We are moving our newsletter to Substack for a better experience!
In Week #207 of the Doctor Penguin newsletter, the following papers caught our attention:
1. Antibiotic Discovery. Discovering new structural classes of antibiotics has been very challenging, as evidenced by the 38-year gap between the introductions of the two most recent classes - the fluoroquinolones in 1962 and the oxazolidinones in 2000.
Wong et al. developed explainable graph neural network models for antibiotic discovery by training on large datasets measuring antibiotic activity and human cell cytotoxicity. Each input compound's chemical structure is represented as a mathematical graph, with atoms as vertices and bonds as edges, and the graph neural network models perform convolution steps to pool together information from neighboring atoms and bonds to predict if a new compound will inhibit bacterial growth or be cytotoxic (i.e., toxic to living cells). By screening over 12 million compounds, the study revealed multiple compounds with antibiotic activity against S. aureus. Of these, one structural class exhibits high selectivity, overcomes resistance, possesses favorable toxicological and chemical properties, and is effective in both the topical and systemic treatment of MRSA in mouse infection models.
Read paper | Nature
2. Pathology. A vision-language generalist AI assistant for human pathology.
Lu et al. developed PathChat, a general-purpose, interactive, vision-language AI assistant for pathology. PathChat combines a pretrained language model and a vision encoder pretrained on 100 million histology images from over 100,000 patient cases and 1.18 million pathology image-caption pairs. It was further finetuned on the currently largest pathology instruction dataset with 257k instructions and corresponding responses. PathChat achieved 87% accuracy on multiple-choice diagnostic questions using publicly available cases of diverse tissue origins and disease models. Additionally, human expert evaluation on open-ended questions showed PathChat generated more accurate and preferred responses compared to the best commercial solution, GPT-4V, and public multimodal models. PathChat can support diverse use cases such as analyzing morphological features, suggesting diagnoses, performing tumor grading following guidelines, and serving as a consultant for human-in-the-loop differential diagnosis.
Read Paper | arXiv preprint
Video demo of PathChat. Credit to Lu et al., original video links can be found in the paper.
3. Autism. Individuals with autism spectrum disorder (ASD) have structural retinal changes that potentially reflect brain alterations.
Kim et al. developed deep ensemble models using retinal photographs to screen for ASD and symptom severity in 958 participants (1890 eyes). The models accurately differentiated ASD from typical development, with a mean AUROC, sensitivity, and specificity of 1.00 (95% CI, 1.00-1.00). They also showed moderate ability to discriminate severe from milder ASD, with a mean AUROC of 0.74. Notably, the model retained an AUROC of 1.00 even when reducing the retinal images to just 10% containing the optic disc. This suggests the optic disc region is crucial for ASD screening. Additionally, the study found retinal photographs may serve as an objective screening tool starting at age 4 years, which is earlier than the average diagnosis age of 60.48 months. Overall, the models demonstrate potential for accessible and objective screening for ASD diagnosis and possibly for symptom severity.
Read Paper | JAMA Network Open
4. Explainability. Could providing AI explanations to clinicians effectively mitigate the harmful effects of model errors and biases on the clinicians?
Jabbour et al. conducted a randomized clinical vignette study evaluating the impact of artificial AI model predictions, with and without explanations, on clinician diagnostic accuracy for patients hospitalized with acute respiratory failure. The results showed that while standard AI models improved clinician accuracy, systematically biased AI models (those consistently misdiagnosing patient subgroups) significantly reduced accuracy. Providing commonly used image-based AI model explanations did not help clinicians recognize the biased models. Thus, explanations did little to remedy the harmful effects of biased models on clinician accuracy. This study highlights the need for clinician training in AI limitations and developing better explanations tailored to clinician needs.
Read Paper | JAMA
-- Emma Chen, Pranav Rajpurkar & Eric Topol