Predicts the radius of gyration of intrinsically disordered proteins from 23 physics-derived sequence features, screening missense mutants in bulk.
No providers recorded yet. Browse all providers
For a disordered protein of length L there are roughly 20L possible missense mutations. Most are benign; a few reshape the conformational ensemble enough to matter for disease, and intrinsically disordered proteins (IDPs) are implicated in Alzheimer's and Parkinson's disease, type II diabetes, and — through p53 — a large share of human cancer. Identifying that small pathogenic subset is the problem, and the standard computational route is too expensive to enumerate: coarse-grained Brownian dynamics on a single-bead-per-residue hydropathy model gives an accurate mean radius of gyration, but running it per mutant across a proteome is out of reach.
This model replaces the simulation with a regression. Swarnadeep Seth and Aniket Bhattacharya, in the Department of Physics at the University of Central Florida, ran Brownian dynamics on roughly 6,500 IDP sequences drawn from MobiDB and used the resulting gyration radii as supervised labels for a multilayer perceptron that reads 23 physics-derived features computed from the FASTA sequence alone. The work was posted as a preprint in July 2024 and published in Biomacromolecules in 2025.
It is deliberately narrow. Where generative approaches such as AI-IDP and IDPForge produce full conformational ensembles, this network predicts one global observable — chain size — cheaply enough that every point mutation in a sequence can be evaluated exhaustively and the handful of large-effect substitutions handed back to simulation.
The network has eight fully connected hidden layers tapering from 128 units down to 4, alternating ReLU and hyperbolic-tangent activations, with four dropout layers at 30%, 30%, 20% and 20%, and a single linear output neuron returning the radius of gyration in nanometres. Training ran for 1,200 epochs at batch size 50 using Adam with a learning rate of 0.0021 and decay rate of 0.0005, minimizing a combined mean squared logarithmic error and Huber loss with early stopping; training and test accuracy stayed within a 90-97% band across repeated runs.
Labels come from Brownian dynamics on the HPS2 coarse-grained model: each amino acid is a single bead carrying a CALVADOS-derived hydropathy value at interaction strength ε = 0.2 kcal/mol, consecutive beads joined by harmonic bonds at 0.38 nm, with charged residues interacting through a Debye-screened Coulomb potential. Source sequences were selected from MobiDB at lengths of 20-300 residues with disorder scores of 99% or above. Inference is four to six orders of magnitude faster than rerunning the simulation.
The intended use is triage. A researcher studying a disease-associated disordered protein can score every possible missense substitution in minutes, isolate the positions where the predicted chain dimension shifts sharply, and commit simulation or experimental effort only to those. Scanning a whole sequence at once localizes mutation-prone regions, which helps in interpreting variants of uncertain significance in disordered regions that structure-based predictors cannot reach. Sequence-to-size prediction also supports the design of disordered linkers with target compaction.
The contribution is a demonstration that a small, fully interpretable network trained on simulation output can stand in for that simulation across a combinatorial search space, and that explainable-AI attribution over physics features recovers the polymer-physics reasoning a practitioner would apply by hand. The caveats are equally clear. Ground truth is simulated rather than measured, so accuracy is bounded by the fidelity of the HPS2 model; the single output is a global size observable, and the authors note explicitly that a large change in the radius of gyration does not by itself imply a pathogenic variant; four of the 33 validation proteins fall outside the 5% error band. No code or trained weights accompany the publication, so the network must be reproduced from the described architecture and training protocol.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.