Use Case7 min readJuly 15, 2026

Green Solvent Substitution Without Losing Solubility

Green solvent replacement in pharma: rank ICH Q3C Class 3 alternatives to DCM and chloroform by predicted solubility before you book lab time.

Oliver Kraft

CovaSyn

Green Solvent Substitution Without Losing Solubility

Everyone agrees the dichloromethane should go. The argument starts when someone asks what replaces it and nobody can say what the process loses in solubility. So the swap gets deferred to the next campaign, and the next one after that.

This article shows how to turn that argument into a ranked shortlist before anyone books lab time, using predicted solubility across an ICH Q3C classified solvent panel. Every number below is a live return from the CovaSolv MCP tools on ibuprofen at 25 C.

Why chlorinated solvents are hard to displace

Two regulatory frameworks drive the decision, and they do different jobs.

ICH Q3C(R8)

sets the residual limits. Class 1 solvents should be avoided outright: 1,2-dichloroethane carries a 5 ppm concentration limit, benzene 2 ppm, carbon tetrachloride 4 ppm. Class 2 solvents are limited by permitted daily exposure: DCM at 6.0 mg/day (600 ppm), chloroform at 0.6 mg/day (60 ppm), THF at 7.2 mg/day (720 ppm). Class 3 solvents are regarded as low toxic potential and are generally acceptable up to 50 mg/day (5000 ppm) without further justification. That last number is the whole point of the exercise: moving a solvent from Class 2 to Class 3 removes an analytical control point and a specification.

The gap between DCM at 600 ppm and a Class 3 solvent at 5000 ppm is the real prize of the swap: one fewer specification and one fewer validated method. Source: ICH Q3C(R8), Impurities: Guideline for Residual Solvents. Class 3 solvents are regarded as low toxic potential and generally acceptable up to 50 mg/day without further justification.

CHEM21

(Prat et al., Green Chem., 2016) is the selection guide most process groups actually use at the bench. It grades solvents Recommended, Problematic, Hazardous and Highly Hazardous on safety, health and environmental axes. Ethanol, IPA, ethyl acetate and acetone sit in the Recommended band. DCM is Hazardous. Chloroform and 1,2-dichloroethane are Highly Hazardous.

Neither framework tells you whether your API will dissolve. That is the gap.

The worked example: what a Class 3 swap actually costs

We ran covasolve_recommend on ibuprofen (SMILES CC(C)Cc1ccc(cc1)C(C)C(=O)O) at 25 C across the full solvent panel, once with scenario="max_solubility" and once with scenario="green_chemistry". Predicted log S is in mol/L. ICH class, cost per litre and boiling point come back in the same call.

Solvent	Predicted log S (mol/L)	ICH Q3C class	Model confidence	Cost/L
THF	0.681	2	0.994	$0.55
1,2-Dichloroethane	0.658	1	0.984	$0.50
DMSO	0.630	3	0.983	$0.70
DCM	0.573	2	0.985	$0.50
Cyclohexanone	0.551	3	0.992	$0.65
IPA	0.502	3	0.999	$0.55
Ethanol	0.476	3	0.987	$0.60
MTBE	0.465	3	0.999	$0.60
Acetone	0.417	3	0.990	$0.40
Ethyl acetate	0.393	3	0.988	$0.45
Chloroform	0.384	2	0.991	$0.55
2-MeTHF	0.378	3	0.987	$1.20
Water	-3.945	3	0.982	$0.01

Read the gaps, not the absolute values.

DCM to IPA costs 0.071 log units. That is roughly 15 percent less solubility. For a dissolution or extraction step with any headroom at all, that is a volume adjustment, not a redesign.
Chloroform is not worth defending. Predicted log S 0.384 is *below* IPA (0.502) and below ethanol, acetone and ethyl acetate. IPA is about 1.3x better on this API. A Class 2 solvent with a 0.6 mg/day PDE is being used here for no solubility benefit at all.
1,2-DCE is the real trade. At 0.658 it is 0.156 log units above IPA, about 1.4x. That is a genuine loss, and it is the case where you would go looking for a mixed-solvent system rather than a straight one-for-one swap.
Two Class 3 solvents bracket DCM. DMSO (0.630) and cyclohexanone (0.551) sit either side of DCM's 0.573. Both dissolve the API. Both have boiling points of 189 C and 155 C, so neither is a sensible isolation solvent. Solubility ranking is a filter, not the decision.

DCM to IPA costs 0.071 log units, roughly 15 percent solubility. Chloroform at 0.384 is below IPA, ethanol, acetone and ethyl acetate, so its Class 2 control burden buys nothing here. Source: Computed with covasolve_recommend on ibuprofen at 25 C, scenario max_solubility. Deployed CovaSolv ensemble_v7; held-out scaffold split R2 0.92, RMSE 0.64 log, 78 percent within 0.5 log on about 5,315 novel-scaffold rows.

What the green scoring actually does

The green_chemistry scenario does not use a different solubility model. It adds a fixed bonus to Class 3 solvents on top of the solubility score, which reshuffles the ranking:

IPA moves from thirteenth on raw solubility to first once ICH class is priced in. Know that this is a rule: the useful output is the shortlist with a solubility penalty attached, not the score. Source: Computed with covasolve_recommend on ibuprofen at 25 C, scenario max_solubility and scenario green_chemistry. The bonus is a fixed rule applied to ICH Q3C Class 3 solvents, not a separate model.

Solvent	max_solubility score	green_chemistry score	Class
IPA	0.602	1.102	3
Ethanol	0.574	1.074	3
Acetone	0.516	1.016	3
Ethyl acetate	0.491	0.991	3
THF	0.780	0.780	2
DCM	0.671	0.671	2

IPA goes from thirteenth on raw solubility to first once the ICH class is priced in. That is the transparent, boring mechanism behind the recommendation, and you should know it is a rule and not a model. The useful output is not the score. It is the ordered shortlist with the predicted solubility penalty attached to each candidate, so the discussion becomes "we lose 15 percent, here is the volume" instead of "it probably won't work".

How to run a solvent swap triage

1. Get the predicted panel for your API with covasolve_recommend, both scenarios. Two calls, seconds each. 2. Cut everything Class 1 and everything Class 2 you cannot justify. 3. Filter the Class 3 survivors on process fit, not solubility: boiling point for distillation and drying, water miscibility if you plan an antisolvent crystallisation, peroxide formation risk for the ethers, cost at scale. 4. Take the top three to four survivors to shake-flask. Confirm the predictions. 5. Only then design the crystallisation. covasolve_antisolvent and covasolve_crystallization will tell you whether the new solvent even supports an isolation route. In the documented ibuprofen case, water-antisolvent addition recovered 99.96 percent from ethanol while cooling crystallisation from ethanol recovered 0.0 percent. A solvent can be a perfectly good dissolution solvent and a useless crystallisation solvent.

Honest limits

The deployed CovaSolv ensemble (ensemble_v7) was evaluated on its real held-out scaffold split: R² 0.92, RMSE 0.64 log units, about 78 percent of predictions within 0.5 log, on roughly 5,315 novel-scaffold rows. For aqueous held-out data the scorecard reads RMSE 0.689, MAE 0.402, R² 0.914, 76.4 percent within 0.5 log. Units are log S in mol/L. Training data: BigSolDBv2.0 (CC-BY 4.0) and OChem.

Now put that next to the worked example. The model's RMSE (0.64 log) is nearly ten times the DCM-to-IPA gap (0.071 log). So:

Do not read fine rank order as truth. The separation between IPA, ethanol, acetone and ethyl acetate here is well inside model error. Treat the top band as an unordered shortlist of four, not a winner and three runners-up.
Large gaps are the reliable signal. Water at -3.945 versus IPA at 0.502 is four orders of magnitude and is not in question. Chloroform sitting below four Class 3 solvents is a big enough gap to act on. A 0.03 log difference is not.
Errors partially cancel within one API. Comparing solvents for the same molecule is more robust than comparing absolute solubilities across molecules, but this is a tendency, not a guarantee, and it is not something we have quantified for your compound.
Check the applicability domain flag. covasolve_predict returns a 95 percent interval and an AD flag per prediction. Novel or large scaffolds get wider intervals. Ibuprofen is an easy case; a 900 Da conjugate is not.
Polymorph, impurity rejection and reaction kinetics are out of scope. A solvent change can shift the crystal form, the impurity purge and the reaction rate. None of that is in a solubility model.
This is triage, not a filing. It ranks experiments. Q3C justification needs measured residual solvent data and validated methods.

Frequently asked questions

What is the best green replacement for dichloromethane in pharma?

There is no single answer, because it depends on the API. For ibuprofen at 25 C, covasolve_recommend puts IPA at predicted log S 0.502 against DCM at 0.573, a loss of about 15 percent solubility for a move from ICH Q3C Class 2 to Class 3. Ethanol, acetone and ethyl acetate follow closely. Run the panel on your own molecule and confirm the top candidates by shake-flask.

What is the difference between ICH Q3C Class 1, 2 and 3 solvents?

Class 1 solvents are known or suspected human carcinogens and should be avoided, with concentration limits such as 5 ppm for 1,2-dichloroethane and 2 ppm for benzene. Class 2 solvents are limited by permitted daily exposure, for example DCM at 6.0 mg/day and chloroform at 0.6 mg/day. Class 3 solvents have low toxic potential and are generally acceptable up to 50 mg/day, or 5000 ppm, without further justification.

Does CHEM21 replace ICH Q3C?

No. They answer different questions. ICH Q3C sets the residual solvent limits a regulator will hold you to. CHEM21 (Prat et al., Green Chem., 2016) is a selection guide that grades solvents Recommended, Problematic, Hazardous or Highly Hazardous on combined safety, health and environmental criteria. Use CHEM21 to choose candidates early and ICH Q3C to define the specification and analytical control.

How accurate is CovaSolv solubility prediction?

On its held-out scaffold split, the deployed ensemble reaches R² 0.92 and RMSE 0.64 log units, with about 78 percent of predictions within 0.5 log across roughly 5,315 novel-scaffold rows. That error is larger than the differences between closely ranked Class 3 solvents, so the model is reliable for shortlisting and for large gaps, and not for declaring a single winner. Every prediction carries a 95 percent interval and an applicability domain flag.

Can I use predicted solubility in a regulatory submission?

No. Treat it as decision support for experiment planning. A Q3C position, a design space or a control strategy needs measured data from validated methods. What the prediction changes is which experiments you run and in what order, which is usually where the time goes.

Why does chloroform rank below ethanol for this API?

Because chlorinated does not mean universally better at dissolving. For ibuprofen, covasolve_recommend returns predicted log S 0.384 for chloroform against 0.476 for ethanol and 0.502 for IPA, a gap large enough to act on. That is worth checking on any legacy process: a Class 2 solvent may be carrying a control burden for a solubility advantage that never existed.