Computational chemistry and AI drug discovery patents

What can be protected

• Molecules found computationallyCompounds and materials identified by modeling, with the support needed to claim them.
• Modeling and simulation methodsMethods claimed as specific technical improvements, such as faster or more accurate simulations.
• Machine-learning pipelinesTraining and screening workflows, where they can be claimed as concrete technical processes.
• Models and data as trade secretsTraining data, model weights and parameters that may be better protected by confidentiality.
• Quantum chemistryMethods that use quantum computers for chemistry. See quantum computing patents.

Where patents in this field go wrong

1. Assuming AI can be an inventor

   U.S. courts have held that only people can be named inventors. AI-assisted inventions can be patented when a person made a significant contribution, so the human role should be documented.

2. Filing on predictions alone

   A molecule that has only been modeled may not be described or enabled well enough to patent. Some experimental data usually makes the application much stronger.

3. Claiming the software too abstractly

   Claims to a general method of calculation can be rejected as abstract ideas. Claims fare better when tied to a specific technical improvement or result.

4. Publishing the model

   Releasing a model or dataset can disclose what you might later want to protect. Decide what to patent, what to publish and what to keep secret before release.

Building a portfolio for computational discovery

1. File on validated results

   Make and test the most promising molecules or materials, then file with that data.

2. Claim the outputs broadly

   Use the model’s predictions to support a genus around the validated examples, while making sure the genus is enabled.

3. Decide on the platform

   Patent pipeline steps that produce a measurable technical improvement. Keep training data, weights and tuning as trade secrets.

4. Record human contributions

   Keep records of who designed experiments, chose candidates and interpreted results. This supports inventorship under current USPTO guidance.

How I work

I work through the model, the data and the chemistry with your team, then help decide what to patent, what to keep as a trade secret and how to document who invented what. See how I work.

Representative experience

Matters are described in general terms to protect client confidentiality.

• Patent prosecution counsel for a drug discovery company’s portfolio on AI-based diagnostic tools

Questions

Can an AI system be named as an inventor?

No. U.S. courts have held that an inventor must be a natural person. Inventions made with AI assistance can still be patented when a person made a significant contribution to the invention.

Can we patent a molecule we have only modeled, not made?

Possibly, but it is harder. The application must describe the molecule and how to make and use it well enough for an expert; experimental data usually strengthens the case considerably.

Should we patent our model or keep it secret?

Often the model, training data and parameters are better kept as trade secrets, while the molecules and materials they produce are patented.

Can we patent a training dataset?

Generally not as such, because data is not an invention. A dataset can be protected as a trade secret, by contract, and sometimes by copyright in its selection and arrangement. Methods of generating or using the data can sometimes be patented.

Will publishing our model’s predictions affect our patents?

It can. Published predictions can become prior art against later applications, especially for specific structures. File on promising candidates before releasing lists of predicted molecules or materials.

Who owns results produced with a vendor’s AI tools?

Usually you do, but check the vendor’s terms for rights to your inputs and outputs and for any use of your data to train its models. For sensitive research, use enterprise terms that keep your data confidential.