Proteins, nucleic acids, modifications — a sequence may look like just a string of letters and numbers, but its real value depends on whether it can be matched back to its numbering, modifications and source. Fail that, and the sequence is unusable.
Protein and nucleic-acid sequences are long strings mixing letters and digits, with none of the redundancy that natural language has for error correction. Read one residue as its neighbour and the biological meaning of the entire sequence can change, while the text looks entirely unremarkable.
So the reliable path is not "let the model recognize it", but to treat layout, numbering and context together as constraints for verification.
Sequence listings frequently span pages and columns, and are even interrupted by figure captions. If a page-break point is treated as two independent sequences, it directly produces two erroneous records — and errors of this kind are very hard to detect downstream.
The same sequence only becomes meaningful when it carries its SEQ ID NO number, modification sites and database cross-references. Extracting the sequence body is only the first step; re-aligning its attached information is what decides usability.
Short peptides, oligonucleotides and antibody heavy/light chain sequences are all within scope. Antibody data is usually also linked to antigens, chain pairing and binding-affinity constants, and this chain can be executed together with sequence extraction.
Extract protein and nucleic-acid sequences, standardize formats, decompose modification information, align SEQ ID NO numbering with body-text citations, and handle cross-page continuation and database cross-references, delivering a sequence dataset ready to load into a database.
SEQ ID NOCross-page continuationModification decompositionThe whole sequence is reconstructed according to continuation rules, and the break-point positions and the basis for continuation are recorded. A page break is not treated as two independent sequences, and we do not stitch by length — continuation must have a basis in layout or numbering, otherwise it manufactures two erroneous records.
The original representation is kept and a separate normalized field is created, recording the modification type and site. Different documents may write the same modification differently, and normalizing outright would mask the differences — and those differences are often the part downstream users care about most.
No. The determination is based on position-by-position alignment; we do not merge things that merely look alike, and differing sites are marked separately. If similar sequences are wrongly merged, downstream loses critical variant information outright.
Entities such as herbs, ingredients, targets and diseases are scattered across several source tables; the focus of processing is cross-table linkage and conflict resolution for the same entity, not single-table extraction. The hard part of such tasks is entity alignment, not text recognition.
Send us a representative patent or sequence listing and we can run a small trial annotation first, so you can see directly how break-point continuation and numbering alignment are handled.