There is no number, and anyone who gives you one without seeing your data is guessing. What there is, is a method: train on 25 %, 50 % and 100 % of what you already have, plot the error against the training set size, and read off whether the curve is still falling. A curve that is still falling steeply says more data will help and roughly how much. A curve that has flattened says the limit is somewhere else, usually in the label noise or in the representation, and more measurements will not move it.
That said, orders of magnitude are useful for planning, and this article gives both: the method, and the ranges that typically apply.
The learning curve, in practice
1. Split off a fixed test set, by scaffold, batch or time, and leave it untouched. 2. Train on random subsets of the remaining data at 10 %, 25 %, 50 %, 75 % and 100 %, several repeats at each size. 3. Plot the test error against training size, with the spread across repeats. 4. Read the shape.
Three shapes and what each means:
- Still falling steeply. More data is the cheapest improvement available. Extrapolating the curve gives a usable estimate of how many measurements buy how much error reduction.
- Flattening. Additional measurements of the same kind will add little. Look at label noise, at whether the representation captures what matters, or at whether the chemistry contains several regimes that need separate treatment.
- Falling but very noisy between repeats. The test set is too small to measure the difference. Fix the evaluation before deciding anything about data volume.
Run this on the data you already have, before asking for a budget. It converts "we need more data" into "300 more measurements in this region reduce the error by about 20 %".
What drives the requirement
Diversity of the target space.
One congeneric series with substituent variation is a small problem. Four unrelated chemotypes are four problems that happen to share a file. Rule of thumb: the requirement scales with the number of distinct regimes, not with the number of compounds.
Label noise.
The error floor of any model is the reproducibility of the measurement. If replicate measurements differ by 0.4 log units, a model error of 0.4 log units is perfect performance and no amount of data will improve it. Measure this first; it also tells you whether averaging replicates is worth more than new compounds.
Signal strength.
A property that depends strongly and smoothly on structure, such as a physicochemical property, is learnable from few hundred points. A property with a cliff, where one substituent changes everything, needs the cliff to be represented in the data, which usually means many more points.
Starting point.
Fine-tuning a model pretrained on public data typically needs an order of magnitude fewer in-house measurements than training from scratch, because the general chemistry comes from the pretraining.
Realistic ranges
With all the caveats above, these ranges match what most projects find:
| Situation | Typical requirement |
|---|---|
| Fine-tuning for one congeneric series, strong signal | 100 to 300 measurements |
| Fine-tuning for a focused in-house endpoint across a few related chemotypes | 300 to 1,000 |
| Training from scratch for a single well-defined property | 1,000 to 10,000 |
| Broad property model across diverse chemistry | tens of thousands |
| Classification with rare positives | enough to have 50 or more positives, which usually sets the total |
Treat these as a planning aid. The learning curve on your own data overrules them.
Getting more out of the data you have
Before collecting more, four things often help more than new measurements:
- Include the failures. Reactions that did not work, compounds that were inactive, batches that failed specification. A dataset of only successes teaches a model that everything succeeds.
- Use the raw value, not the threshold. "Above 10 µM" discards the information in a continuous measurement. Where only thresholded data exist, censored regression uses them better than classification does.
- Average replicates properly. Where a compound was measured several times, use the mean and record the spread. This reduces label noise without any new experiment.
- Fix the split. Many projects think they need more data when they actually have a leaking split: near-duplicates in training and test, which makes the model look good and the next prediction bad.
Choosing the next measurements, not just more of them
When the learning curve says more data will help, which measurements matter. Two strategies beat random selection:
- Fill the sparse regions. Measure where the model's uncertainty is highest or where your chemical space is thin. These points carry the most information per experiment.
- Close the loop in small cycles. Measure a batch, retrain, let the updated model propose the next batch. Five cycles of twenty experiments usually beat one hundred chosen up front, because each cycle learns from the last. This is the same mechanism as a design of experiments campaign, applied to a model instead of a response surface.
Both need the experiments to be cheap enough to run in batches. Where each measurement takes a week, the learning curve plus a one-shot design is the pragmatic choice.
Limits of the statement
- A learning curve extrapolation is an extrapolation. It is reliable for roughly a factor of two beyond your current data size, not for an order of magnitude.
- The curve is specific to the test set definition. Changing the split rule changes the curve, which is why the split is fixed before the exercise starts.
- Data volume does not fix a wrong target. If the measured property is not the property that matters for the decision, a larger dataset makes a more confident irrelevant model.
- None of this addresses whether you are allowed to use the data. Ownership, confidentiality and consent for customer or partner data are separate questions with their own answers.
Frequently asked questions
How many compounds do you need to train a QSAR model?
There is no single number. For fine-tuning inside one congeneric series with a strong structure-property relationship, 100 to 300 measurements can be enough; training a broad model from scratch typically needs thousands. The learning curve on your own data gives the answer for your case.
What is a learning curve in this context?
A plot of model error on a fixed test set against the size of the training set. Its shape says whether more data of the same kind will reduce the error, and roughly by how much.
Why does assay noise set a limit?
Because the model is fitted to the measured labels. If the same sample measured twice differs by 0.4 units, the model cannot be more accurate than that, and a reported error at that level is already at the ceiling.
Is it better to measure more compounds or more replicates?
It depends on where the error comes from. High label noise argues for replicates, which lower the noise floor; a thin or patchy chemical space argues for new compounds. Comparing the learning curve with the replicate spread shows which.
Does active learning reduce the data requirement?
Often substantially, because the model proposes the measurements that carry the most information. It requires experiments that can be run in several small batches rather than all at once.