Skip to main content

Table 6 Comparisons of performance with state-of-the-art methods on classification datasets, splitting the datasets by scaffold splitting in a ratio of 8:1:1 (higher values are better)

From: Double-head transformer neural network for molecular property prediction

Methods

PCBA

BACE

Tox21

SIDER

MolNet [1]

\(0.136 \pm 0.004\)

/

\(0.829 \pm 0.006\)

\(0.648 \pm 0.009\)

RF on Morgan [19]

/

\(0.804 \pm 0.035\)

\(0.582 \pm 0.031\)

\(0.540 \pm 0.013\)

FFN on Morgan [19]

\(0.189 \pm 0.005\)

\(0.843 \pm 0.052\)

\(0.722 \pm 0.041\)

\(0.608 \pm 0.035\)

FFN on Morgan Counts [19]

\(0.195 \pm 0.003\)

\(0.849 \pm 0.047\)

\(0.725 \pm 0.052\)

\(0.595 \pm 0.033\)

FFN on RDKit [19]

\(0.161 \pm 0.005\)

\(0.833 \pm 0.046\)

\(0.788 \pm 0.046\)

\(0.618 \pm 0.031\)

DMPNN [19]

\(0.707 \pm 0.002\)

\(0.759 \pm 0.0291\)

\(0.779 \pm 0.037\)

\(0.602 \pm 0.024\)

Ours

\({{ {0.715}}} \pm {{ {0.004}}}\)

\({{ {0.774}}} \pm {{ {0.014}}}\)

\(0.772 \pm 0.023\)

\({{ {0.661}}} \pm {{ {0.046}}}\)