TL;DR: A new computational study reports AI-first screening for CASP4 blockers

TL;DR: On Sept. 14, 2026, researchers reported an AI-accelerated pipeline that combines the ChemBERTa chemical language model with molecular dynamics and MM/PBSA free-energy analysis to nominate existing drugs as possible CASP4 inhibitors.

The work appeared in the Journal of Chemical Information and Modeling under the title AI Accelerated Chemical Screening Integrates ChemBERTa to Identify Repositionable CASP4 Inhibitors via MD Simulations and MM/PBSA Analysis, according to ACS Publications. The listed authors are Mubashir Hassan, Sidra Ghayour Bhatti, Muhammad Yasir, Wanjoo Chun and Andrzej Kloczkowski, according to ACS Publications. The ACS Publications listing gives a publication timestamp of Sept. 14, 2026, according to ACS Publications. The title itself describes the approach as integration of ChemBERTa screening with molecular dynamics simulations and MM/PBSA analysis focused on repositionable inhibitors of CASP4, according to ACS Publications. The study is computational and is framed around drug repositioning rather than synthesis of entirely new chemical matter, according to ACS Publications.

What was published on Sept. 14

The publication venue is the Journal of Chemical Information and Modeling, a journal published by the American Chemical Society, according to ACS Publications. The ACS Publications record places the article in volume 66, issue 19, article page 12740, according to ACS Publications. The byline names five contributors spanning computational chemistry and life-science affiliations, according to ACS Publications. The title identifies CASP4, also described in related literature as caspase-4, as the protein target, according to ACS Publications. The title identifies ChemBERTa as the artificial-intelligence component used for initial chemical screening, according to ACS Publications. The title identifies molecular dynamics simulations as the structural stability check and MM/PBSA analysis as the binding-energy estimation step, according to ACS Publications. The use of the term repositionable signals that the screened collection emphasized approved or otherwise characterized compounds that could be redirected toward a new target, according to ACS Publications.

Why CASP4 draws attention in inflammation research

CASP4 encodes human caspase-4, one of the inflammatory caspases that operate in pyroptosis, according to Trends in Pharmacological Sciences. Pyroptosis is an inflammatory and lytic form of programmed cell death, distinct from largely noninflammatory apoptosis that preserves plasma-membrane integrity during execution, according to Trends in Pharmacological Sciences. Human caspases-4 and -5, with caspase-11 as the corresponding enzyme in mice, cleave gasdermin D, according to Trends in Pharmacological Sciences. Cleavage of gasdermin D generates membrane pores that enable release of cytokines such as interleukin-1 beta and interleukin-18 and ultimately drive cell lysis, according to Trends in Pharmacological Sciences. Apoptosis is mediated mainly by initiator caspases-8 and -9 and executioner caspases-3 and -7, while pyroptosis relies on the inflammatory caspase-gasdermin axis, according to Trends in Pharmacological Sciences. Programmed cell death contributes to the pathogenesis of several diseases, which is why caspases have long appeared to be unusually attractive drug targets, according to Trends in Pharmacological Sciences. Their catalytic pockets are readily inhibited in biochemical assays, and inhibitors have shown target engagement in vitro, in vivo and in some clinical settings, according to Trends in Pharmacological Sciences.

No approved direct caspase inhibitor yet

No direct caspase inhibitor has become an approved medicine, according to Trends in Pharmacological Sciences. Caspase inhibitors have repeatedly entered clinical trials but have not reached clinical practice due to toxicity and limited efficacy, according to Trends in Pharmacological Sciences. The caspase-1 inhibitor pralnacasan was halted after liver abnormalities emerged in long-duration animal studies, according to Trends in Pharmacological Sciences. Belnacasan, another caspase-1 inhibitor, entered trials in epilepsy and inflammatory conditions without establishing persuasive clinical efficacy, according to Trends in Pharmacological Sciences. Broader pan-caspase inhibitors, including nivocasan and emricasan, failed for different reasons, including toxicity, insufficient efficacy and disease-stage mismatch, according to Trends in Pharmacological Sciences. Emricasan repeatedly moved proximal biomarkers in the expected direction, showing that caspases could be inhibited in patients, while later clinical outcomes remained unchanged, according to Trends in Pharmacological Sciences. Those outcomes do not show that caspases are undruggable, but they show that a successful next-generation inhibitor needs a targeting logic that identifies the relevant cell state and a clinical setting in which rescuing that state can still alter disease outcomes, according to Trends in Pharmacological Sciences. A computational repositioning strategy is therefore notable because it seeks starting points with prior human or animal exposure data rather than beginning from untested chemistry, according to the IJPS Journal review of AI and machine-learning models in early drug discovery.

A natural brake on caspase-4 points to therapeutic logic

A separate study published online in September 2026 described an endogenous regulatory mechanism that restrains caspase-4 and its mouse counterpart caspase-11, according to Cell Chemical Biology. Under physiological conditions, dipeptidyl peptidase 3 cleaves phosphatidylethanolamine binding protein 1 to generate an N-terminal fragment, according to Cell Chemical Biology. That fragment directly binds and suppresses caspase-4/11 and limits the response intensity of the non-canonical inflammasome, according to Cell Chemical Biology. The PEBP1 N terminus appeared in the serum of mice and patients with sepsis, indicating clinical relevance, according to Cell Chemical Biology. Mice deficient in DPP3 exhibited stronger inflammatory cytokine responses and had poor survival in a lipopolysaccharide-induced sepsis model, according to Cell Chemical Biology. Promoting DPP3 activity effectively constrained the response intensity of sepsis in mice and increased their survival, according to Cell Chemical Biology. The non-canonical inflammasome is a protein complex involved in bacterial infections, and its activation leads to excessive inflammatory responses during sepsis, according to Cell Chemical Biology. The authors concluded that pharmacologically promoting DPP3 activity or mimicking the PEBP1 inhibitory fragment could represent a therapeutic strategy to temper pathological inflammation in sepsis, according to Cell Chemical Biology. The finding underscores why dampening caspase-4 signaling is biologically plausible even as direct synthetic inhibitors remain unapproved, according to Cell Chemical Biology and Trends in Pharmacological Sciences.

How ChemBERTa reads chemistry as language

ChemBERTa is a language model for SMILES strings of molecules, according to the ChemRxiv preprint on plasma protein binding prediction under scaffold split. SMILES is a one-dimensional textual representation of chemical structure, and chemistry-aware Transformers treat SMILES as a chemical language, according to the IJPS Journal review and a 2026 practitioner summary of molecular AI. The workflow converts SMILES into tokens, passes tokens through a Transformer, produces embeddings and then predicts a molecular property, according to that practitioner summary. The same approach can potentially extend to ADMET, toxicity, bioactivity and drug-likeness prediction, according to that practitioner summary. In the IJPS Journal taxonomy, ChemBERTa, SMILES-Transformer and MolFormer belong to the Transformer and large-language-model class applied to sequence-to-sequence lead optimization and retrosynthesis, according to the IJPS Journal. Their stated advantage is capture of long-range syntax and contextual chemical grammar, according to the IJPS Journal. Graph neural networks, by contrast, directly preserve topological atom-bond graph connectivity for property prediction and binding-affinity tasks, according to the IJPS Journal. Traditional machine-learning methods such as random forest, support-vector machines and XGBoost retain high baseline accuracy on small tabular bioassay datasets, according to the IJPS Journal.

What AI screening changes about early discovery

Traditional drug discovery is described as arduous and high-risk, with multi-billion-dollar research expenditures, timelines exceeding a decade and high early-stage attrition, according to the IJPS Journal. A standard preclinical pipeline requires synthesizing and assaying thousands of chemical analogues to isolate a single viable clinical candidate, with early-stage attrition exceeding 90% due to poor pharmacokinetics, off-target toxicity or suboptimal potency, according to the IJPS Journal. Virtual screening computationally triages millions to billions of compounds to enrich active hit rates prior to physical assaying, according to the IJPS Journal. Conventional high-throughput screening yields hit rates typically between 0.01% and 0.14%, while modern deep-learning scoring functions and directed message-passing neural networks achieve enrichment rates exceeding 15% to 50% in verified biological assays, according to the IJPS Journal. In one widely cited case, researchers trained a directed message-passing neural network on 2,335 diverse molecules to predict growth inhibition against Escherichia coli, according to the IJPS Journal. The model subsequently screened over 107 million structures from the ZINC15 database and the Drug Repurposing Hub in under four days, according to the IJPS Journal. Out of 99 computationally prioritized candidates tested in wet-lab assays, 51 demonstrated potent growth inhibition, an empirical hit rate of 51.5% versus a historical high-throughput baseline near 0.05%, according to the IJPS Journal. The lead candidate Halicin illustrated how deep learning could surface structurally novel antibiotics, according to the IJPS Journal.

How dynamics and MM/PBSA vet a virtual hit

Molecular docking provides an initial assessment of how a small molecule may fit a binding site, while molecular dynamics tests whether that modeled complex remains stable over time, according to the ChemRxiv preprint on reversible covalent cathepsin S inhibitors. That preprint evaluated 100 alpha-cyanoacrylamide derivatives by rigid docking and advanced top candidates to pharmacokinetic assessment and initial 25-nanosecond simulations, according to the ChemRxiv preprint. Two compounds, F7 and F48, were then examined with density functional theory, covalent and non-covalent induced-fit docking and 100-nanosecond simulations, according to the ChemRxiv preprint. Complex stability and binding behavior were examined with root-mean-square deviation, root-mean-square fluctuation, hydrogen-bond analysis, free-energy landscape analysis, per-residue energy decomposition and molecular mechanics Poisson-Boltzmann surface area calculations, according to the ChemRxiv preprint. F7 maintained a stable binding mode with persistent interactions with key active-site residues and favorable accommodation of its alkyl substituent within the S2 pocket, according to the ChemRxiv preprint. F7 showed lower conformational fluctuations, stronger residue-level energetic contributions and a more favorable MM/PBSA binding free energy of minus 33.47 kilocalories per mole than F48 at minus 27.72 kilocalories per mole, according to the ChemRxiv preprint. The authors identified F7 as a suitable scaffold for further experimental investigation rather than as a validated drug, according to the ChemRxiv preprint. The CASP4 paper adopts the same conceptual sequence of AI prioritization followed by dynamics and MM/PBSA estimation, but directed at repositionable CASP4 inhibitors, according to ACS Publications.

Rigor checks that matter for scaffold-hopping claims

Evaluating models by scaffold split, which separates molecules by Bemis-Murcko scaffold and places each entire group into only training, validation or test sets, makes prediction harder than random splitting, according to the ChemRxiv preprint on plasma protein binding prediction. Under scaffold splitting the model predicts on unseen cores, so the task becomes more difficult than under regular random splitting, according to the ChemRxiv preprint. In that plasma-protein-binding benchmark, graph neural networks and ChemBERTa performed better than tree models on a mean basis, and fusion models performed better than individual graph or ChemBERTa-only models on a mean basis, according to the ChemRxiv preprint. Random splitting on a random-forest model exhibited a better R-squared score than scaffold-split evaluation, according to the ChemRxiv preprint. The lesson for repositioning is that performance on familiar chemical series does not guarantee recognition of distant scaffolds, according to the ChemRxiv preprint. Transformer frameworks are also being adapted to specialized affinity tasks, including a RoBERTa-based framework integrating RNA-BERTa and ChemBERTa-v2 to model molecule-RNA binding affinities, according to Briefings in Bioinformatics. That framework included six RNA class-specific models covering aptamers, repeats, ribosomal RNAs, riboswitches, microRNAs and viral RNAs plus a general model for other classes, according to Briefings in Bioinformatics. On an independent diverse test set the model achieved area-under-receiver-operating-characteristic values of 0.57 to 0.59, comparable to other evaluated methods, according to Briefings in Bioinformatics. Application to 3,492 approved drugs identified 2,859 compounds with predicted binding affinities at pKd of at least 6 across 294 RNA targets, suggesting utility for repurposing and prioritization for further investigation, according to Briefings in Bioinformatics.

What the CASP4 result does and does not establish

The Sept. 14 CASP4 report establishes a computational nomination: known compounds prioritized by a chemical language model and rescored by physics-based simulation, according to ACS Publications. It does not by itself establish clinical efficacy, safety or regulatory readiness, a distinction emphasized by the recent history of caspase inhibitors, according to Trends in Pharmacological Sciences. Even inhibitors that moved biomarkers in patients later left clinical outcomes unchanged, according to Trends in Pharmacological Sciences. Recent work on gasdermin D-mediated delivery, in which membrane pores admit otherwise membrane-impermeable inhibitors into pyroptotic cells, illustrates how cell-state selectivity may matter as much as raw potency, according to Trends in Pharmacological Sciences. The endogenous DPP3-PEBP1 brake on caspase-4/11 further suggests that timing, context and pathway specificity will shape any future intervention, according to Cell Chemical Biology. For now, the practical value of the CASP4 screen is efficiency: narrowing vast chemical space to a short list for biochemical testing, structural follow-up and independent replication, according to the IJPS Journal and ACS Publications. Wet-lab validation, selectivity profiling across caspases, pharmacokinetic assessment and disease-relevant models remain the required next layers before any repositioned candidate could be considered a therapeutic lead, according to the IJPS Journal, the ChemRxiv cathepsin S study and Trends in Pharmacological Sciences.

Sources referenced