Thermodynamics-Informed Accurate pKa Prediction and Protonation State Generation in PlayMolecule AI

This paper introduces AcepKa, a thermodynamically consistent application within the PlayMolecule AI platform that leverages the Uni-pKa framework and a retrained Uni-Mol backbone to accurately predict pKa values and generate dominant protonation states, while also featuring engineering advancements like a high-speed GPU-accelerated conformer generator and 3D-aware protonation tools for drug discovery.

The Gist

Imagine you are trying to design a key (a drug molecule) to fit into a very specific lock (a protein in your body). For the key to work, it needs to be the right shape and have the right "magnetic charge" to stick to the lock.

In the world of chemistry, this "magnetic charge" is determined by something called pKa. Think of pKa as a molecule's mood ring. Depending on the environment (like the pH level of your blood or stomach), the molecule can gain or lose tiny particles called protons. When it gains or loses these, its electrical charge changes, which changes its shape and how it behaves.

If you get the charge wrong, your drug key might not fit the lock at all, or it might get stuck in the wrong place.

The Problem: The "Chameleon" Molecule

Many drug molecules are like chameleons. They don't just have one fixed state; they can exist in many different "outfits" (protonation states) at the same time.

• Old methods tried to guess the outfit by looking at just one part of the molecule, like guessing a person's whole wardrobe by looking only at their shoes. This often led to mistakes because it ignored how the different parts of the molecule influence each other.
• Super-accurate methods (Quantum Mechanics) try to calculate every single possibility from scratch, but they are so slow and expensive that they are like trying to count every grain of sand on a beach to find a specific one.

The Solution: AcepKa (The "Thermodynamic Chef")

The paper introduces a new tool called AcepKa, built into a platform called PlayMolecule AI. Think of AcepKa as a master chef who doesn't just guess the recipe but understands the entire kitchen.

Here is how it works, using simple analogies:

1. The "Protonation Ensemble" (The Full Menu)

Instead of guessing just one outfit, AcepKa generates a complete menu of every possible outfit the molecule could wear. It looks at the molecule and says, "Okay, if it's acidic, it might look like this. If it's basic, it might look like that." It creates a "family" of all possible versions of the molecule.

2. The "Thermodynamic Consistency" (The Balanced Scale)

This is the secret sauce. In the past, tools would guess the outfits independently, which sometimes broke the laws of physics (like a scale that doesn't balance).
AcepKa uses a framework called Uni-pKa. Imagine a balance scale. If you add weight to one side (a proton), the scale must tip in a specific, predictable way. AcepKa ensures that all the guesses fit together perfectly on this scale. It calculates the "energy cost" of every outfit and uses physics to determine which ones are actually likely to happen in the real world.

3. The "Speedster" Engine (AceConfgen)

To do this, the computer needs to twist and turn the molecule into thousands of 3D shapes to see which one fits best.

• The Old Way: Like a person walking through a maze, trying every path one by one.
• The AcepKa Way: They built a custom engine called AceConfgen. This is like a super-fast drone that can fly through the maze 40 times faster than the competition. It uses special graphics cards (GPUs) to do the heavy lifting in seconds instead of hours.

4. The "PlayMolecule AI" (The Smart Assistant)

AcepKa isn't just a tool you have to code; it's built into a friendly platform called PlayMolecule AI.

• Imagine you have a robot scientist assistant. You can just chat with it: "Hey, take this drug molecule, put it inside this protein pocket, and tell me what charge it has at body temperature."
• The robot automatically runs AcepKa, checks the 3D shape, and shows you the result instantly. It can even warn you: "Hey, this molecule is right on the edge of changing its charge. You might need to test both versions!"

Why Does This Matter?

In drug discovery, time is money, and mistakes are costly.

• Before: Scientists might spend weeks guessing the wrong shape of a drug, only to find out in the lab that it doesn't work.
• Now: With AcepKa, they can simulate the "mood ring" of the drug instantly, ensuring they only build keys that are guaranteed to fit the lock.

In summary: AcepKa is a super-smart, ultra-fast tool that uses the laws of physics to predict exactly how a drug molecule will behave in the human body, helping scientists design better medicines faster and with fewer mistakes.

Technical Summary

1. Problem Statement

Accurate prediction of acid dissociation constants (pKa) and the determination of dominant protonation states are critical challenges in drug discovery. These properties dictate a molecule's solubility, membrane permeability, and binding affinity to protein targets.

• Complexity: Many drug-like molecules are polyprotic, featuring multiple ionizable sites that influence each other through inductive and resonance effects, creating coupled equilibria.
• Limitations of Existing Methods:
  • Traditional/Template-based methods: Rely on empirical corrections and local descriptors, lacking generalizability across diverse chemical spaces.
  • Deep Learning (e.g., MolGpKa, pKasolver): Often frame pKa as a site-specific regression task. This simplification neglects the global protonation network, leading to thermodynamic inconsistencies where predicted values violate fundamental thermodynamic cycles.
  • Quantum Mechanical (QM) methods: While physically rigorous, they are computationally prohibitive due to the need for exhaustive conformational sampling and solvent modeling.

2. Methodology

The authors introduce AcepKa, an application integrated into the PlayMolecule AI platform. It is built upon the Uni-pKa framework, which unifies statistical mechanics with representation learning.

Core Theoretical Framework

Instead of predicting pKa as a scalar regression target, AcepKa models the complete protonation ensemble:

• Microstate vs. Macrostate: It distinguishes between microstates (specific protonation forms) and macrostates (ensembles of protonated/deprotonated species).
• Thermodynamic Consistency: The model predicts the standard Gibbs free energy ($G$) for every microstate. Macro-pKa values and population distributions are then derived analytically using the Boltzmann distribution (Eq. 1 & 2 in the paper). This ensures that the sum of populations and equilibrium constants strictly adhere to thermodynamic laws.

Architecture Components

1. Microstate Enumerator: A rule-based module using SMARTS patterns to combinatorially generate all valid microstates (net charge between -2 and +2) for an input molecule.
2. Uni-Mol Backbone: A 3D molecular representation learning framework using a Transformer-based architecture invariant to SE(3) transformations (rotation/translation). It takes 3D coordinates and atom types as input to predict the standard Gibbs free energy for each microstate.
3. FE2pKa Module: Converts the predicted free energies into macro-pKa values and pH-dependent microstate populations.

Engineering & Workflow Enhancements

• Training Strategy: The Uni-Mol backbone was pre-trained on ~1 million molecules (ChEMBL) and fine-tuned on high-quality datasets (DataWarrior, i-BonD). The training involves supervised pKa prediction and self-supervised tasks (masked atom prediction, 3D coordinate recovery).
• AceConfgen (Conformer Generation): A proprietary, GPU-accelerated conformer generator. It uses fused kernels for force/energy calculations in FP32 precision.
  • Performance: Achieves a 40x speedup compared to NVIDIA's nvMolKit on consumer hardware (e.g., RTX 4090) while maintaining comparable accuracy.
• 3D Modality: Unlike standard tools that only accept SMILES, AcepKa can accept bound ligand poses (e.g., from crystal structures or docking) and apply protonation states directly to that specific geometry.

3. Key Contributions

1. Thermodynamically Consistent Prediction: By modeling free energies of the entire ensemble rather than individual sites, AcepKa guarantees thermodynamic consistency, resolving a major flaw in previous deep learning approaches.
2. High-Performance Conformer Generation: The development of AceConfgen significantly reduces the computational bottleneck of generating 3D conformers for microstates, enabling high-throughput screening.
3. Integration with AI Workflows: Seamless integration into PlayMolecule AI, allowing natural language interaction via an LLM agent ("co-scientist") that can orchestrate complex workflows (e.g., preparing ligands for docking or MD simulations).
4. 3D-Aware Protonation: The ability to predict protonation states for ligands in specific bound poses, crucial for structure-based drug design.

4. Results

• Benchmark Performance: AcepKa achieved State-of-the-Art (SOTA) performance on standard public benchmarks (Novartis Acid/Base, SAMPL6/7/8).
  • It outperformed industry-grade tools like ChemAxon Marvin and Schrödinger Epik.
  • It demonstrated superior accuracy in distinguishing subtle electronic effects and handling complex ionization patterns (e.g., RMSE of ~0.65–0.73 on SAMPL datasets).
• Conformer Generation Efficiency:
  • On the Platinum 2017 benchmark (4,548 molecules), AceConfgen generated 227,400 conformers in 1.4 minutes on an RTX 4090.
  • In contrast, nvMolKit took nearly 58 minutes and failed on 2 molecules.
  • Accuracy (RMSD) was comparable, with 72.6% of AceConfgen's conformers within 1.0 Å of the reference bioactive pose.
• Application Modes: The tool supports both Single-molecule mode (providing full population distributions and macro-pKa) and Library mode (returning the dominant state for large compound libraries).

5. Significance

AcepKa represents a paradigm shift in computational pKa prediction by bridging the gap between the physical rigor of QM methods and the speed of machine learning.

• Scientific Robustness: It ensures that predictions are physically valid and thermodynamically consistent, which is essential for reliable drug design.
• Practical Utility: By integrating high-performance computing (GPU acceleration) and user-friendly AI interfaces, it makes advanced thermodynamic modeling accessible to medicinal chemists.
• Impact on Drug Discovery: The ability to accurately model protonation states in both solution and protein-bound environments directly improves the reliability of binding affinity predictions, solubility estimates, and the overall success rate of computational drug discovery campaigns.