# torchdrug.metrics#

## Basic Metrics#

### AUROC#

area_under_roc(pred, target)[source]#

Area under receiver operating characteristic curve (ROC).

Parameters
• pred (Tensor) – predictions of shape $$(n,)$$

• target (Tensor) – binary targets of shape $$(n,)$$

AUROC()#

alias of torchdrug.metrics.area_under_roc

### AUPRC#

area_under_prc(pred, target)[source]#

Area under precision-recall curve (PRC).

Parameters
• pred (Tensor) – predictions of shape $$(n,)$$

• target (Tensor) – binary targets of shape $$(n,)$$

AUPRC()#

alias of torchdrug.metrics.area_under_prc

### R2#

r2(pred, target)[source]#

$$R^2$$ regression score.

Parameters
• pred (Tensor) – predictions of shape $$(n,)$$

• target (Tensor) – targets of shape $$(n,)$$

### Accuracy#

accuracy(pred, target)[source]#

Classification accuracy.

Suppose there are $$N$$ sets and $$C$$ categories.

Parameters
• pred (Tensor) – prediction of shape $$(N, C)$$

• target (Tensor) – target of shape $$(N,)$$

### Matthews Correlation Coefficient#

matthews_corrcoef(pred, target)[source]#

Matthews correlation coefficient between prediction and target.

Definition follows matthews_corrcoef for K classes in sklearn. For details, see: https://scikit-learn.org/stable/modules/model_evaluation.html#matthews-corrcoef

Parameters
• pred (Tensor) – prediction of shape :math: (N, K)

• target (Tensor) – target of shape :math: (N,)

### Pearson Correlation Coefficient#

pearsonr(pred, target)[source]#

Pearson correlation between prediction and target.

Parameters
• pred (Tensor) – prediction of shape :math: (N,)

• target (Tensor) – target of shape :math: (N,)

### Spearman’s Rank Correlation Coefficient#

spearmanr(pred, target)[source]#

Spearman correlation between prediction and target.

Parameters
• pred (Tensor) – prediction of shape :math: (N,)

• target (Tensor) – target of shape :math: (N,)

Classification accuracy for categories with variadic sizes.

Suppose there are $$N$$ samples, and the number of categories in all samples is summed to $$B$$.

Parameters
• input (Tensor) – prediction of shape $$(B,)$$

• target (Tensor) – target of shape $$(N,)$$. Each target is a relative index in a sample.

• size (Tensor) – number of categories of shape $$(N,)$$

Area under receiver operating characteristic curve (ROC) for sets with variadic sizes.

Suppose there are $$N$$ sets, and the sizes of all sets are summed to $$B$$.

Parameters
• pred (Tensor) – prediction of shape $$(B,)$$

• target (Tensor) – target of shape $$(B,)$$.

• size (Tensor) – size of sets of shape $$(N,)$$

Area under precision-recall curve (PRC) for sets with variadic sizes.

Suppose there are $$N$$ sets, and the sizes of all sets are summed to $$B$$.

Parameters
• pred (Tensor) – prediction of shape $$(B,)$$

• target (Tensor) – target of shape $$(B,)$$.

• size (Tensor) – size of sets of shape $$(N,)$$

Top-k precision for sets with variadic sizes.

Suppose there are $$N$$ sets, and the sizes of all sets are summed to $$B$$.

Parameters
• pred (Tensor) – prediction of shape $$(B,)$$

• target (Tensor) – target of shape $$(B,)$$

• size (Tensor) – size of sets of shape $$(N,)$$

• k (LongTensor) – the k in “top-k” for different sets of shape $$(N,)$$

### F1 Max#

f1_max(pred, target)[source]#

F1 score with the optimal threshold.

This function first enumerates all possible thresholds for deciding positive and negative samples, and then pick the threshold with the maximal F1 score.

Parameters
• pred (Tensor) – predictions of shape $$(B, N)$$

• target (Tensor) – binary targets of shape $$(B, N)$$

## Chemical Metrics#

### SA#

SA(pred)[source]#

Synthetic accesibility score.

Parameters

pred (PackedMolecule) – molecules to evaluate

### QED#

QED(pred)[source]#

Quantitative estimation of drug-likeness.

Parameters

pred (PackedMolecule) – molecules to evaluate

### Chemical Validity#

chemical_validity(pred)[source]#

Chemical validity of molecules.

Parameters

pred (PackedMolecule) – molecules to evaluate

### LogP#

logP(pred)[source]#

Logarithm of partition coefficient between octanol and water for a compound.

Parameters

pred (PackedMolecule) – molecules to evaluate

### Penalized LogP#

penalized_logP(pred)[source]#

Logarithm of partition coefficient, penalized by cycle length and synthetic accessibility.

Parameters

pred (PackedMolecule) – molecules to evaluate