MediMate DDInter conditional source-grade baseline
This CPU linear classifier estimates the DDInter-reported grade only after a pair is already recorded as an interaction in DDInter and its source grade is Unknown. It cannot detect interactions among arbitrary pairs, establish that an unlisted pair has no interaction, or determine whether a treatment is appropriate. The accompanying source-backed lookup and MCP tools return recorded assertions before any estimate and preserve unknown and ambiguous identities.
Training data and architecture
The dataset release contains 160,235 unique DDInter source-recorded pairs, of which 29,813 have no graded label. This model used 48,212 source-graded training pairs after source-native structure checks and parent/drug-identity partitioning. Unknown, unlisted, unsupported-structure and cross-partition pairs were excluded from supervision; no synthetic negative pairs were used.
For each drug, the model uses a 2,048-bit Morgan radius-2 fingerprint. The unordered pair feature concatenates fingerprint sum and intersection, then applies L2 normalization. A class-weighted one-vs-rest linear SGD classifier predicts DDInter's Major, Moderate or Minor label. config.json records the architecture. model.npz contains numeric arrays, without pickle; report.json records the source and model hashes and evaluation.
Evaluation
The once-evaluated 1,633-pair parent/drug-identity-disjoint holdout yielded macro-F1 0.3729, accuracy 0.5426, and Major recall 103/266 = 0.3872. The majority-class baseline reached accuracy 0.7802 and macro-F1 0.2922 on the same holdout. The registered research bar was macro-F1 0.80 and Major recall 0.95; this model missed both. The reference grades come from one source, not independent clinical adjudication, and the split is not scaffold-disjoint. See report.json and the dataset's linear_research_report.json for confusion matrices and exclusions.
The source-backed lookup is useful for retrieving known assertions and their provenance; the classifier is an experimental review-priority signal only. Do not use its output as a medication alert, as a no-interaction claim, or as a substitute for source evidence or clinical review. Vietnamese brand, salt, formulation and multi-source identity resolution are outside this public model's coverage.
Run
Install medimate-ddi and download the dataset repository to DATASET_DIR. Download this repository to MODEL_DIR, then run:
medimate-ddi-lookup --dataset DATASET_DIR --model MODEL_DIR Naltrexone Abacavir
The example returns DDInter's recorded assertion. Estimated grades appear only for recorded pairs whose source grade is Unknown and both drugs have verified structures. The MCP server intentionally serves exact source assertions without loading this classifier:
medimate-ddi-mcp --dataset DATASET_DIR
The model and DDInter-derived data are distributed under DDInter's CC BY-NC-SA 4.0 terms. Attribute the DDInter team and original publication, and preserve noncommercial/share-alike obligations. See the official DDInter terms. The MIT license in the code repository applies to code only. This artifact contains no MIMIC data or MIMIC-derived weights.
- Downloads last month
- 28