Designing sequences for specific protein backbones is a keystep in creating new functional proteins. Here, we introduceGeoSeqBuilder, a deep learning framework that integratesprotein sequence generation with side chain conformationprediction to produce the complete all-atom structures fordesigned sequences. GeoSeqBuilder uses spatial geometricfeatures from protein backbones and explicitly includesthree-body interactions of neighboring residues. GeoSeqBuilderachieves native residue type recovery rate of 51.6%,comparable to ProteinMPNN and otherleadingmethods,while accurately predicting side chain conformations. Wefirst used GeoSeqBuilder to design sequences for thioredoxinand a hallucinated three-helical bundle protein. All the15 tested sequences expressed as soluble monomeric proteinswith high thermal stability, and the 2 high-resolutioncrystal structures solved closely match the designed models.The generated protein sequences exhibit low similarity(minimum 23%) to the original sequences, with significantlyaltered hydrophobic cores. We further redesigned the hydrophobiccore of glutathione peroxidase 4, and 3 of the 5designs showed improved enzyme activity. Although furthertesting is needed, the high experimental success ratein our testing demonstrates that GeoSeqBuilder is a powerfultool for designing novel sequences for predefined proteinstructures with atomic details. GeoSeqBuilder is availableat https://github.com/PKUliujl/GeoSeqBuilder.