The recognized smiles are inaccurate, all S are wrong for sample.pdf

[jelymewang]

Is there a solution to the inaccurate smiles recognition?

[dahuilangda]

Thanks for the feedback!

The issue with inaccurate SMILES recognition—especially the incorrect sulfur atoms (S)—is due to limitations in the current recognition tool, MolScribe.

I'm working on upgrading the system to allow tool switching, so you'll soon be able to choose between:

• MolScribe
• MolVec
• [MolNexTR](https://github.com/CYF2000127/MolNexTR)

This will give you more flexibility to use the recognition backend that works best for your data.

Appreciate your patience—stay tuned for the update

[jelymewang]

I replaced to MolNexTR, the effect is very good, no need to download new dependencies, git clone to project and modify a few lines of code to use

[jelymewang]

Thanks for the feedback!

The issue with inaccurate SMILES recognition—especially the incorrect sulfur atoms (S)—is due to limitations in the current recognition tool, MolScribe.

I'm working on upgrading the system to allow tool switching, so you'll soon be able to choose between:

• MolScribe
• MolVec
• [MolNexTR](https://github.com/CYF2000127/MolNexTR)

This will give you more flexibility to use the recognition backend that works best for your data.

Appreciate your patience—stay tuned for the update

Yes, the incorrect sulfur atoms (S)

[dahuilangda]

That's awesome to hear—thanks for confirming that MolNexTR works well!

If you're open to it, would you mind submitting your modification as a pull request?
It would be really helpful to integrate your changes into the project and benefit other users as well.

Appreciate your contribution!

[jelymewang]

I have submitted it. Please review it.

[dahuilangda]

The pull request has been merged—thank you for your contribution! Looking forward to any further feedback.

[jelymewang]

The speed is too slow at the moment. I want to improve it to access the API once per batch. Do you have any good ideas?

[jelymewang]

My idea is to segment a batch first and then give it to the api

[dahuilangda]

Thank you so much for your contribution – fantastic work! I’ve merged your PR and then:

• Cleaned up the MolNexTR integration: removed unused files and kept only the core modules
• Moved MolNexTR into utils/ for better organization
• Added an engine‑switch option in the pipeline so you can choose between MolScribe, MolVec and MolNexTR

I’ll also be adding batch/parallel compound‑ID recognition support in the next release. Really appreciate your help on this!

[jelymewang]

I have successfully applied batch processing, but currently, due to issues with Paddle, the program is prone to freezing. It is still an unstable version, likely due to issues with my Paddle version and memory not being released. When there are a large number of pages, the program freezes.

[jelymewang]

At the same time, in some chemical patents, some molecular structure diagrams are relatively unclear. In such cases, SMILES recognition is very inaccurate. I tried super-resolution, but it did not improve much. Do you have any better solutions?

[dahuilangda]

Hi @jelymewang , thanks for the update and for getting batch processing working! A couple of suggestions:

1. Speed improvements
   To cut down on per‐call overhead, batch your inputs at the vision‐model stage when extracting Compound IDs.

2. Handling low‑quality diagrams
   Inaccurate SMILES on blurry or low‑res images is usually a limitation of the parser. Try swapping the engine (MolScribe, MolVec or MolNexTR) via the engine‐switch option—one of them may handle those tricky diagrams better.

[jelymewang]

I also encountered a problem, that is, the inaccurate activity extraction caused by the error in converting PDF to MD. Can I directly use visual understanding or multimodal model to understand the relationship between numbering and activity?