicity predictors, can blur them together.
This post walks through ADEPT (Automated Dynamics-Aware Evolutionary Profiling Tool), a workflow we built to separate these signals. By the end, you'll have scored your own protein in three different ways depending on how much setup you want to do: web tool, Python package, or Colab notebook.
The Problem: Conservation Is Doing Two Jobs At Once
Pick any residue that shows up as "highly conserved" in a multiple sequence alignment. Why is it conserved? Two very different reasons are possible:
It holds the protein together. Mutate it, and the fold collapses. Think buried hydrophobic core residues.
It enables the protein to move. Mutate it, and the protein still folds, but it can't do its job. Think hinge residues, catalytic glycines, allosteric switches.
Classical conservation scores can't tell these apart. AlphaMissense, ESM-based predictors, and other static-structure models lean heavily on the first signal and tend to underweight the second. That's why some variants in flexible-but-functionally-essential regions get labeled "benign" or "ambiguous" when they're actually pathogenic.
ADEPT separates these two signals by adding the missing dimension: dynamics, measured directly from Molecular Dynamics (MD) trajectories.
Two Scores, Two Biophysical Constraints
The framework produces two orthogonal metrics for every residue:
RCS: Rigid Conserved Score A residue that is both highly conserved and structurally rigid (low RMSF in MD). These are your stability scaffold: hydrophobic core residues, packed side chains, structural anchors. Mutations here tend to trigger unfolding.
DCS: Dynamic Conserved Score A residue that is highly conserved but structurally mobile (high RMSF in MD). This is the rare, interesting population: residues that evolution clearly cares about, even though they move. They cluster on flexible hinges, loops, and allosteric switches, and they're enriched in glycines and prolines, the two residues that classical biophysics teaches us govern backbone flexibility.
The key insight: a residue can be highly conserved for either reason, and the implications for variant interpretation are completely different.
Why This Matters: A Worked Example
We validated ADEPT against 737 human variants from ClinVar, and the cleanest results came from cases where standart predictors and ADEPT disagreed.
Take NARS1 (asparaginyl-tRNA synthetase) and PGK1 (phosphoglycerate kinase). Both are enzymes with significant conformational dynamics. Both have known pathogenic variants in flexible regions. And in both cases, standart predictors classified specific pathogenic variants as benign or ambiguous, while ADEPT flagged them through a high DCS, correctly identifying that they sit in evolutionarily-locked dynamic regions.
The lesson is that a model trained on static structural patterns will systematically miss a class of pathogenicity that lives in motion. ADEPT is designed to recover exactly that class.
How to Run ADEPT on Your Own Protein
There are three ways to use ADEPT, ranging from "I just want an answer" to "I need to integrate this into a pipeline."
Option 1: The Web Tool (No Setup, 5 Minutes)
Best for: a single protein, exploratory analysis, students and clinicians who don't want to touch Python.
Head to the web server:
You'll need two inputs:
RMSF data: per-residue root-mean-square fluctuation values from an MD trajectory (CSV format).
Conservation data: per-residue conservation scores from an MSA (CSV format, e.g., from ConSurf, Rate4Site, or any equivalent).
Optionally, upload the PDB structure to get a 3D visualization colored by RCS and DCS at the end.
The tool returns a per-residue table, ranked lists of top RCS and DCS residues, and a structural mapping you can rotate in-browser.
If you don't have MD data yet, you can generate it cheaply on Google Compute Engine or AWS: I wrote a full guide for that two years ago and it still applies.
Option 2: The Python Package (For Pipelines)
Best for: batch processing, reproducibility, integration into a larger bioinformatics workflow.
Install it from PyPI:
pip install adept-evo
Then run the CLI:
adept --name MyProtein \
--rmsf data/rmsf.csv \
--data data/conservation.csv \
--pdb structure.pdb
This produces the same outputs as the web tool: per-residue RCS and DCS scores, ranked residue lists, and a PDB file with the scores written into the B-factor column for visualization in PyMOL or ChimeraX.
For programmatic use inside a notebook or script, import the package directly:
from adept_evo import ADEPT
profile = ADEPT(
name="MyProtein",
rmsf_file="data/rmsf.csv",
conservation_file="data/conservation.csv",
pdb_file="structure.pdb"
)
profile.compute()
profile.export("results/")
The full API is documented in the GitHub repo.
Option 3: Colab (For Trying It Without Installing Anything)
Best for: testing the workflow on your own data without local setup, teaching.
A ready-to-run Colab notebook is available:
Click "Open in Colab," upload your RMSF and conservation files when prompted, and run the cells top to bottom. Total runtime is usually under 10 minutes for a protein in the 400–600 residue range.
Interpreting Your Results
Once you have RCS and DCS scores, here's how I read them:
High RCS, low DCS → Structural scaffold. Mutations here are likely loss-of-function via destabilization. Most existing predictors will agree with you on these; ADEPT confirms.
High DCS, low RCS → The interesting class. These are evolutionarily-locked dynamic residues. Variants here may be pathogenic through loss of motion, not loss of fold. Cross-check against ClinVar, and pay special attention if other predictors call them benign: that's exactly the disagreement pattern that surfaced NARS1 and PGK1 in our validation.
High on both → Rare but possible. These tend to be residues that anchor a flexible region's pivot point: important to inspect by hand.
Low on both → Probably tolerant to substitution.
For drug discovery applications, the high-DCS class is also where you should look for cryptic allosteric pockets. These are residues that motion-driven, evolutionarily-essential, and often surface-accessible: a rare combination that conventional pocket-detection tools miss.
What's Next
We released ADEPT as an open framework because the value compounds when more people apply it to their own systems. A few specific things I'd love to see:
Application to GPCRs and ion channels, where dynamic conservation should be especially informative.
Cross-comparison with deep mutational scanning data: does DCS predict positions that are intolerant in DMS but tolerant in classical conservation?
Integration with VUS reinterpretation pipelines in clinical genetics.
If you use ADEPT in your work, I'd genuinely like to hear what you find, including the cases where it disagrees with the result you expected. Those are the most useful data points for refining the method.
Resources
Web tool (ADEPT): https://www.karagolresearch.com/adept
GitHub repository: https://github.com/karagol-taner/Dynamics-aware-Evolutionary-Profiling
Python package: pip install adept-evo
Colab notebook: ADEPT.ipynb
Preprint / manuscript: https://doi.org/10.64898/2026.01.16.699945
Previous post — Molecular Dynamics on cloud computing: Link
If you'd like to cite the framework:
Karagöl, T., & Karagöl, A. (2026). Dynamics-aware Evolutionary Profiling Uncouples Structural Rigidity from Functional Motion to Enable Enhanced Variant Interpretation.
Comments