Highway　bridges　supported　by　the　unbonded　laminated　rubber　bearings　(ULRBs)　are　critical　components　of　the　transportation　infrastructure　systems.　Accurate　and　efficient　assessment　of　damage　states　and　seismic　response　prediction　for　individual　bridges　at　regional　scale　is　essential　for　evaluating　seismic　vulnerability　and　enhancing　infrastructure　resilience.　However,　such　assessments　are　typically　resource-intensive,　requiring　significant　labor,　time,　and　computational　power.　This　study　introduces　an　approach　based　on　machine　learning　(ML)　algorithms　to　develop　predictive　models　for　rapid　damage　states　assessment　and　seismic　response　prediction,　incorporating　various　types　of　input　data.　The　ML　includes　five　ensemble　learning　algorithms　and　three　deep　neural　networks　(DNNs),　notably　TabNet,　a　novel　architecture　optimized　for　tabular　data　processing.　Using　a　comprehensive　dataset　generated　through　nonlinear　time-history　analyses　(NTHA)　of　120　ULRB-supported　highway　bridges　subjected　to　320　ground　motions,　a　total　of　38,400　data　samples　were　obtained　for　model　training　and　evaluation.　Results　demonstrate　that　TabNet　outperformed　other　models　in　predicting　damage　states,　achieving　accuracies　of　93.6　%　and　90.5　%　across　two　test　sets.　For　predicting　the　displacement　of　ULRBs,　CatBoost　demonstrated　superior　performance,　achieving　R2　values　of　0.949　and　0.905　for　the　two　test　sets.　Furthermore,　model　interpretability　was　enhanced　using　SHAP　analysis,　which　identified　the　ULRB　friction　coefficient,　elastic　stiffness,　and　spectral　accelerations　at　0.8　s　and　1.0　s　as　key　predictors　for　bridge　damage　states　and　seismic　response.　This　study　represents　the　first　attempt　to　apply　TabNet　for　damage　assessment　and　response　prediction　in　highway　bridges.　The　results　provide　a　valuable　reference　for　applying　ML　techniques　to　evaluation　of　highway　bridge　performance　under　earthquake　events.