with a very small value of b is poor at learning latent representations. In fact, [52] shows that in b-VAE , small values of b tend to encourage disentangled representations and form latent clusters. Finally, it should be clarified that the validity, novelty, and diversity of generated molecules are usually used for a " sanity check " of generative models in molecular design. These three measures should not be used as dominant performance indicators of a design algorithm, as the algorithm should be evaluated in terms of the quality of solutions. The property distributions of generated samples can be important indicators of the proximity of generated samples to actual samples. Figure 3 shows the distributions of QED, SAS, 0.2 2 0.5 QED 0.5 ZINC Atoms 20 10 20 β = 1 (a) Bonds 2 Rings β = 0.1 1 246 -10 -5 05 10 SAS LOGP β = 0.01 00 0 ZINC β = 1 β = 0.1 β = 0.01 CF NO Other 2 0.5 QED PCBA Atoms 20 10 20 1 β = 1 (c) Bonds 2 Rings 0.5 1 SAS β = 0.1 ZINC β = 1 β = 0.1 β = 0.01 Single Double (b) Triple 0.2 ZINC β = 1 β = 0.1 β = 0.01 Tri Quad Pent Hex 510 -10 -5 05 10 LOGP β = 0.01 00 0 PCBA β = 1 β = 0.1 β = 0.01 CF NO Other PCBA β = 1 β = 0.1 β = 0.01 Single Double (d) Triple PCBA β = 1 β = 0.1 β = 0.01 Tri Quad Pent Hex FIGURE 3 Property and structural feature distributions over 20K randomly sampled molecules by FragVAE without evolution. (a) Property distributions over ZINC. (b) Structural feature distributions over ZINC. (c) Property distributions over PCBA. (d) Structural feature distributions over PCBA. MAY 2022 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 21