Papers, preprints, chapters and posters

Teaching LLMs to Generate Challenging MILP Instances via Solver Feedback

Singla, J., Pareek, P., Jawanpuria, P., & Singla, P.

Submitted to ICLR 2027, arXiv:2609.37356

Summary: Benchmarking a mixed-integer linear programming (MILP) solver needs feasible, hard instances, but most generators do not learn hardness; they inherit it from seed instances or search for it per family and size. Here a frozen SCIP solver scores each instance a language model writes by branch-and-bound nodes and the relaxation gap left after root cuts, so weaknesses the solver repairs earn little, while a validity gate rejects hardness bought with unbounded continuous variables, aggregated big-M links or extreme coefficient ranges. Trained by reinforcement learning without seed instances, OptiScribe-12B raises median nodes over base Gemma-4-12B 1.7–5.0× on capacitated facility location (CFL) and 1.9–4.5× on max-cut, and CFL feasibility by 9–19 points; the gains hold beyond training sizes and under held-out solvers, HiGHS and Gurobi. Multiple knapsack stays easy for every model.

Cross-Model Agreement as a Deployment-Time Reliability Signal for Automatic Polyp Segmentation

Gupta, S., & Singla, J.

arXiv preprint, arXiv:2609.10495

Summary: In real-time colonoscopy there is no ground truth at inference, so a segmentation model that fails does so silently. RBQE scores a prediction by how far a second, independently trained referee agrees with it on the same image, needing no reference mask. Independent training alone turns out to be enough: a referee identical in architecture and differing only in random initialisation reaches ROC-AUC 0.923, while a cross-architecture SegFormer-B0 referee reaches 0.960 and beats test-time augmentation by 0.055 under the same protocol. Rejecting the low-agreement cases raises the mean Dice of what is kept, at the cost of one extra forward pass.

From genome to function: Identification and characterization of bifunctional pyrophosphate-fructose-6-phosphate 1-phosphotransferase from Candidatus Liberibacter asiaticus

Sandra, K. P., Arun, E., Rathore, K., Gubyad, M., Ghosh, D. K., Sharma, A. K., & Singla, J.

bioRxiv, 2026-09

Summary: Candidatus Liberibacter asiaticus causes citrus greening, and long association with its host has left central carbon metabolism broken in two places — no phosphoglucose isomerase in glycolysis, no transaldolase in the pentose phosphate pathway. PFP is the enzyme that bridges them, and it turns out to do two jobs. Recombinant CLas PFP was expressed, purified and measured: it binds fructose 6-phosphate and sedoheptulose 7-phosphate with comparable affinity, Kₘ 66.6 and 57.5 µM, which confirms the second reaction and with it the sedoheptulose bisphosphate route. The protein loses conformational stability above 50 °C and at pH 11.

Coarse composition suffices: tabular in-context learning for multi-activity antimicrobial peptide profiling

Kumar, R., Pal, A., Solanki, D., Pareek, P., Singh, J., & Singla, J.

bioRxiv, 2026-08

Summary: Antimicrobial peptides usually act on more than one class of pathogen, so the question worth asking of a candidate is which activities it has, not whether it is antimicrobial at all. The leading methods on the ESCAPE benchmark answer it with multimodal, structure-conditioned deep models. Pairing 330 interpretable sequence descriptors with TabPFN — a tabular foundation model that predicts in a single forward pass, with no gradient training and no hyperparameter search — beats them, reaching 77.8% five-label macro average precision against a previous best of 72.1%, with the largest gains on remote homologues. Predicted structure turns out to be unnecessary at inference, and ten global physicochemical scalars alone recover 91% of the full-feature performance.

RECAST: Reconstruction-Anchored Semi-Supervised Training for Low-Resource Machine Translation

Bathla, D., Singla, P., & Singla, J.

Submitted to AAAI-27

Summary: Low-resource translation has far more monolingual text than parallel text, so most of the data carries no reference to train against. RECAST translates a sentence and then reconstructs it, and uses the agreement between the original and the reconstruction as the training signal — the round trip stands in for the reference the unlabelled half does not have.

FiRe: Frequency Reparameterization as a Preconditioner for Periodic Implicit Neural Representations

Shukla, H., Verma, R., & Singla, J.

arXiv preprint, arXiv:2606.29414

Summary: Periodic implicit neural representations such as SIREN give every neuron the same global frequency, which spends the representational budget badly when local signal content varies. FiRe gives each neuron a bounded, input-dependent frequency through a separate low-rank gate, leaving the activation function itself untouched. The gate acts as an implicit preconditioner that improves the conditioning of the problem at initialisation, so optimisation converges sooner at a fixed budget.

Morphotype-Resolved 3D Morphometry Reveals a Structure–Density–Location Coupling in Mitochondrial Networks

Singh, A., Yadav, A., Deshmukh, A., Singh, A., Verma, R., White, K., & Singla, J.

Structure, Accepted

Summary: Mitochondrial networks in whole β-cells are resolved into distinct morphotypes and measured in three dimensions from soft X-ray tomograms. Across the population, structure, density and subcellular location are found to vary together rather than independently — a coupling that a population-averaged measurement would miss.

Pathology-Aware Multi-View Contrastive Learning for Patient-Independent ECG Reconstruction

Youssef, Y., & Singla, J.

arXiv preprint, arXiv:2603.17248

Summary: Reconstructing a full 12-lead ECG from a handful of leads is ill-posed: anatomy differs from patient to patient, so a model fitted to one person's geometry does not transfer to the next. Methods that ignore the underlying pathology lose exactly the morphology the precordial leads are read for. This regularises the latent space against a pathological manifold instead, aligning waveform features with clinical labels by supervised contrastive learning so that anatomy is treated as the nuisance variable it is. On PTB-XL the patient-independent RMSE falls by about 76% against the previous best, and the gain holds on the PTB Diagnostic Database.

Samasamayik: A Parallel Dataset for Hindi–Sanskrit Machine Translation

Karthika, N. J., Suryanarayanan, K., Purohit, J., Ramakrishnan, G., Singla, J., & Gourishetty, A. K.

arXiv preprint, arXiv:2603.24307

Summary: The same argument carried to a second language pair: 92,196 Hindi–Sanskrit parallel sentences, aggregated from spoken tutorials, children's magazines, radio conversations and instruction material rather than from classical texts. Fine-tuning ByT5, NLLB and IndicTrans-v2 on it gives significant in-domain gains while holding up on the usual test sets, and a comparison against existing corpora shows minimal semantic and lexical overlap — the resource is new, not a repackaging.

Structural and functional characterization of thermostable EstS1 esterase for BHET degradation

Verma, S., Aggarwal, D., Ashar, M., Pandey, A. K., Sutradhar, A., Pandey, S., Sircar, D., Singla, J., & Kumar, P.

Journal of Structural Biology, 108342

Summary: BHET is the bottleneck intermediate in breaking PET plastic back down to its monomers. Crystal structures of the thermostable esterase EstS1 from Sulfobacillus acidophilus catch BHET and its products in the active-site tunnel, and the biochemistry shows the enzyme carries the reaction through to terephthalate rather than stalling at MHET — a single enzyme for a step usually split between two.

Robust mitochondria segmentation and morphological profiling using soft X-ray tomography

Yadav, A., Singh, A., Deshmukh, A., Bharadwaj, P., Baliyan, A., White, K., & Singla, J.

Journal of Structural Biology, 108291

Summary: A pipeline for turning soft X-ray tomograms into measurable biology: mitochondria are separated from the surrounding volume, the mask is split into individual organelles, and each one is profiled for size, shape and connectivity. The emphasis is on segmentation that holds up across volumes rather than being tuned to a single dataset.

YOLO-LAN: Precise Polyp Detection via Optimized Loss, Augmentations and Negatives

Gupta, S., & Singla, J.

arXiv preprint, arXiv:2509.19166

Summary: Polyps missed during colonoscopy are the ones that become colorectal cancer. YOLO-LAN is a detection pipeline tuned on the three things its name names — an M2IoU loss, wide-ranging augmentations, and negative frames containing no polyp at all, so the model sees the clinic rather than a curated set. On Kvasir-seg and BKAI-IGH NeoPolyp the gain lands where it matters, at strict IoU: detections that are precisely placed, not merely present.

DSEM-NIDS: Enhanced Network Intrusion Detection System Using Deep Stacking Ensemble Model

Mahmoud, L., Liyanage, M., Singla, J., & Gangopadhyay, S.

IEEE Open Journal of the Computer Society, 1–12

Summary: A single intrusion detector carries its own bias and variance into every decision, and on traffic at internet scale that shows up as false alarms and missed attacks. DSEM-NIDS stacks several different base models under a nested meta-learner that learns how to combine them, rather than averaging or voting. It is evaluated across four datasets — 5G-NIDD, UNR-IDD, N-BaIoT and NSL-KDD — spanning 5G, enterprise, IoT and classical traffic.

Structural Virology: The Key Determinants in Development of Antiviral Therapeutics

Handa, T., Saha, A., Narayanan, A., Ronzier, E., Kumar, P., Singla, J., & Tomar, S.

Viruses, 17(3), 417

Summary: A review of what structure buys antiviral development. Resolving capsids, envelope proteins, replication machinery and host-interaction interfaces at atomic or near-atomic resolution is what turned viral proteases, polymerases and integrases into drug targets for HIV, SARS-CoV-2 and influenza. X-ray crystallography, NMR, cryo-EM and cryo-electron tomography each contribute, and the structural diversity and high mutation rates of viruses are precisely why that work has to keep pace.

Mechanistic and structural insights into EstS1 esterase: A potent broad-spectrum phthalate diester degrading enzyme

Verma, S., Choudhary, S., Kumar, K. A., Mahto, J. K., Mishra, I., Prakash, V. B., et al.

Structure, 33(2), 247–261.e3

Summary: Phthalate diesters are important pollutants and act as endocrine disruptors, and while certain bacterial esterases degrade them to monoesters, the structural and mechanistic basis was largely unexplored. High-resolution structures of the thermostable, pH-tolerant EstS1 — apo, and with substrates, products and analogs bound — show a catalytic tunnel carrying substrate in and product out, and a central Ser-His-Asp triad running a bi-bi ping-pong mechanism through a tetrahedral intermediate. Mutating Met207 to alanine abolishes DEHP binding at the active site altogether.

Designing and bioengineering of CDRs with higher affinity against receptor-binding domain (RBD) of SARS-CoV-2 Omicron variant

Singh, V., Choudhary, S., Bhutkar, M., Nehul, S., Ali, S., Singla, J., et al.

International Journal of Biological Macromolecules, 290, 138751

Summary: Nanobodies raised against the original Wuhan RBD bind the Omicron RBD poorly, and raising a new one for every variant is slow. Three of them — H11-H4, C5 and H3 — are rebuilt instead: close to two thousand mutations across their complementarity-determining regions, screened by protein-protein docking and then tested for complex stability under molecular dynamics, leaving seventeen candidates. Those were expressed and measured by isothermal titration calorimetry, and five H11-H4 mutants bind Omicron RBD at Kᴅ 8.8–27 µM against the parent's 32 µM. The pipeline is the result: repurposing an existing nanobody is cheaper than finding a new one.

Structure-guided mutations in CDRs for enhancing the affinity of neutralizing SARS-CoV-2 nanobody

Singh, V., Bhutkar, M., Choudhary, S., Nehul, S., Kumar, R., Singla, J., et al.

Biochemical and Biophysical Research Communications, 734, 150746

Summary: The same question approached from a single structure rather than a whole library: 112 structure-guided single mutations across the CDRs, docked against the spike RBD, narrowed to 32, then to nine under molecular dynamics, then to six that calorimetry confirms bind more tightly than the native nanobody. The one that matters is Leu106Thr, which neutralises SARS-CoV-2 pseudovirus at IC₅₀ 0.03 µM where the parent needs 0.77 µM — roughly twenty-five-fold, from one residue.

In-silico characterization of a hypothetical protein of Sulfobacillus sp. hq2 for degradation of phthalate diesters

Verma, S., Singh, A., Kumar, P., & Singla, J.

International Journal of Biological Macromolecules, 280, 136006

Summary: Phthalate plasticizers leach out of PVC and disrupt endocrine function, but few esterases degrade a wide enough range of them, and thermostability — needed for in-situ use — narrows the field further. This study takes an unannotated protein, POB10642.1 from a thermostable Sulfobacillus strain, and argues that it is one: 42.67% sequence identity and 0.557 Å RMSD to the known EstS1 esterase, a Tm of 55–66 °C, two surface cavities forming a catalytic tunnel, and stable binding to eight different phthalate diesters across 100 ns of simulation.

Characterization of haloacid dehalogenase superfamily acid phosphatase from Staphylococcus lugdunensis

Kaur, H., Rode, S., Sandra, K. P., Mahto, J. K., Alam, M. S., Gupta, D. N., et al.

Archives of Biochemistry and Biophysics, 753, 109888

Summary: Staphylococcus lugdunensis carries more than ten haloacid dehalogenase superfamily proteins in its genome and, before this, not one of them had been characterised. SLHAD1 turns out to be a metal-dependent acid phosphatase that strips phosphate from intermediates of glycolysis, gluconeogenesis, nucleotide and thiamine metabolism; the substrate preferences and the genomic context together point to thiamine metabolism as its physiological job. The crystal structure was solved at 1.7 Å with a homodimer in the asymmetric unit — of a form missing 49 C-terminal residues, because the protein cleaves itself at one specific point over time.

Sāmayik: A Benchmark and Dataset for English–Sanskrit Translation

Maheshwari, A., Gupta, A., Krishna, A., Singh, A. K., Ramakrishnan, G., Kumar, G. A., & Singla, J.

LREC-COLING 2024, 14298–14304

Summary: Sanskrit is a classical language still in sustenance, but the digitised Sanskrit that exists is overwhelmingly poetry, which leaves models poorly matched to how the language is written now. Sāmayik is around 53,000 English–Sanskrit sentence pairs in contemporary prose, curated from language instruction material, teaching pedagogy and online tutorials. Models trained on it improve significantly on out-of-domain contemporary text, outperforming models trained on classical-era poetry.

Database Evolution, by Scientists, for Scientists: A Case Study

Schuler, R. E., Singla, J., Vallat, B., White, K. L., Berman, H. M., & Kesselman, C.

2023 IEEE 19th International Conference on e-Science, 1–10

Summary: Scientific database designs go stale within months of use, and there is little evidence that scientists have the tools or the process to evolve them without leaning on a database administrator. This paper offers a simplified evolution methodology aimed at scientists, then follows one scientist evolving a research database for cell modeling as new requirements arrived, with a detailed analysis of what they actually did — showing that a complex information system can be evolved by the researcher who depends on it.

On Learning with LAD

Jothishwaran, C. A., Srivastava, B., Singla, J., & Gangopadhyay, S.

arXiv preprint, arXiv:2309.16630

Summary: Logical analysis of data yields two-class classifiers as Boolean functions in disjunctive normal form. Those rules are found by optimization, which ought to invite overfitting, and yet it does not happen. The justification offered here is an estimate of the VC dimension for LAD models whose hypothesis sets are DNFs with a small number of cubic monomials — the capacity to overfit simply is not there — illustrated and confirmed empirically.

A New Visual Design Language for Biological Structures in a Cell

Singla, J., Burdsall, K., Cantrell, B., Halsey, J. R., McDowell, A., McGregor, C., et al.

Structure, 30(4), 485–497.e3

Summary: Part of the effort to build a spatiotemporal model of the pancreatic β-cell is an immersive experience, World in a Cell. It needs a way to draw molecules and organelles in crowded cellular environments that a machine can also animate in real time: a visual design language built from tetrahedral blocks, which carries the structural features while staying cheap enough for animation and user interaction.

Soft X-ray tomography to map and quantify organelle interactions at the mesoscale

Loconte, V., Singla, J., Li, A., Chen, J. H., Ekman, A., McDermott, G., Sali, A., White, K. L., & Larabell, C. A.

Structure, 30(4), 510–521.e3

Highlighted in Nature — “Soft X-rays capture the dance of the organelles”

Summary: Soft X-ray tomography sees an intact cell with no stain at all: the contrast is the differential absorption of the carbon-rich compounds in each organelle. Here it follows insulin vesicles and their interaction with mitochondria in β cells during secretion, quantifying morphology, composition and relative position from one reconstruction.

An intensity-based post-processing tool for 3D instance segmentation of organelles in soft X-ray tomograms

Li, A., Zhang, S., Loconte, V., Liu, Y., Ekman, A., Thompson, G. J., Sali, A., Stevens, R. C., White, K., Singla, J., & Sun, L.

PLOS ONE, 17(9), e0269887

Summary: A semantic mask says which voxels are mitochondrion and which are insulin vesicle; it does not say where one organelle ends and the next begins. This post-processing tool separates touching instances from the raw tomogram's intensity together with the mask and the organelle's own morphology, handling sphere-like vesicles and columnar mitochondria alike — more accurately, on synthetic and β-cell data, than connected-region labelling or watershed.

Auto-segmentation and time-dependent systematic analysis of mesoscale cellular structure in β-cells during insulin secretion

Li, A., Zhang, X., Singla, J., White, K. L., Loconte, V., Hu, C., et al.

PLOS ONE, 17(3), e0265567

Summary: The bottleneck in soft X-ray tomography is not the imaging but the labour-intensive manual segmentation that follows it. This pipeline pairs semantic segmentation with a first-applied instance step to give separate organelle masks at high Dice and Recall, then locates those organelles by their radial distribution. Across INS-1E β-cells at several time points it recovers the effect of glucose on insulin vesicles and mitochondria.

Bayesian metamodeling of complex biological systems across varying representations

Raveh, B., Sun, L., White, K. L., Sanyal, T., Tempkin, J., Zheng, D., et al.

Proceedings of the National Academy of Sciences, 118(35), e2104559118

Summary: Modelling a whole cell means integrating models that agree on neither scale, granularity, data type, nor even mathematical representation. Bayesian metamodeling divides that problem: each input model is converted to a standardized statistical form as a probabilistic graphical model, coupled to the others through their mutual relations with the physical world, and then harmonized against them. The proof of principle assembles six such models — from vesicle trafficking to systemic insulin response — into one account of glucose-stimulated insulin secretion, resolving the places where they conflict.

A community approach to whole-cell modeling

Singla, J., & White, K. L.

Current Opinion in Systems Biology, 26, 33–38

Summary: A model of a whole cell has to hold everything at once — structure across scales, metabolic networks, protein signalling — which puts it beyond any one discipline and any one group. This review takes stock of where the field has reached and argues that the binding constraint is not the modelling but the data: the inputs such a model needs have to be organised, annotated and archived before they can be used at all, and that is work no single lab can do. The case it makes is for community standards, drawing on earlier efforts that solved the same problem by agreeing on them.

Assessment of scoring functions to rank the quality of 3D subtomogram clusters from cryo-electron tomography

Singla, J., White, K. L., Stevens, R. C., & Alber, F.

Journal of Structural Biology, 213(2), 107727

Summary: Template-free clustering of subtomograms can turn up complexes nobody thought to look for — but only if the clusters can be ranked, and real ones carry both alignment error and contamination from misclassified complexes. Of the more than fifteen scoring functions assessed here, most proved sensitive to signal-to-noise ratio and needed Gaussian filtering first; two did not, and spectral-SNR-based Fourier shell correlation is also the fastest to compute.

Visualizing subcellular rearrangements in intact β cells using soft X-ray tomography

White, K. L., Singla, J., Loconte, V., Chen, J. H., Ekman, A., Sun, L., et al.

Science Advances, 6(50), eabc8262

Highlighted in Science — “Subcellular map of vesicle maturation”

Summary: Whole pancreatic β cells reconstructed in three dimensions by soft X-ray tomography at time points after glucose-stimulated insulin secretion. Glucose rapidly changes the density of insulin packing, increases mitochondrial volume, and brings vesicles closer to mitochondria; exendin-4 prolongs those effects and raises vesicle maturation.

Visualizing insulin vesicle neighborhoods in β cells by cryo-electron tomography

Zhang, X., Carter, S. D., Singla, J., White, K. L., Butler, P. C., Stevens, R. C., & Jensen, G. J.

Science Advances, 6(50), eabc8258

Summary: Subcellular neighborhoods — local mixtures of organelles and proteins in particular ratios — matter most in cells that secrete, yet their part in insulin vesicle maturation was unclear. Several distinct tomograms per single β cell put it in reach: sampling from near the nucleus outward to the plasma membrane, both the volume fraction occupied by insulin vesicles and their average diameter rise.

De novo structural pattern mining in cellular electron cryotomograms

Xu, M., Singla, J., Tocheva, E. I., Chang, Y. W., Stevens, R. C., Jensen, G. J., & Alber, F.

Structure, 27(4), 679–691

Highlighted in Nature Methods — “Template-free visual proteomics”

Summary: A cellular cryotomogram holds a plethora of macromolecular complexes at once, and the usual way to pull them out is to match against a template of something already known — which by construction cannot find anything new. Multi-Pattern Pursuit works the other way round, discovering complexes de novo from the heterogeneous particles extracted from entire tomograms, and the classes it finds are good enough to seed more targeted refinement.

Opportunities and challenges in building a spatiotemporal multi-scale model of the human pancreatic β cell

Singla, J., McClary, K. M., White, K. L., Alber, F., Sali, A., & Stevens, R. C.

Cell, 173(1), 11–19

Summary: A predictive model of an entire eukaryotic cell — its dynamic structure from atomic to cellular scales — is a grand challenge sitting at the intersection of biology, chemistry, physics and computer science. This sets out the vision for one such model of the pancreatic β cell, a target relevant to understanding and modulating the pathogenesis of diabetes, and argues that the experimental and modelling methods it needs are best developed as a community effort.

Understanding the functional impact of copy number alterations in breast cancer using a network modeling approach

Srihari, S., Kalimutho, M., Lal, S., Singla, J., Patel, D., Simpson, P. T., Khanna, K. K., & Ragan, M. A.

Molecular BioSystems, 12(3), 963–972

Summary: Copy-number alterations account for most of the expression variation between breast tumours, but the genes they move are not all next to the alteration — some are affected from across the genome. This integrates copy-number and expression profiles in one network to separate the two, and the genes it picks out reconstruct all ten METABRIC subtypes, each driven by its own set. Alongside known drivers such as CCND1, ERBB2 and MDM2 it recovers BRF2 and SF3B3; knocking BRF2 down costs ER-/HER2+ cells their viability and leaves normal breast cells alone, which makes it a candidate target rather than a correlate.

Inferring synthetic lethal interactions from mutual exclusivity of genetic events in cancer

Srihari, S., Singla, J., Wong, L., & Ragan, M. A.

Biology Direct, 10(1), 1–18

Summary: Synthetic lethality is a good therapeutic idea with a poor record in prediction: most candidate pairs come from screens in model organisms and either fail in human cells or involve genes human tumours never alter. This infers the pairs from the tumours directly, on the argument that two genes almost never altered in the same tumour are the pair a cell cannot afford to lose at once. Across 3,980 breast, prostate, ovarian and uterine samples from TCGA it finds 718 genes likely to be synthetic lethal with six DNA-damage response genes, and those genes land in the top quartile of essentiality in ten DDR-deficient cell lines — TLK2 among them.