OBJECTIVE: Machine learning models have utilized semantic features, deep features, or both to assess lung nodule malignancy. However, their reliance on manual annotation during inference, limited interpretability, and sensitivity to imaging variations hinder their application in real-world clinical settings. Thus, this research aims to integrate semantic features derived from radiologists' assessments of nodules, guiding the model to learn clinically relevant, robust, and explainable imaging features for predicting lung cancer. METHODS: We obtained 938 low-dose CT scans from the National Lung Screening Trial (NLST) with 1,261 nodules and semantic features. Additionally, the Lung Image Database Consortium dataset contains 1,018 CT scans, with 2,625 lesions annotated for nodule characteristics. Three external datasets were obtained from UCLA Health, the LUNGx Challenge, and the Duke Lung Cancer Screening. For imaging input, we obtained 2D nodule slices in nine directions from 50×50×50mm nodule crop. We converted structured semantic features into sentences using Gemini. We fine-tuned a pretrained Contrastive Language-Image Pretraining (CLIP) model with a parameter-efficient fine-tuning approach to align imaging and semantic text features and predict the one-year lung cancer diagnosis. RESULTS: Our model outperformed the state-of-the-art (SOTA) models in the NLST test set with an AUROC of 0.901 and AUPRC of 0.776. It also showed robust results in external datasets. Using CLIP, we also obtained predictions on semantic features through zero-shot inference, such as nodule margin (AUROC: 0.807), nodule consistency (0.812), and pleural attachment (0.840). CONCLUSION: By incorporating semantic features into the vision-language model, our approach surpasses the SOTA models in predicting lung cancer from CT scans collected from diverse clinical settings. It provides explainable outputs, aiding clinicians in comprehending the underlying meaning of model predictions. The code is available at https://github.com/luotingzhuang/CLIP_nodule.

BACKGROUND: Network Component Analysis (NCA) has been used to deduce the activities of transcription factors (TFs) from gene expression data and the TF-gene binding relationship. However, the TF-gene interaction varies in different environmental conditions and tissues, but such information is rarely available and cannot be predicted simply by motif analysis. Thus, it is beneficial to identify key TF-gene interactions under the experimental condition based on transcriptome data. Such information would be useful in identifying key regulatory pathways and gene markers of TFs in further studies. RESULTS: We developed an algorithm to trim network connectivity such that the important regulatory interactions between the TFs and the genes were retained and the regulatory signals were deduced. Theoretical studies demonstrated that the regulatory signals were accurately reconstructed even in the case where only three independent transcriptome datasets were available. At least 80% of the main target genes were correctly predicted in the extreme condition of high noise level and small number of datasets. Our algorithm was tested with transcriptome data taken from mice under rapamycin treatment. The initial network topology from the literature contains 70 TFs, 778 genes, and 1423 edges between the TFs and genes. Our method retained 1074 edges (i.e. 75% of the original edge number) and identified 17 TFs as being significantly perturbed under the experimental condition. Twelve of these TFs are involved in MAPK signaling or myeloid leukemia pathways defined in the KEGG database, or are known to physically interact with each other. Additionally, four of these TFs, which are Hif1a, Cebpb, Nfkb1, and Atf1, are known targets of rapamycin. Furthermore, the trimmed network was able to predict Eno1 as an important target of Hif1a; this key interaction could not be detected without trimming the regulatory network. CONCLUSIONS: The advantage of our new algorithm, relative to the original NCA, is that our algorithm can identify the important TF-gene interactions. Identifying the important TF-gene interactions is crucial for understanding the roles of pleiotropic global regulators, such as p53. Also, our algorithm has been developed to overcome NCA's inability to analyze large networks where multiple TFs regulate a single gene. Thus, our algorithm extends the applicability of NCA to the realm of mammalian regulatory network analysis.

MOTIVATION: Data from DNA microarrays and ChIP-chip binding assays often form the basis of transcriptional regulatory analyses. However, experimental noise in both data types combined with environmental dependence and uncorrelation between binding and regulation in ChIP-chip binding data complicate analyses that utilize these complimentary data sources. Therefore, to minimize the impact of these inaccuracies on transcription analyses it is desirable to identify instances of gene expression-ChIP-chip agreement, under the premise that inaccuracies are less likely to be present when separate data sources corroborate each other. Current methods for such identification either make key assumptions that limit their applicability and/or yield high false positive and false negative rates. The goal of this work was to develop a method with a minimal amount of assumptions, and thus widely applicable, that can identify agreement between gene expression and ChIP-chip data at a higher confidence level than current methods. RESULTS: We demonstrate in Saccharomyces cerevisiae that currently available ChIP-chip binding data explain microarray data from a variety of environments only as well as randomized networks with the same connectivity density. This suggests a high degree of inconsistency between the two data types and illustrates the need for a method that can identify consistency between the two data sources. Here we have developed a Gibbs sampling technique to identify genes whose expression and ChIP-chip binding data are mutually consistent. Compared to current methods that could perform the same task, the Gibbs sampling method developed here exceeds their ability at high levels (>50%) of transcription network and gene expression error, while performing similarly at lower levels. Using this technique, we show that on average 73% more gene expression features can be captured per gene as compared to the unfiltered use of gene expression and ChIP-chip-derived network connectivity data. It is important to note that the method described here can be generalized to other transcription connectivity data (e.g. sequence analysis, etc.). AVAILABILITY: Our algorithm is available on request from the authors and soon to be posted on the web. See author's homepage for details, http://www.seas.ucla.edu/~liaoj/

UNLABELLED: Network component analysis (NCA) is a method to deduce transcription factor (TF) activities and TF-gene regulation control strengths from gene expression data and a TF-gene binding connectivity network. Previously, this method could analyze a maximum number of regulators equal to the total sample size because of the identifiability limit in data decomposition. As such, the total number of source signal components was limited to the total number of experiments rather than the total number of biological regulators. However, networks that have less transcriptome data points than the number of regulators are of interest. Thus it is imperative to develop a theoretical basis that allows realistic source signal extraction based on relatively few data points. On the other hand, such methods would inherently increase numerical challenges leading to multiple solutions. Therefore, solutions to both the problems are needed. RESULTS: We have improved NCA for transcription factor activity (TFA) estimation, based on the observation that most genes are regulated by only a few TFs. This observation leads to the derivation of a new identifiability criterion which is tested during numerical iteration that allows us to decompose data when the number of TFs is greater than the number of experiments. To show that our method works with real microarray data and has biological utility, we analyze Saccharomyces cerevisiae cell cycle microarray data (73 experiments) using a TF-gene connectivity network (96 TFs) derived from ChIP-chip binding data. We compare the results of NCA analysis with the results obtained from ChIP-chip regression methods, and we show that NCA and regression produce TFAs that are qualitatively similar, but the NCA TFAs outperform regression in statistical tests. We also show that NCA can extract subtle TFA signals that correlate with known cell cycle TF function and cell cycle phase. Overall we determined that 31 TFs have statistically periodic TFAs in one or more experiments, 75% of which are known cell cycle regulators. In addition, we find that the 12 TFAs that are periodic in two or more experiments correspond to well-known cell cycle regulators. We also investigated TFA sensitivity to the choice of connectivity network we constructed two networks using different ChIP-chip p-value cut-offs. AVAILABILITY: The NCA Toolbox for MATLAB is available at http://www.seas.ucla.edu/~liaoj/download.htm.