COMPARISON REPORT

Solubility Methods for ML-era Drug Design & Development

The quality of a predictive model is bounded by the quality of the data it learns from. In solubility prediction, dataset heterogeneity — caused by mixed methods, non-standard protocols, and incomplete annotation — is a persistent source of model error. This report compares the top experimental methods for solubility measurement — single-particle analysis (SPA^®), shake-flask, potentiometric titration, and kinetic solubility — with explicit emphasis on the criteria that matter for ML pipelines: cost, time, data quality, universal applicability, and standardization.

1. Why Machine Learning Needs Better Solubility Data

Aqueous solubility is among the most heavily modelled physicochemical endpoints in cheminformatics, yet published QSPR/ML benchmarks rarely exceed an RMSE of ~0.6–0.9 log units on independent test sets [1], [2]. Similarly for organic solvents, a recent Nature Communications article showed that state-of-the-art deep-learning predictors of organic-solvent solubility have reached the aleatoric limit (0.5–1 log S) of the available test data — meaning that the dominant bottleneck is not the algorithm — it is the experimental training data [3]. Public datasets aggregate measurements obtained over decades using shake-flask, potentiometric titration, and kinetic methods. These protocols differ in equilibration time, solid form (crystalline / amorphous / salt), solvent composition (pH, ionic strength, water content), experimental conditions (temperature etc.), and detection limit, producing inter-laboratory variability of 0.5–1.5 log units even on the same compound [4], [1], [2]. Saal & Petereit’s classic comparison of kinetic vs. thermodynamic data shows a scatter of two orders of magnitude on a single dataset [5].

Recent work further shows that data fidelity, in addition to data volume, are the binding constraints. Amrihesari et al. (2025) demonstrated that ML models trained on high-fidelity experimental polymer-solubility data substantially outperform identical architectures trained on low-fidelity data of comparable size [6]. Llinas & Avdeef and the curated AqSolDB dataset reach the same conclusion for small-molecule aqueous solubility [1], [2].

For a measurement platform to be useful in modern ML-driven discovery it must therefore satisfy seven criteria, summarised below — and crucially, satisfy them simultaneously.

Table 1. The criteria a solubility measurement platform must satisfy to support modern ML-based predictive modelling.

2. The Four Top Methods Compared

2.1 SPA^® (machine vision-based)

Direct optical observation of powder particles dissolving under controlled experimental conditions. Returns equilibrium solubility, intrinsic dissolution rate, and particle morphology in minutes. Sample amount in the µg range per assay, with ~1 mg of total material typically covering ≥20 datapoints across different solvent and vehicle combinations [10], [12]. Detection is image-based and requires no analyte-specific calibration [8]. Solid form changes are verified in situ during the measurement.

2.2 Shake-flask (industry standard)

Excess solid is suspended in solvent and agitated until the dissolved phase reaches equilibrium with the solid, after which phases are separated by filtration or centrifugation and the dissolved fraction is quantified [13], [14]. Returns equilibrium solubility. Sample amount is in the mg to g range per datapoint, with equilibration typically taking 24–72 h. Quantification requires analyte-specific HPLC, UV, or MS calibration. Result depends on equilibration time, sedimentation time, and separation and dilution technique [15]. Solid form is usually verified post measurement.

2.3 Potentiometric acid/base titration

An ionisable compound is dissolved at a pH where it is fully soluble and titrated until precipitation occurs; the inflection in the titration curve yields the pH–solubility profile [16]. Returns equilibrium solubility of the unionized form. Sample amount is approximately µg range per assay, with measurement time in hours per compound. Detection uses a pH electrode and requires no per-analyte external calibration. Applicable only to ionisable molecules within a usable pKa window in aqueous media. The precipitating solid phase is not characterized during the measurement.

2.4 Kinetic solubility (laser nephelometry / DMSO solvent shift)

A small volume of DMSO stock solution is injected into aqueous buffer and precipitation is detected by turbidity (laser nephelometry) or UV absorbance [17]. Returns a non-equilibrium solubility, corresponding to a precipitation-onset concentration of the unknown precipitated form [4], [18], [5]. Sample amount is in the µg range per well, with minutes per assay and high throughput in 96- or 384-well plate formats. Detection is generic (turbidity or UV) and does not require per-analyte calibration. Readouts are sensitive to inject volume, mixing, and detection threshold, and the DMSO co-solvent contributes to the measured signal. The precipitating solid is typically amorphous or ill-defined and is generally not characterized.

3. Side-by-Side Comparison Through the ML-Pipeline Lens

Table 2. Comparison of SPA^®, shake-flask, potentiometric titration and kinetic solubility methods through the criteria that matter for ML predictive-model development.

4. Coverage Across Solvent and Modality Space

A predictive model is only as broad as the labelled chemistry-solvent space it was trained on. Below is the practical coverage of the four methods across the solvent and modality axes that drug discovery and formulation development actually traverse.

Table 3. Practical applicability of each method across the solvent and modality axes traversed in drug discovery and formulation development. SPA^® is the only method that operates universally across all rows from a single protocol [10], [12].

5. Why SPA^® Matters Specifically for ML Predictive Models

5.1 Data quality — data noise is the binding constraint

SPA^® data is produced by a single machine vision-based protocol with the solid form changes verified in situ by microscopy, eliminating the two largest contributors to literature variance: protocol heterogeneity with poor annotation, and unrecognised solid-form transitions [7], [8], [9]. Confirmed inter-laboratory variability is well below baseline of conventional methods [15], [4], [1], [5].

5.2 Material amount — datasets that previously could not exist

Approximately 1 mg of material yields ≥20 SPA^® datapoints — roughly 1000× less material per datapoint than shake-flask requires [10]. This is the difference between “we cannot measure this compound experimentally” and “we can” for natural products, peptides, HPAPIs, scarce intermediates, and early-stage NMEs across solvent space. ML datasets covering novel chemotypes — exactly where computational predictors are weakest — become feasible to construct.

5.3 Time & cost — data density and active learning

Sub-hour turnaround makes SPA^® compatible with active-learning loops in which a model proposes the next compound to measure and is retrained as data arrive. Shake-flask, with 1–3 day cycles and gram-scale material requirements, cannot close that loop. Cost-per-data-point falls because significantly less API is consumed.

5.4 Universality — solvent and modality space from a single SOP

A dataset built from one platform has no covariate shift in the measurement variable. Models trained on SPA^® data generalise across the neutral/ionisable, salt/cocrystal, crystalline/amorphous, and aqueous/biorelevant/organic/formulation-vehicle axes simultaneously — axes that currently require fusing 3–5 different assay datasets, each with its own bias [10], [12]. Crucially, SPA^® addresses the entire solubility and formulation decision tree in a single workflow: aqueous buffers, biorelevant media, process solvents, co-solvent systems, micellar media, emulsions, lipid-based drug delivery systems (LBDDS), cyclodextrin complexes, nanosuspensions, and amorphous solid dispersions [12].

5.5 Standardization — eliminating the protocol confound

Modern ML treatment of structured experimental data shows performance is dominated by protocol variance in the training set [6]. SPA^® centralises measurement around a single platform with a fixed schema. Combined with The Solubility Company’s licensing model [11], the same SOP and data schema can be deployed across sites, producing data that is portable across organisations — a property essentially absent from literature datasets and other method capabilities.

6. Position in the ML-Era Drug Discovery Stack

SPA^® does not replace HTS kinetic screening or shake-flask in their established niches. It occupies the previously unattained slot of thermodynamically meaningful, single-platform, sub-milligram solubility data — across modalities, solvents, stages, and sites — suitable for early development, and the training and benchmarking of predictive models. In a typical ML-driven discovery workflow we therefore expect:

·         Kinetic / nephelometry HTS for early ten thousands-scale discovery rank-ordering.

·         SPA^® for thousands-scale, model-quality thermodynamic data — used to train, calibrate, and validate in-silico predictors.

·         Shake-flask retained for late-stage regulatory anchoring on a small number of compounds.

Conclusion

Across the criteria that govern the usefulness of solubility data for machine-learning predictive models — data quality, material amount, cost, time, universality across solvents and modalities, and standardization — the SPA^® platform either matches or substantially outperforms shake-flask, potentiometric titration, and kinetic methods. Its single-protocol, sub-milligram, machine vision-based design eliminates the dominant sources of label noise in literature solubility datasets, and its applicability to organic solvents, biorelevant fluids, and the full range of formulation vehicles makes it the only method that can produce a single, internally-consistent dataset spanning the entire chemical, solvent and modality space that modern drug discovery actually requires. With both fee for service and direct licensing options, SPA^® is positioned to become the standardised data-generation layer of the ML-era drug-discovery stack.

References

[1] Llinas, A.; Avdeef, A. Solubility Challenge Revisited After Ten Years, with Multilab Shake-Flask Data, Using Tight (SD ~0.17 log) and Loose (SD ~0.62 log) Test Sets. Journal of Chemical Information and Modeling 2019, 59 (6), 3036–3040. DOI: 10.1021/acs.jcim.9b00345

[2] Sorkun, M. C.; Khetan, A.; Er, S. AqSolDB, a Curated Reference Set of Aqueous Solubility and 2D Descriptors for a Diverse Set of Compounds. Scientific Data 2019, 6, 143. DOI: 10.1038/s41597-019-0151-1

[3] Attia, L.; Burns, J. W.; Doyle, P. S.; Green, W. H. Data-Driven Organic Solubility Prediction at the Limit of Aleatoric Uncertainty. Nature Communications 2025, 16, 7497. DOI: 10.1038/s41467-025-62717-7

[4] Bergström, C. A. S.; Avdeef, A. Perspectives in Solubility Measurement and Interpretation. ADMET and DMPK 2019, 7 (2), 88–105. DOI: 10.5599/admet.686

[5] Saal, C.; Petereit, A. C. Optimizing Solubility: Kinetic Versus Thermodynamic Solubility Temptations and Risks. European Journal of Pharmaceutical Sciences 2012, 47 (3), 589–595. DOI: 10.1016/j.ejps.2012.07.019

[6] Amrihesari, M. et al. Role of High-Fidelity vs. Low-Fidelity Experimental Data in Machine-Learning Model Performance for Predicting Polymer Solubility. Macromolecular Rapid Communications 2025. DOI: 10.1002/marc.202500454

[7] Hokkala, E.; Strachan, C. J.; Agopov, M.; Semjonov, K.; Heinämäki, J.; Yliruusi, J.; Svanbäck, S. Fast Imaging-Based Single Particle Analysis Method for Solubility Determination. International Journal of Pharmaceutics 2022, 624, 121976. DOI: 10.1016/j.ijpharm.2022.121976

[8] Svanbäck, S.; Ehlers, H.; Yliruusi, J. Optical Microscopy as a Comparative Analytical Technique for Single-Particle Dissolution Studies. International Journal of Pharmaceutics 2014, 469 (1), 10–16. DOI: 10.1016/j.ijpharm.2014.04.036

[9] Štukelj, J.; Svanbäck, S.; Agopov, M.; Löbmann, K.; Strachan, C. J.; Rades, T.; Yliruusi, J. Direct Measurement of Amorphous Solubility. Analytical Chemistry 2019, 91 (11), 7411–7417. DOI: 10.1021/acs.analchem.9b01378

[10] The Solubility Company. SPA® Technology. https://www.thesolubilitycompany.com/spa-technology/ (accessed April 2026).

[11] The Solubility Company. New Licensing and Partnership Models for a Data-Driven Era. Press release, 9 October 2025. https://www.thesolubilitycompany.com/the-solubility-company-announces-new-licensing-and-partnership-models-for-data-driven-era/ (accessed April 2026).

[12] The Solubility Company. Streamlining Preclinical Formulation via Novel Solubility Profiling. Capabilities presentation, April 2026. Solvent, modality and formulation-vehicle coverage of the SPA® platform.

[13] Black, S.; Dang, L.; Liu, C.; Wei, H. On the Measurement of Solubility. Organic Process Research & Development 2013, 17 (3), 486–492. DOI: 10.1021/op300336n

[14] Veseli, A.; Žakelj, S.; Kristl, A. A Review of Methods for Solubility Determination in Biopharmaceutical Drug Characterization. Drug Development and Industrial Pharmacy 2019, 45 (11), 1717–1730. DOI: 10.1080/03639045.2019.1665062

[15] Baka, E.; Comer, J. E. A.; Takács-Novák, K. Study of Equilibrium Solubility Measurement by Saturation Shake-Flask Method Using Hydrochlorothiazide as Model Compound. Journal of Pharmaceutical and Biomedical Analysis 2008, 46 (2), 335–341. DOI: 10.1016/j.jpba.2007.10.030

[16] Glomme, A.; März, J.; Dressman, J. B. Comparison of a Miniaturized Shake-Flask Solubility Method with Automated Potentiometric Acid/Base Titrations and Calculated Solubilities. Journal of Pharmaceutical Sciences 2005, 94 (1), 1–16. DOI: 10.1002/jps.20212

[17] Bevan, C. D.; Lloyd, R. S. A High-Throughput Screening Method for the Determination of Aqueous Drug Solubility Using Laser Nephelometry in Microtiter Plates. Analytical Chemistry 2000, 72 (8), 1781–1787. DOI: 10.1021/ac9912247

[18] Alsenz, J.; Kansy, M. High Throughput Solubility Measurement in Drug Discovery and Development. Advanced Drug Delivery Reviews 2007, 59 (7), 546–567. DOI: 10.1016/j.addr.2007.05.007

[19] Plum, J.; Bavnhøj, C. G.; Eliasen, J. N.; Rades, T.; Müllertz, A. Comparison of Induction Methods for Supersaturation: Amorphous Dissolution versus Solvent Shift. European Journal of Pharmaceutics and Biopharmaceutics 2020, 152, 35–43. DOI: 10.1016/j.ejpb.2020.04.017