Can pKa Prediction Ever Truly Replace Experimental Measurement?
In the high-stakes world of drug discovery, few properties are as critical as pKa. This single number dictates how a drug dissolves, crosses cell membranes, and ultimately reaches its target. For years, the industry has debated a fundamental question: Can modern computational predictions finally replace the time-consuming, expensive process of experimental measurement?
The short answer is no. While artificial intelligence and machine learning have made remarkable strides, they cannot replace the wet lab. Instead, we are witnessing the emergence of a new, symbiotic relationship where the laboratory and the algorithm depend on each other more than ever before. But the story is deeper than just "AI vs. Lab." It is about complexity. As drug molecules become more intricate—featuring multiple ionizable groups, long-range interactions, self-assembly tendencies, and chemical instability—the gap between a simple "predicted pKa" and "physical reality" widens. Only rigorous, adaptive experimentation can bridge that gap.
The Illusion of Simplicity: Why "Acid" and "Base" Labels Fail
Modern machine learning models have reduced pKa prediction errors significantly, often achieving Mean Absolute Errors of 0.5–0.7 units. However, these models often rely on the same flawed conceptual frameworks that plague traditional software: the idea that a pKa can be "assigned" to a specific functional group. Recent research by Robert Fraczkiewicz (Simulations Plus Inc.), published in the Journal of Medicinal Chemistry (2025), highlights that for multiprotic compounds—molecules with multiple acidic and basic groups—the traditional labels of "acid" and "base" are often physically meaningless.
Consider Glutaric Acid or Cysteine. These molecules have symmetric or near-symmetric ionizable groups. Standard software might predict two different pKa values for two identical carboxylic groups, implying they behave differently. In reality, the molecule exists as a population of microstates. The observed "apparent pKa" is a thermodynamic average of these microstates, not a property of a single group. If a researcher tries to "fix" a drug's solubility by modifying a specific group based on a predicted pKa, they might be targeting a phantom property that does not exist in isolation. The ionization of one group electrostatically influences the others. In Tyramine, for example, the ionization of the amine group shifts the pKa of the phenol group by nearly 0.6 units due to long-range electrostatic effects, even though they are separated by a flexible chain. This phenomenon was detailed in studies by Tejwani, Stouch, and Anderson (Journal of Pharmaceutical Sciences, 2009), which demonstrated that long-range electrostatic interactions can amplify shifts in microscopic ionization constants. Predictive models often miss these microscopic interactions because they treat functional groups as independent fragments. Only experimental determination of microconstants, can reveal the true ionization landscape.
The Hidden Trap: Self-Assembly and Dimerization
Even if a model gets the microstates right, it often fails to account for solution behavior. Many modern drugs are so lipophilic that they do not just dissolve as single molecules; they aggregate. A striking example is Clofazimine, an antibiotic used for leprosy and repurposed for coronaviruses. Literature reported its pKa anywhere from 6.08 to 9.11, a massive discrepancy that baffled researchers for decades.
The breakthrough came from a study by Tatjana Ž. Verbić and colleagues (University of Belgrade and St. John's University), published in Molecular Pharmaceutics (2023). They revealed that Clofazimine forms dimers in aqueous solution. The dimer is more soluble than the monomer, creating a "reverse cosolvent dependence" that confused standard titration methods. This led to wildly different pKa values depending on the concentration and solvent used. If a molecule aggregates, the "pKa" you measure is not just an ionization constant; it is a composite of ionization and dimerization equilibrium. No standard prediction algorithm accounts for this unless it is explicitly trained on dimerization data, which rarely exists in public databases.
The Fragility of "Ground Truth": When the Lab Fights Back
It is tempting to view experimental measurement as the ultimate truth. But the lab bench is not immune to error. If an analyst is not meticulous, the "measured" pKa can be just as misleading as a bad pKa prediction. Instability and decomposition are silent killers of data integrity.
A landmark study on Quercetin and Fisetin, led by Elisabet Fuguet and Clara Rafols at the University of Barcelona (published in Talanta, 2023), highlighted the difficulty of measuring polyphenolic compounds. These molecules have four to five acidic groups with very close pKa values. At high pH, these compounds decompose. If a researcher titrates slowly using standard UV-metric methods, the molecule breaks down before the curve is finished, creating "ghost" pKa values or shifting the entire profile. To solve this, the University of Barcelona team switched to Fast UV technique, completing titrations in just six minutes to outrun the decomposition. They even had to titrate in opposite directions—low-to-high versus high-to-low—to prove which data was real and which was an artifact of degradation. They discovered that Quercetin decomposes above pH 10.5, while Fisetin holds steady until pH 11. If a researcher relies on a prediction for a compound like Quercetin, they might get a number, but without the experimental evidence of stability, that figure lacks the necessary context to accurately predict the compound's behavior.
The Data Flywheel: Why We Need More Than Just Water
This brings us to the most critical, yet often overlooked, aspect of the prediction versus measurement debate. Machine learning models do not learn from thin air; they learn from experimental data. To build a truly robust model that mimics the human body, we need more than just a single pKa value in pure water. The human body is a complex, multi-environment system. The stomach is highly acidic, the blood is neutral with high ionic strength, and the cell membrane is a lipid-rich, non-aqueous environment.
To train an AI that can predict how a drug behaves in these specific environments, we need experimental pKa values measured in different solvent mixtures to simulate membrane permeability, varying ionic strengths to mimic blood plasma versus intracellular fluid, and temperature variations to account for fever or hypothermia scenarios. If we stop measuring pKa values in these diverse media to save money or time, we starve the ML models of the data they need to learn. The result is models that are accurate in water but fail miserably in the body. Every time an analyst measures a pKa in a complex solvent system, or determines a microconstant for a multiprotic drug, they are not just characterizing one drug; they are feeding the future intelligence of the entire industry.
A Symbiotic Future: The Cycle of Intelligence
We are currently standing at a unique intersection in scientific history. On one side, generative AI is designing new molecules at an unprecedented pace, creating thousands of novel structures every day that have never existed before. On the other side, machine learning models are hungry for the high-quality experimental data required to understand these new structures.
This creates a powerful, necessary cycle. As we design more complex drugs with multiple ionizable groups, potential for dimerization, and instability, we generate a massive need for data. These new molecules require fresh, high-quality experimental data in complex media, accounting for microstates, to train the models. Conversely, the lab cannot measure every single virtual molecule. It needs ML to guide the experiments, predicting which compounds are likely to degrade, aggregate, or have confusing microstates before the analyst picks up a pipette.
The bottom line is clear: the lab needs ML to prevent complications and prioritize work, but ML needs the lab to be fed. We are not replacing scientists with robots. We are building a system where AI amplifies the value of the wet lab. The more we discover, the more we need to measure. The more we measure, the smarter our AI becomes.
