Background: Since the finishing of the sequencing of the Arabidopsis thaliana genome, the Arabidopsis community and the annotator centers have been working on the improvement of gene annotation at the structural and functional levels. In this context, we have used the large CATMA resource on the Arabidopsis transcriptome to search for genes missed by different annotation processes. Probes on the CATMA microarrays are specific gene sequence tags (GSTs) based on the CDS models predicted by the Eugene software. Among the 24 576 CATMA v2 GSTs, 677 are in regions considered as intergenic by the TAIR annotation. We analyzed the cognate transcriptome data in the CATMA resource and carried out data-mining to characterize novel genes and improve gene models. Results: The statistical analysis of the results of more than 500 hybridized samples distributed among 12 organs provides an experimental validation for 465 novel genes. The hybridization evidence was confirmed by RT-PCR approaches for 88% of the 465 novel genes. Comparisons with the current annotation show that these novel genes often encode small proteins, with an average size of 137 aa. Our approach has also led to the improvement of pre-existing gene models through both the extension of 16 CDS and the identification of 13 gene models erroneously constituted of two merged CDS. Conclusion: This work is a noticeable step forward in the improvement of the Arabidopsis genome annotation. We increased the number of Arabidopsis validated genes by 465 novel transcribed genes to which we associated several functional annotations such as expression profiles, sequence conservation in plants, cognate transcripts and protein motifs.