A Better Starting Point for Formulation: Seven New Tools on Neurosnap

Explore solubility, ionization, mixture behavior, and drug–polymer compatibility with a growing collection of scientific screening tools.

A formulation is a meeting of molecules, conditions, and practical constraints. Change the pH, switch a solvent, adjust the temperature, or choose another polymer, and the same compound can present a very different scientific problem.

That is what makes formulation research so interesting—and why a single prediction rarely answers everything.

We have added seven tools to Neurosnap to help scientists explore different parts of that picture: ESOL (RDKit), Dimorphite-DL, MolGpKa, Pytherm, ThermoPack, teqp, and COSMOPharm. Together, they offer ways to frame molecular screening and thermodynamic questions before committing to the next round of experiments.

The goal is practical: make it easier to choose a useful calculation, understand what its output means, and decide what evidence to collect next. Several integrations remain experimental, and their individual validation limits matter.

Start with the question

Figure 1. A question-led map of the new collection. These are complementary scientific modules; their inputs and outputs do not automatically form one coupled prediction pipeline.

Aqueous solubility, protonation, liquid mixing, and polymer compatibility describe different aspects of a formulation. Keeping those questions separate helps make the answers more useful.

For example, a neutral-molecule solubility estimate can help with an early screen. Investigating a pH effect needs ionization information as well. Estimating how a solvent mixture divides into liquid and vapor phases requires a different thermodynamic model again.

Here is where each addition fits.

A first look at molecular solubility and ionization

ESOL (RDKit) provides a fast aqueous solubility estimate for neutral organic molecules, along with supporting molecular descriptors. It is useful for an initial comparison of compounds when you want a simple, structure-based starting point. Our implementation uses the published Delaney equation with RDKit descriptors, including an RDKit logP substitution, so the result is an ESOL-style screening estimate. Crystal form, salts, and pH-dependent effects need separate consideration.

Dimorphite-DL explores candidate protonation states over a chosen pH interval. It uses empirical ionization rules to enumerate possible structures and their formal charges. This is useful when preparing molecules for downstream calculations or examining which ionization states deserve attention. An enumerated state does not come with a predicted equilibrium population.

MolGpKa addresses a related question: where are the recognized acidic and basic sites, and what pKa values does its model predict for them? The graph-convolutional approach provides site-level estimates rather than a complete titration curve. Our integration remains beta, with unresolved accuracy issues for basic sites; predictions need independent checking before they guide formulation decisions.

Figure 2. An educational illustration for an idealized monoprotic acid with pKa = 5. These analytical curves are not measurements or outputs from the new services. The solubility relationship assumes equilibrium with a neutral solid, ideal dilute behavior, constant intrinsic solubility, and no competing phases or aggregation. The services do not automatically combine ESOL and pKa into a pH–solubility prediction.

The illustration explains why these molecular questions belong together. The proportion of ionized material can change substantially across a pH range, and a neutral-molecule estimate alone cannot describe the entire solubility picture. The useful next step is to compare the relevant hypotheses with measurements under defined conditions.

When the mixture becomes the main character

Pytherm brings liquid-phase nonideality into view. The Neurosnap service evaluates activity coefficients and excess Gibbs energy using UNIQUAC or classic UNIFAC with supplied model parameters. It also offers a binary solid-solubility calculation using the solute’s melting point and fusion enthalpy. This can help explore how composition and molecular interactions affect a mixture. The current solubility mode assumes a pure crystalline solute and omits heat-capacity corrections and liquid–liquid phase separation. See the Pytherm project for the broader scientific framework.

ThermoPack helps investigate vapor–liquid behavior for a supported set of small-molecule mixtures. Its Neurosnap modes calculate temperature–pressure flash equilibrium, bubble pressure, or dew pressure using Peng–Robinson or Soave–Redlich–Kwong models. These questions are relevant when temperature, pressure, and solvent composition influence processing or separation decisions. The available cubic models are engineering approximations; strongly associating pharmaceutical mixtures may need a different treatment. The ThermoPack project provides the underlying thermodynamic library.

teqp offers another way to interrogate an equation of state. Our service accepts supplied critical constants and acentric factors, then evaluates cubic-model pressure and compressibility at a specified state or solves pure-fluid liquid–vapor coexistence below the critical temperature. It is useful when you want explicit control over the fluid parameters being tested. The current mixture models use zero binary interaction parameters, and mixture property evaluation does not establish phase stability. The teqp project supports a broader range of thermodynamic development than this initial service exposes.

A molecular view of drug–polymer compatibility

COSMOPharm focuses on equilibrium drug solubility in dry PVA, PVP K12, and PVP K30. It combines generated molecular surface-charge profiles, an approximate polymer representation, and supplied crystal properties to estimate saturation loading across temperature.

That makes it relevant to scientists exploring polymer carriers for amorphous solid dispersions. Our adaptation remains experimental: validation across four systems and 22 archived measurements showed useful agreement for several combinations and a substantial unresolved discrepancy for paracetamol–PVP K12. The COSMOPharm article describes the scientific foundation; our dedicated Neurosnap post walks through the approach, tests, and current limitations.

Turn the result into a better experiment

Figure 3. Example ways to organize a study. Choose tools according to the question and available inputs, then test the relevant prediction experimentally. The arrows represent research planning, not automatic transfer of outputs between services.

For an early molecular screen, you might begin with ESOL and explore candidate protonation states with Dimorphite-DL. For a liquid-mixture problem, select an activity model or equation-of-state calculation that matches your components and conditions. For a drug–polymer study, bring measured crystal properties to COSMOPharm and use the resulting curve to help plan loading experiments.

Each path asks something different of the inputs. Molecular tools need the appropriate chemical structure and state. Thermodynamic models need compatible component, group, or interaction parameters. Polymer calculations need a clear account of the drug’s crystal form and the assumptions attached to the chosen polymer grade.

Storage stability remains a separate experimental question, involving effects such as crystallization kinetics, degradation, moisture, and processing history. Model comparisons and equilibrium estimates can inform that work, while stability studies establish what happens to the actual material over time.

The new formulation collection gives Neurosnap users more ways to investigate the problem before the next experiment. Start with the behavior you want to understand, choose the tool that addresses it, and let the calculation help sharpen the study you run next.

Scientific credit belongs to the developers and authors of ESOL, RDKit, Dimorphite-DL, MolGpKa, Pytherm, ThermoPack, teqp, and COSMOPharm. Figures are original Neurosnap illustrations; Figure 2 shows an explicitly idealized analytical relationship rather than benchmark or service-performance data.