Temperature transferable Machine Learned Coarse Grained model for proteins

This paper introduces a thermodynamically informed, temperature-transferable machine-learned coarse-grained framework for proteins that explicitly decomposes the potential of mean force into energetic and entropic components, enabling accurate prediction of temperature-dependent properties like heat capacity across different thermal regimes without retraining.

The Gist

Imagine you are trying to understand how a complex origami crane moves and folds. To do this perfectly, you would need to track every single crease, fold, and piece of paper in real-time. In the world of biology, this is like tracking every single atom in a protein. It's incredibly accurate, but it's also so computationally heavy that it's like trying to count every grain of sand on a beach to understand how the tide moves. It takes too long and requires too much power.

Scientists use a shortcut called "Coarse-Grained" (CG) modeling. Instead of tracking every grain of sand, they group them into buckets. Instead of tracking every atom, they group amino acids (the building blocks of proteins) into single "beads." This makes the simulation run much faster, like watching a time-lapse video instead of real-time footage.

However, there's a catch with the current shortcuts. Most of these models are like a map drawn for a specific day's weather. If you draw a map of a city based on how traffic flows at 8:00 AM, that map works great at 8:00 AM. But if you try to use that same map to predict traffic at 5:00 PM or on a rainy Tuesday, it falls apart. The current "Machine Learned" models for proteins are usually trained at just one specific temperature. If you change the temperature (which happens constantly in nature), the model gets confused and predicts the wrong behavior.

The New Solution: A "Weather-Aware" Map

This paper introduces a new way to build these protein maps that understands temperature. The authors, Jacopo Venturin and Cecilia Clementi, created a system that doesn't just learn what the protein looks like, but why it behaves that way at different temperatures.

Here is how they did it, using a simple analogy:

1. The Two Ingredients of a Protein's Behavior
Think of a protein's behavior as a recipe made of two ingredients:

• Energy (The Cost): How much effort it takes to hold a shape.
• Entropy (The Chaos): How much the protein likes to wiggle and explore different shapes.

Old models mixed these two ingredients into a single "flavor" pot. They learned the final taste but couldn't tell you how much sugar (energy) or salt (entropy) was in it. Because they couldn't separate them, they couldn't predict what would happen if you changed the heat (temperature).

The new model separates the pot. It has two distinct containers: one for Energy and one for Entropy. It learns them separately but forces them to follow a strict rule (a thermodynamic law) that says, "If you change the temperature, the relationship between energy and chaos must change in this specific way."

2. The "Smart" Architecture
Usually, if you want a computer to learn a rule, you just tell it, "Try to follow this rule, and if you mess up, we'll give you a penalty." But the authors built the rule into the structure of the computer brain itself.

Imagine building a car where the engine is physically designed so that it cannot run backwards. You don't need to teach the driver not to go backwards; the car literally can't do it. Similarly, this new model is built so that it is physically impossible for it to violate the laws of thermodynamics. This ensures that even if the model has to guess (extrapolate) for a temperature it has never seen before, it won't make a wild, impossible guess.

3. The "Post-It Note" Fix
The team tested this on a tiny protein called "Chignolin" (a 10-piece puzzle piece of a larger protein). They simulated it at five different temperatures, from cool (300 K) to hot (400 K).

They found that their new model could perfectly predict how the protein folded and unfolded across all these temperatures, whereas the old "single-temperature" models failed miserably when the temperature changed.

But here is the cleverest part: The model was so good at learning the shape and movement that the only thing left to fix for perfect accuracy was a simple "global shift" in energy. It's like if you had a perfect map of a city, but the elevation numbers were off by exactly 10 feet. You don't need to redraw the whole map; you just need to add a "Post-It note" that says, "Add 10 feet to everything."

The authors showed that they could calculate this tiny correction after the model was already trained, without having to re-teach the whole system. This allowed them to accurately predict the protein's "heat capacity" (how much heat it takes to warm the protein up), a measurement that the old models simply could not do.

In Summary
This paper presents a new tool for studying proteins that is "temperature-aware."

• Old way: A model trained for one temperature is useless for another.
• New way: A model that separates "energy" from "chaos" and forces them to follow the laws of physics.
• Result: It can accurately predict how proteins behave from cold to hot, and it can even tell us how much heat they absorb, all without needing to be retrained for every new condition.

This is a step toward making computer simulations of life more reliable, allowing scientists to study complex biological systems across a wide range of conditions without hitting a wall of computational cost.

Technical Summary

Technical Summary: Temperature Transferable Machine Learned Coarse-Grained Model for Proteins

Problem Statement
Coarse-grained (CG) molecular simulations are essential for studying large biological systems over long timescales, yet their accuracy has historically been limited by the difficulty of defining effective interactions. Recent advances in machine learning (ML) have improved CG potentials by approximating the Potential of Mean Force (PMF) using Graph Neural Networks (GNNs). However, a critical limitation persists: standard Machine Learned Coarse-Grained (MLCG) models are typically trained at a single thermodynamic state. They implicitly encode both energetic and entropic contributions within a single effective potential, lacking temperature transferability. Consequently, these models cannot consistently extrapolate or interpolate across different temperatures, nor can they predict temperature-dependent thermodynamic quantities such as heat capacity ($c_V$), which require an explicit separation of energy and entropy.

Methodology
The authors propose a thermodynamically informed MLCG framework that explicitly decomposes the CG PMF ($W$) into its energetic ($U_W$) and entropic ($S_W$) components, adhering to the relation $W = U_W - TS_W$.

1. Thermodynamic Constraint by Construction:
   Instead of using soft penalty terms in the loss function, the authors enforce an exact thermodynamic relation between energy and entropy directly within the neural network architecture. By utilizing the relation $\frac{\partial U_W}{\partial T} - T \frac{\partial S_W}{\partial T} = 0$, they design a constrained network where the unconstrained model's output is transformed to satisfy this condition. This ensures that the model remains physically grounded during extrapolation.

2. Joint Training Objective:
   The model is trained using a multi-objective loss function that combines:

   • Force Matching (FM): Minimizing the difference between the CG forces and the mapped atomistic forces to approximate the full PMF.
   • Energy Matching (EM): Minimizing the difference between the predicted energetic component ($U_W$) and the reference atomistic potential energy.
     This joint optimization allows the model to simultaneously learn the correct free energy landscape and its energetic decomposition.

3. Temperature-Dependent Priors and Shifts:
   To ensure physical stability (preventing unphysical bond stretching or overlaps) and account for mean energy shifts, the framework incorporates:

   • Temperature-dependent priors for bonded terms (bonds, angles, dihedrals), derived analytically for Gaussian-distributed features.
   • A global energy shift ($U_W^{shift} = aT + b$) to align the atomistic energy fluctuations with the training data. The authors note that while a linear shift stabilizes training, higher-order polynomial corrections can be applied post-hoc to improve thermodynamic predictions without retraining the GNN.

4. Dataset:
   The framework was validated on the Chignolin (CLN025) protein using an extensive dataset of $\sim$250 $\mu$s of atomistic Molecular Dynamics (MD) simulations across five temperatures (300 K to 400 K). The CG representation maps each residue to a single $C_\alpha$ bead.

Key Results

• Thermodynamic Consistency: Unlike single-temperature baselines, which fail to maintain the correct folded/unfolded state balance when tested outside their training temperature, the temperature-dependent model accurately reproduces the free energy profiles of the atomistic reference across the entire 300–400 K range. This holds true for both interpolation (350 K) and extrapolation (300 K, 400 K).
• Heat Capacity Prediction: The study demonstrates that the framework can recover the isochoric heat capacity ($c_V$), a quantity inaccessible to standard MLCG models. While the initial linear energy shift was insufficient for precise $c_V$ prediction, applying a simple post-hoc scalar correction (fitting higher-order polynomials to the mean energy) allowed the model to accurately match the atomistic heat capacity without retraining the GNN.
• Decoupling of Structure and Energy: The results suggest that the structural features learned by the GNN are decoupled from the global energetic baseline. The heat capacity is primarily governed by a structure-independent energetic shift, which can be refined efficiently via scalar adjustments.

Significance and Claims
The paper claims to provide a physically grounded pathway toward thermodynamically transferable MLCG simulations. By enforcing thermodynamic consistency through architectural constraints rather than loss regularization, the model guarantees physically consistent behavior across temperature regimes. The work highlights that explicitly decomposing the PMF into energetic and entropic components is essential for:

1. Accurately extrapolating free energy landscapes to untrained temperatures.
2. Predicting temperature-dependent observables like heat capacity.
3. Enabling a modular approach where structural learning (GNN) can be separated from global energetic shifts, allowing for high-precision thermodynamic predictions via inexpensive post-hoc corrections.

The authors conclude that this approach overcomes the limitations of current single-state MLCG models, offering a robust method for simulating complex biomolecular systems across varying thermodynamic conditions.