Machine learning models are increasingly employed to automatically extract features from weakly annotated bioimaging data such as microscopy and cytomorphology images. Existing approaches rely solely on automated feature extraction, which leads to inconsistent classification performance across different weakly annotated datasets. To address this, we propose a model‑agnostic knowledge augmentation method that enables ML models to leverage domain knowledge beyond what automated extraction can capture.
The core idea is to inject prior biological knowledge as constraints or auxiliary features into any ML model while keeping the model architecture unchanged. Implementation steps include: (1) building a knowledge graph from literature and experiments that describes the localization relationship of TDP‑43 between nucleus and cytoplasm; (2) designing a knowledge embedding layer that maps graph information into vectors of the same dimensionality as image features; (3) during training, concatenating or fusing the knowledge vectors with image features via an attention mechanism, guiding the model to learn representations aligned with biological meaning.
We evaluated seven mainstream ML architectures (including CNNs, Transformers, GNNs, etc.). Adding knowledge augmentation consistently improved classification accuracy. When combined with masked‑autoencoder‑based self‑supervised learning, the pipeline achieved the best overall performance, boosting precision by 2.9% on three in‑house microscopy datasets and by an average of 6.8% on three public cytomorphology benchmarks.
A concrete application targets the ALS‑related phenotype of TDP‑43 mislocalization from nucleus to cytoplasm. Using three independent microscopy datasets of motor neurons derived from human induced pluripotent stem cells, the model accurately quantified this phenotype and revealed several biological insights: (1) dose‑dependent effects of the proteasomal inhibitor MG132 and the oxidative stressor sodium arsenite (SA) on TDP‑43 mislocalization; (2) variable pathology between sporadic and familial ALS donors; (3) lack of protective effect from edaravone and niclosamide against SA‑induced TDP‑43 mislocalization.
In summary, our pipeline injects domain knowledge into weakly annotated images, enhancing classification robustness and providing a more reliable tool for in‑vitro cellular image analysis, thereby deepening our understanding of ALS pathology and potential therapeutics. Review