Skip to content

Applications

TGNN-Solv is a solubility model first. The most defensible application layer is therefore not "general chemistry AI", and not a full pharmacology simulator, but decision support for workflows where explicit solvent and temperature choices matter.

The maintained Applications workspace in the Experiment Lab exposes three surfaces built around that principle:

  • synthesis-route solvent screening
  • developability and oral dose-pressure proxies
  • solvent-swap / precipitation screening

1. Synthesis Route Screening

The route-facing workflow accepts a sequence of intermediates with:

  • compound SMILES
  • reaction temperature
  • isolation temperature
  • candidate solvents

For each step, the app scores solvents by whether they look:

  • loadable at the hot endpoint
  • significantly less soluble at the cold endpoint
  • likely to create a useful temperature-swing crystallization window

This is deliberately narrower than retrosynthesis. The model is not planning bond disconnections. Instead it acts as a solvent-selection layer inside a human- or tool-designed route.

2. Developability and Oral Dose Proxies

The developability workflow uses water plus a few explicit formulation-relevant solvent surrogates to answer questions like:

  • does aqueous solubility look dose-limiting?
  • how much cosolvent leverage exists relative to water?
  • is the problem plausibly "preformulation-hard" rather than "hopeless in water only"?

The app reports:

  • predicted ln(x2) and x2
  • water-relative uplift
  • an approximate aqueous molarity proxy
  • a rough 250 mL max-supported-dose proxy for water

That is intentionally framed as a proxy, not a full BCS or PBPK result.

3. PK / PD Scope

Equilibrium solubility is relevant to PK, but it is only one piece of the chain. TGNN-Solv can support:

  • preformulation solvent ranking
  • precipitation and solvent-swap screening
  • early oral dose-pressure heuristics

It does not directly predict:

  • permeability
  • dissolution kinetics
  • precipitation in vivo
  • metabolism or clearance
  • distribution or exposure
  • pharmacodynamics

So the right interpretation is:

  • useful upstream of PK/PD models
  • not a replacement for PK/PD models

4. Solvent-Swap and Workup Design

The solvent-swap screen looks at moving a compound from a donor solvent into a poorer target medium and estimates how strong the crash-out pressure appears to be. This is useful for:

  • antisolvent ideas
  • workup solvent exchange
  • choosing between direct cooling and solvent-shift isolation

5. GUI Integration

Everything above is available in Experiment Lab -> Applications.

The app uses the same checkpoint-selection and inference-family logic as the main inference workspace:

  • TGNN-Solv
  • DirectGNN

That means application screens can be compared across model families without inventing a separate inference stack.