Technology · EMX
Biomedical entity recognition, without a neural network
EMX is the entity matcher we built for Atlas. It finds genes, diseases, drugs and seven other entity types in scientific text, links each mention to the vocabularies that define it, and does it on a CPU in milliseconds, with the same output every time.
- median per full-text article, on CPU, with identifiers
- 55 ms
- F1 on tmVar v3 genes with identifier (HunFlair2: 76.75)
- 76.8
- entity types, against 5 for HunFlair2
- 10
- vocabularies, 71.2 million concepts
- 13
What is biomedical entity normalization?
Biomedical entity normalization finds the genes, diseases, chemicals and other entities named in a text and links each one to an identifier in a reference vocabulary, such as NCBI Gene, MONDO or MeSH. It is what lets a system know that LY3537982 and olomorasib are the same molecule, or that two papers discuss the same gene under different names. It usually takes two steps and a trained model; EMX does both in one pass, from a compiled lexicon.
With Atlas
- One pass: each mention comes with its position and every candidate identifier
- Ten entity types, from genes to procedures
- Byte-identical output from one run to the next
- Runs on CPU, with no GPU and no external API
Why we built it
Atlas first extracted entities by sending each document to a general-purpose language model. It worked, with four costs: about four seconds per abstract, a bill for every document, results that can change between two runs, and entity names without identifiers, which leaves the linking across papers to another step.
Entity recognition is a task where most of the knowledge already exists, in the vocabularies the field maintains. We wanted to use it directly: fast, deterministic, with identifiers from the start, and keep language models for what they do better, such as reading results, relations and contradictions.
How it works
EMX is written in Rust. It compiles 71.2 million concepts and 51.8 million written forms from 13 scientific vocabularies, among them NCBI Gene, MONDO and MeSH, into a lexical index searched with an Aho-Corasick automaton, then applies deterministic context rules.
Each mention keeps its exact position, its type and its candidate identifiers in the vocabularies. Nothing is learned, so an error is fixed in the lexicon or in a rule, and the fix can be measured.
Entity types: gene / protein, disease, chemical / drug, cell type, cell line, organism, anatomy, cellular component, procedure, phenomenon.
Benchmarks: recognition quality
F1 scores (0 to 100) against human-annotated public corpora and one internal full-text set. A mention counts as correct only if both its boundaries and its type are exact. The “setting” column says where a system was trained on the corpus, which makes its score an upper bound rather than a fair comparison.
| Dataset | Setting | EMX | HunFlair2 | scispaCy | DeepSeek v4.1 flash |
|---|---|---|---|---|---|
| tmVar v3, genes + identifierHunFlair2 figure as published by Sänger et al. (2024); DeepSeek returns no identifier | HunFlair2's own cross-corpus evaluation | 76.8 | 76.75 | — | — |
| tmVar v3, gene recognition | Same corpus, recognition only | 84.0 | 87.9 | — | 52.4 |
| BC5CDR test | Out of domain for EMX and HunFlair2; in domain for scispaCy | 81.6 | 81.1 | 75.7 | 74.6 |
| NCBI Disease test | In domain for HunFlair2 | 75.3 | 93.1 | 57.7 | 74.9 |
| GNormPlus | In domain for HunFlair2 | 73.0 | 81.3 | — | — |
| Atlas full-text gold, 5 shared typesGold built by a consensus of language models, not human experts; DeepSeek received the guideline used to build it | 150 passages of recent full-text articles, rebuilt without DeepSeek's vote | 74.0 | 68.5 | 32.7 | 75.5 |
| Quaero (French) | Out of domain for all; none is built for French | 26.7 | 20.2 | 17.3 | — |
On the task that matters for Atlas, finding a gene and identifying it, EMX matches HunFlair2 on the corpus its authors chose for their own cross-corpus evaluation (76.8 against 76.75). HunFlair2 still finds more genes, and wins clearly on the corpora it was trained on.
Benchmarks: speed
Median and 95th-percentile time per full-text article, on the same 500 articles (10 used for warm-up). Each system runs on the hardware where it is fastest.
| System | Task | Hardware | Median | p95 |
|---|---|---|---|---|
| EMX | Recognition + identifiers | CPU, 4 threads (cloud) | 55 ms | 216 ms |
| EMX | Recognition + identifiers | CPU, 4 threads (Intel i7-7700) | 59 ms | 192 ms |
| scispaCy | Recognition | CPU (Intel i7-7700) | 1.6 s | 4.7 s |
| HunFlair2 | Recognition | GPU NVIDIA L4 | 5.5 s | 17.5 s |
| DeepSeek v4.1 flash | Recognition, chunks sent in parallel | API | 20.4 s | 38.5 s |
| HunFlair2 | Recognition + normalization (100 articles) | GPU NVIDIA L4 | 35 s | 93 s |
At equal task, recognition and identifiers on the same cloud machine, EMX is about 640 times faster than HunFlair2 on a full-text article and about 1,700 times faster on an abstract (3.6 ms against 6.1 s), with no GPU. On an abstract, DeepSeek answers in 3.3 to 4.0 seconds through its API, without identifiers.
Where EMX is not better
- HunFlair2 finds more genes: 87.9 against 84.0 F1 for recognition on tmVar v3, with higher recall in general.
- On corpora HunFlair2 was trained on, the gap is real: 93.1 against 75.3 on NCBI Disease, 81.3 against 73.0 on GNormPlus.
- Choosing the right gene among species and homonyms remains the main weakness: 57.4 F1 with identifier on GNormPlus, 42.1 on NLM-Gene.
- A lexicon cannot find a name it has never seen, such as a new compound code or a typo; a neural network can generalize.
- Generic types are covered but weak: procedure 39.9, phenomenon 37.6, anatomy 45.1 F1 on the Atlas gold.
- EMX needs 8 to 9.4 GiB of RAM and a 9.5 GiB lexicon on disk, against 1.7 GiB of RAM and 194 MB for scispaCy.
- None of the four systems handles sequence variants such as P155S; that is the job of specialized tools like tmVar.
Method and caveats
- The Atlas full-text gold (150 passages, 10 types) is a consensus of four language models, not an expert annotation. DeepSeek is one of them; to score it, the gold is rebuilt from the other three, but DeepSeek still received the guideline used to build it.
- EMX’s context rules were tuned on the calibration half of tmVar v3 and of the Atlas gold. The validation halves, never seen during tuning, show no overfitting (78.6 and 72.6 F1, against 75.9 and 71.7 on calibration).
- One EMX configuration is used everywhere, the one Atlas runs; there is no tuning per corpus, and its conventions (longest term) are not those of every corpus.
- On the public disease and chemical corpora, our scorer counts each mention once per document. Differences between systems are fair, but absolute values do not compare with published figures. tmVar v3 follows the HunFlair2 protocol and compares directly.
- DeepSeek was measured with an exhaustive annotation prompt, its best case for this task, not as Atlas calls it in production. scispaCy was measured with its BC5CDR model only.
- The EMX lexicon combines vocabularies under different licenses and is used internally; it is not distributed, so EMX’s figures cannot be reproduced without recompiling it from the sources. The Atlas gold is not published.
Status in Atlas
EMX is being integrated into Atlas. Once in production, it will identify entities in the papers and trials of each project, and language models will work on top of its identifiers. The sources Atlas reads are listed on data sources, and what the entities become in a project on the knowledge graph page.
References
- Sänger M. et al. (2024). HunFlair2 in a cross-corpus evaluation of biomedical named entity recognition and normalization tools. Bioinformatics. doi:10.1093/bioinformatics/btae564
- Corpora: NCBI Disease (Doğan et al. 2014), BC5CDR (Li et al. 2015; Wei et al. 2015), GNormPlus (Wei, Kao, Lu 2015), tmVar v3 (Wei et al. 2022), NLM-Gene (Islamaj et al. 2021), BioRED (Luo et al. 2022), MedMentions (Mohan & Li 2019), Quaero (Névéol et al. 2014).
- Systems: HunFlair2 (flair 0.14.0 and 0.15.1), scispaCy 0.6.2 with en_ner_bc5cdr_md 0.5.4, DeepSeek v4.1 flash at temperature 0.
See what Atlas does with these entities
We set up a project on your disease and target with you, and walk through the papers, trials and evidence it finds.
Request a demoBenchmarks of 2 Nov 2026, emx 0.1.0, lexicon v42. Speed measured on an AWS g6.2xlarge (NVIDIA L4 GPU, 8 vCPU AMD EPYC 7R13) and on an Intel Core i7-7700; quality measured on the same machine with the same scorer for every system, except where a published figure is cited.
