Assessing the Robustness of Gene-Expression Findings on Infliximab Response in Ulcerative Colitis

Background: Public gene expression studies may reuse patients across datasets. I examined how patient reuse, probe selection and data scaling affect two analyses of infliximab response in ulcerative colitis (UC). Methods: I linked pretreatment expression records to patients and later outcomes. Random-forest performance was compared after excluding the test record alone, all records from its patient, or the test record plus matched unrelated records. I assessed the area under the receiver operating characteristic curve (AUROC) and uncertainty by resampling patient groups with predictions fixed. Separately, I tested probe selection and scaling for a published 1,007-gene score in 24 GSE16879 UC records. Further checks assessed training-only probe selection and individual-patient influence. Results: I found 23 identical expression profiles across two datasets, representing 22 patients. The prediction analysis included 79 records from 55 identified patient groups. Excluding all records from the patient being tested reduced AUROC by 0.042 for week-8 healing and 0.050 for week-30 response. However, uncertainty intervals included no difference, and the comparisons were sensitive to one patient. Selecting probes using training data alone did not change the patient-excluded model’s predictions. In a separate analysis, changing probe selection and scaling reduced the difference in gene scores between responders and nonresponders, but nonresponders had higher average scores under all four settings. Conclusions: Model comparisons were sensitive to individual patients, and gene-score differences depended on probe selection and scaling. These findings highlight the need to check patient overlap and preprocessing when evaluating gene-expression studies of treatment response. They do not establish how much overlap inflates performance or whether the models are clinically useful.