Mean squared error (MSE) measures the error in numerical predictions. For each prediction, subtract the observed value and square the difference; then average those squared errors. Lower MSE means a better fit, and squaring gives larger errors more weight. section 4.1.4 section 4.2.1
The formula averages squared differences between observed values y and predictions ŷ across n data points. section 4.1.4
MSE can be both a training loss and an evaluation metric. For solubility prediction, linear regression selects weights to minimize training MSE. For a neural network predicting drug-target activity, gradients of the MSE loss guide weight updates. section 4.2.1 section 8.2
Training a PyTorch model
Use nn.MSELoss() as the criterion. For each batch, calculate predictions, compute the loss against the targets, call loss.backward(), and update the parameters with the optimizer. Track validation loss alongside training loss to guide training choices. section 8.1.2 section 8.2.3
Listing 8.8 Evaluate the Model
def evaluate_model(model, test_loader, criterion, device):
model.eval()
total_loss = 0.0
predictions = []
actuals = []
with torch.no_grad():
for inputs, targets in test_loader:
inputs, targets = inputs.to(device), targets.to(device)
outputs = model(inputs)
loss = criterion(outputs, targets)
total_loss += loss.item()
predictions.extend(outputs.cpu().numpy())
actuals.extend(targets.cpu().numpy())
predictions = np.array(predictions).flatten()
actuals = np.array(actuals).flatten()
mse = mean_squared_error(actuals, predictions)
mae = mean_absolute_error(actuals, predictions)
print(f'Test Loss: {total_loss/len(test_loader):.4f}')
>>> Test Loss: 0.9235
print(f'Mean Squared Error: {mse:.4f}')
>>> Mean Squared Error: 0.9249
print(f'Mean Absolute Error: {mae:.4f}')
>>> Mean Absolute Error: 0.7274This evaluation listing collects test predictions and actual values, then calculates MSE and MAE with scikit-learn. section 8.2.3
MSE is in squared target units. Taking its square root gives root mean squared error (RMSE), in the original target units; mean absolute error (MAE) is another useful summary that does not square residuals. For a solubility model, an MSE of 2.29 corresponds to an RMSE of 1.51 log-molarity units. section 4.1.4
Interpreting affinity predictions
For binding affinity in pKd units, an MSE of about 0.2 gives an RMSE of about 0.45 pKd units, corresponding to roughly a 2.8-fold difference in dissociation constant. Interpret the score in its dataset’s units: KIBA scores are not pKd values. section 11.6.2
A low MSE alone does not establish that a model is ready for screening. Compare results with other metrics and consider the prediction task: ordinary MSE weights errors equally across the affinity range, even when errors among high-affinity binders matter more. A differently weighted loss may better reflect that priority. section 8.1.2 section 11.5.6
















