Fig. 3: Result of the molecular designs using DyRAMO.

Result of the molecular designs with adjusted reliability levels by DyRAMO (Dynamic Reliability Adjustment for Multi-objective Optimization) (a, c, e) and without considering the prediction reliability (b, d). a, b, The left panel shows the evolution of the predicted properties, with the predicted values scaled from zero to one. The right panel shows the evolution of the maximum value of Tanimoto similarity (MTS) between the designed molecules and the training data of each property. The dotted line represents the set reliability levels. c–e Examples of designed molecules and their predicted properties (Pred.) and prediction reliability, i.e., the MTS with the training data, of each property (Rel.). c Examples of molecules designed by DyRAMO with well-optimized predicted properties. The shown molecules were selected according to the following procedure. Molecules that exceeded the set reliability levels were obtained from ten times molecular designs with high DSS (Degree of Simultaneous Satisfaction of prediction reliability and multiple property optimizations) scores. Molecules that passed two filters, the rule of five filter and the PubChem filter, were extracted. The rule of five filter is a filter based on Lipinski’s rule of five⁸⁰, and the PubChem filter⁸ is a filter based on the frequency of occurrence of molecular patterns in the PubChem database. Subsequently, k-means clustering (k=20) was conducted, and the molecules with the highest reward from each cluster were selected. Other designed examples are shown in Fig. S3. The highlighted areas of molecules represent the quinazoline substructure, a characteristic substructure of known epidermal growth factor receptor (EGFR) inhibitors. Inhibitory activity against EGFR (negative logarithm of the half-maximal inhibitory concentration: pIC₅₀), metabolic stability (remaining percentage in one hour), and membrane permeability (μcm s⁻¹) are colored in blue, green, and red, respectively. Predictions are obtained from a single model run. d Examples of molecules designed without reliability consideration. The molecular designs for 10,000 molecules were conducted three times, and designed molecules that passed both the rule of five filter and the PubChem filter were extracted. K-means clustering was performed on these extracted molecules, and the molecule with the largest reward in each cluster was selected. Five molecules are shown here, and the others are shown in Fig. S4. e Designed molecules in the case where approved drugs were excluded from the training data of the property prediction models. Clustering was conducted to select molecules in the same manner as in (c). DyRAMO reproduced gefitinib, one of the approved drugs against EGFR (the left end of e). Five molecules are shown here, with the remaining molecules presented in Fig. S5.