EVENT DOSSIER
The researchers proposed a new molecular characterization method called WEECFP-SuRGE. This method is based on the Transformer architecture and does not require external pre-training. Its core consists of a 1024-dimensional parameter-free continuous molecular fingerprint (WEECFP) and a rotational encoding based on molecular shortest path graph distance parameters (SuRGE). In the TDC ADMET ranking, this method ranked 2nd overall in its hybrid model and 1st among methods without external pre-training. It also achieved first place in five benchmark tests, including Pgp and liposolubility. Additionally, it outperformed all classic fingerprint baselines in the regression task of the MoleculeNet dataset. WEECFP has near-lossless tokenization capabilities, allowing for the restoration of the original SMILES with greedy overlap reconstruction in 99.9% cases; its graph distance encoding can achieve high correlation with real-world paired distances at a spatial complexity of O(S).
WEECFP-SuRGE 无外部预训练分子模型发布
研究人员提出了 WEECFP-SuRGE，这是一种无需外部预训练、基于 Transformer 架构的分子指纹方法。其核心组件包括一个 1024 维的无参数连续分子指纹 WEECFP，以及一种由分子最短路径图距离参数化的旋转编码 SuRGE。该方法的 7 模型混合体在 TDC ADMET 排行榜上表现优异：整体排名第 2（仅次于 MapLight+GNN），在无外部预训练方法中排名第 1，并在 Pgp、脂溶性等 5 个基准测试中获得第 1 名；在 MoleculeNet 数据集的回归任务中击败了所有经典指纹基线。此外，WEECFP 的分词具有近乎无损的特性，能在 99.9% 的情况下通过贪婪重叠重建恢复原始 SMILES，且其图距离编码能以 O(S) 的空间复杂度实现与真实成对距离高度相关（Pearson r…