Recent advances and breakthroughs have facilitated the design of protein binders towards specific epitopes. However, discriminating binders from non-binders has been challenging. Combining physics-based and deep-learning based approaches, we aim to further decrease the false-positive rate. Here, we used RFdiffusion, LigandMPNN, BindCraft and iterative rounds of AF3 and Boltz2 refolding cycles to design protein binders.