Thermodynamic Cycles for Protein Stability and Binding

This direction is merged into another direction: structural input sensitivity. Its last version stays readable and citable here.

Version 1 of 15 papers17 conceptsfrozen 2026-10-04

When we use AI to predict how a mutation affects a protein, we are usually asking a question about stability: "Does this change make the protein more or less likely to hold its shape?" Traditionally, deep learning models have tried to answer this by looking only at the final, folded structure. However, a growing body of research suggests that this is only half the story.

To accurately predict stability and binding, models must account for the "reference state"—the protein in its unfolded or unbound form. By framing protein prediction as a thermodynamic cycle, researchers are bridging the gap between generative AI and classical physics.

The Missing Half of the Equation

In thermodynamics, the stability of a protein ($\Delta G$) isn't an absolute value; it is the difference in energy between two states: the folded state and the unfolded state. Similarly, binding affinity is the difference between the bound complex and the individual components existing in isolation.

Most "inverse folding" models—AI trained to predict which amino acid sequences fit a given 3D structure—focus exclusively on the folded state. They calculate the likelihood of a sequence given a structure, $P(\text{sequence} | \text{structure})$. While these models are powerful, they often fail to generalize because they ignore the energetic cost of the unfolded state. If a mutation makes the folded state slightly more stable but makes the unfolded state much more stable, the net effect is actually a decrease in overall protein stability.

Modeling the Unfolded State

To fix this, Jes Frellsen, Kresten Lindorff-Larsen, and their colleagues in Zero-shot protein stability prediction by inverse folding models: a free energy interpretation introduced a formal thermodynamic "bridge." They argue that to predict the change in stability ($\Delta \Delta G$) caused by a mutation, we must subtract the log-likelihood of the sequence in the unfolded state from its log-likelihood in the folded state.

Since simulating a truly unfolded, "floppy" protein is computationally expensive, they use a clever statistical shortcut: they model the unfolded state using the amino acid distributions found in intrinsically disordered proteins. This correction accounts for the "background noise" of amino acid preferences, significantly improving the accuracy of zero-shot stability predictions without requiring new training data.

Boltzmann Alignment and the Unbound State

This logic extends beyond single proteins to how proteins interact with each other or with drugs. In Boltzmann-Aligned Inverse Folding Model as a Predictor of Mutational Effects on Protein-Protein Interactions, researchers including Xiaoran Jiao, Weian Mao, and Wengong Jin describe a framework called Boltzmann Alignment.

The intuition here is that a protein-protein complex exists in equilibrium with its unbound parts. To predict how a mutation affects binding, the model calculates the difference in sequence likelihood between two scenarios:

  1. The Bound State: The two proteins are modeled together as a single complex.
  2. The Unbound State: The two proteins are modeled as independent chains, physically separated from each other.

By subtracting the energy of the independent chains from the energy of the complex, the model isolates the specific energetic contribution of the binding interface. This "thermodynamic cycle" approach ensures the model doesn't just learn what a "good" interface looks like, but specifically what makes the interface stronger than the individual proteins being apart.

Generalizing to New Molecules

This physics-inspired inductive bias is also proving vital for drug discovery. In Accurate and generalizable protein-ligand binding affinity prediction with geometric deep learning, Guang Yang developed IPBind, which applies these principles to protein-ligand interactions.

By explicitly predicting the atomic energy contributions for both the bound complex and the isolated ligand and protein, IPBind mimics the physical process of binding. This approach led to a 20% improvement in performance when testing on entirely new protein families. The reason is simple: while the specific geometry of a binding pocket might change, the underlying physics of the thermodynamic difference between "together" and "apart" remains constant across all of biology.

Why It Works

The success of these methods reveals a core principle: AI models perform better when their architecture mirrors the physical constraints of the problem.

By treating the outputs of generative models as proxies for Gibbs free energy and explicitly calculating the difference between states, these researchers have turned "black box" sequence predictors into physically consistent thermodynamic simulators. This shift from modeling states to modeling differences between states is the key to creating AI that can truly understand the stability and interactions of the molecular world.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_thermodynamic_cycles_for_protein_stabili_4b4f73dd,
  title = {Thermodynamic Cycles for Protein Stability and Binding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/thermodynamic-cycles-for-protein-stability-and-binding/rdr_4b4f73ddfb35478a90df3b733b539e04},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_thermodynamic_cycles_for_protein_stabili_4b4f73dd_v1,
  title = {Thermodynamic Cycles for Protein Stability and Binding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/thermodynamic-cycles-for-protein-stability-and-binding/rdr_4b4f73ddfb35478a90df3b733b539e04/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}