DNA interstrand crosslinks (ICLs) covalently join opposing strands, blocking both replication and transcription, therefore making ICL-inducing compounds highly toxic and ideal anti-cancer agents. While incisions surrounding the ICL are required to remove damaged DNA, it is currently unclear which endonucleases are needed for this key event. SNM1A has been shown to play an important function in human ICL repair, however its suggested role has been limited to exonuclease activity and not strand incision. Here we show that SNM1A has endonuclease activity, having the ability to cleave DNA structures that arise during the initiation of ICL repair. In particular, this endonuclease activity cleaves single-stranded DNA. Given that unpaired DNA regions occur 5′ to an ICL, these findings suggest SNM1A may act as either an endonuclease and/or exonuclease during ICL repair. This finding is significant as it expands the potential role of SNM1A in ICL repair.

SNM1A is a nuclease required to repair DNA interstrand cross-links (ICLs) caused by some anticancer compounds, including cisplatin. Unlike other nucleases involved in ICL repair, SNM1A is not needed to restore other forms of DNA damage. As such, SNM1A is an attractive target for selectively increasing the efficacy of ICL-based chemotherapy. Using a fluorescence-based exonuclease assay, we screened a bioactive library of compounds for inhibition of SNM1A. Of the 52 compounds initially identified as hits, 22 compounds showed dose–response inhibition of SNM1A. An orthogonal gel-based assay further confirmed nine small molecules as SNM1A nuclease activity inhibitors with IC 50 values in the mid-nanomolar to low micromolar range. Finally, three compounds showed no toxicity at concentrations able to significantly potentiate the cytotoxicity of cisplatin. These compounds represent potential leads for further optimization to sensitize cells toward chemotherapeutic agents inducing ICL damage.

3-Deoxy-D-arabinoheptulosonate-7-phosphate (DAHP) synthase catalyzes the first step in the shikimate biosynthetic pathway and is an antimicrobial target. We used an inhibitor-in-pieces approach, based on the previously reported inhibitor DAHP oxime, to screen inhibitor fragments in the presence and absence of glycerol 3-phosphate to occupy the distal end of the active site. This led to DAHP hydrazone, the most potent inhibitor to date, K i = 10 ± 1 nM. Three trifluoropyruvate (TFP)-based inhibitor fragments were efficient inhibitors with ligand efficiencies of up to 0.7 kcal mol −1 /atom compared with 0.2 kcal mol −1 /atom for a typical good inhibitor. The crystal structures showed the TFPbased inhibitors binding upside down in the active site relative to DAHP oxime, providing new avenues for inhibitor development. The ethyl esters of TFP oxime and TFP semicarbazone prevented E. coli growth in culture with IC 50 = 0.21 ± 0.01 and 0.77 ± 0.08 mg mL −1 , respectively. Overexpressing DAHP synthase relieved growth inhibition, demonstrating that DAHP synthase was the target. Growth inhibition occurred in media containing aromatic amino acids, suggesting that growth inhibition was due to depletion of some other product(s) of the shikimate pathway, possibly folate.

α-Carboxyketose synthases, including 3-deoxy-Darabinoheptulosonate 7-phosphate synthase (DAHPS), are longstanding targets for inhibition. They are challenging targets to create tight-binding inhibitors against, and inhibitors often display half-of-sites binding and partial inhibition. Half-of-sites inhibition demonstrates the existence of inter-subunit communication in DAHPS. We used X-ray crystallography and spatially resolved hydrogen−deuterium exchange (HDX) to reveal the structural and dynamic bases for inter-subunit communication in Escherichia coli DAHPS(Phe), the isozyme that is feedback-inhibited by phenylalanine. Crystal structures of this homotetrameric (dimer-ofdimers) enzyme are invariant over 91% of its sequence. Three variable loops make up 8% of the sequence and are all involved in inter-subunit contacts across the tight-dimer interface. The structures have pseudo-twofold symmetry indicative of inter-subunit communication across the loose-dimer interface, with the diagonal subunits B and C always having the same conformation as each other, while subunits A and D are variable. Spatially resolved HDX reveals contrasting responses to ligand binding, which, in turn, affect binding of the second substrate, erythrose-4phosphate (E4P). The N-terminal peptide, M1−E12, and the active site loop that binds E4P, F95−K105, are key parts of the communication network. Inter-subunit communication appears to have a catalytic role in all α-carboxyketose synthase families and a regulatory role in some members.

The Vision Transformer (ViT) model is built on the assumption of treating image patches as "visual tokens" and learning patch-to-patch attention. The patch embedding based tokenizer is a workaround in practice and has a semantic gap with respect to its counterpart, the textual tokenizer. The patch-to-patch attention suffers from the quadratic complexity issue, and also makes it non-trivial to explain learned ViT models. To address these issues in ViT models, this paper proposes to learn patch-to-cluster attention (PaCa) based ViT models. Queries in our PaCa-ViT are based on patches, while keys and values are based on clustering (with a predefined small number of clusters). The clusters are learned end-to-end, leading to better tokenizers and realizing joint clustering-for-attention and attention-for-clustering when deployed in ViT models. The quadratic complexity is relaxed to linear complexity. Also, directly visualizing the learned clusters can reveal how a trained ViT model learns to perform a task (e.g., object detection). In experiments, the proposed PaCa-ViT is tested on CIFAR-100 and ImageNet-1000 image classification, and MS-COCO object detection and instance segmentation. Compared with prior arts, it obtains better performance in classification and comparable performance in detection and segmentation. It is significantly more efficient in COCO due to the linear complexity. The learned clusters are also semantically meaningful and shed light on designing more discriminative yet interpretable ViT models.