跳到主要内容
Learn Motion
探索课程产品能力价格

下载 Learn Motion

用手机扫描二维码,下载 App。

App Store

扫码从 App Store 下载 Learn Motion打开商店页面

Google Play

扫码从 Google Play 下载 Learn Motion打开商店页面
登录
Learn Motion

How AI Designs a Drug Molecule from Scratch

1Why Designing a Drug Molecule Is Hard2Turning Molecules into Something a Machine Can Read3How Generative Models Propose New Molecules4Scoring and Filtering the Candidates5Testing, Learning, and Improving the Design
Testing, Learning, and Improving the Design

The Design–Test–Refine Loop

2 / 3
Watch the arrow that comes back. The forward path is familiar by now: generate candidates, score them, filter them, synthesize the shortlist, measure it. The return path is the new part, and it carries two different things. One is data — every measured result becomes a training example, so the model gets a chance to compare what it predicted with what actually happened. The other is criteria. If a whole round of candidates binds well but degrades in cells, that is not one molecule's problem; it is a signal that stability was weighted too lightly. So the next round adjusts the weights, and the shortlist shifts. Follow the loop around once more and notice that each pass narrows the uncertainty without removing it, because every new round proposes candidates the model has never seen measured.
0:00 / 0:00

The pipeline does not end when a candidate is tested. Measured results return to the model and to the design criteria, and the next round begins from a better-informed position. This is what makes the process a cycle rather than a one-way flow.

Two distinct things travel back along the return path. The first is data: each measured outcome — the real binding value, the real stability, the real cellular behaviour — becomes a new training example. The model that predicted the candidate's properties now has one more case where it can compare what it said with what actually happened. Over many rounds, this corrects systematic errors, particularly for chemical regions where the original training data was thin.

The second is criteria. Suppose several candidates from one round bind well but all degrade quickly in cells. That pattern is not a property of any single molecule; it is a signal that the design criteria underweighted stability. The next round adjusts the filters and the scoring weights so that stability counts for more. The model's numbers improve because the data improved, and the shortlist improves because the criteria improved.

The loop repeats: generate, score, filter, synthesize, measure, feed back. Each pass narrows the uncertainty. It does not eliminate it, because every new round introduces candidates the model has still never seen measured.

上一页2 / 3下一页

Learn Motion

生成课程,认真学会。

运营方:Wuhan Daoyin Technology Co., Ltd.

联系我们: [email protected]
隐私政策用户服务协议

© 2026 Learn Motion