VAV1 itself is a compelling but difficult target. As a guanine nucleotide exchange factor, it functions as a molecular switch for Rho-family GTPases, transmitting signals downstream of the T-cell receptor and B-cell receptor complexes. When immune cells encounter antigen, VAV1 activation triggers a cascade of events: cytoskeletal rearrangement that enables cell movement and immune synapse formation, transcriptional programs that drive proliferation, and the release of cytokines such as interleukin-2 and interferon-gamma that amplify inflammatory responses. Because VAV1 sits near the top of this signaling hierarchy, dampening its activity can broadly restrain pathological immune activation, which is precisely what autoimmune diseases like rheumatoid arthritis and inflammatory bowel disease require. At the same time, its role in normal immune function means that any therapeutic strategy must carefully calibrate the degree and duration of suppression.

The historical difficulty in drugging VAV1 stems from its protein architecture. Like many signaling adaptors and enzymes, VAV1 lacks the deep, well-defined hydrophobic pockets that classical small-molecule inhibitors exploit. Its catalytic domains interact with partners through relatively flat, extended surfaces, and the protein’s activity is regulated through conformational changes and phosphorylation rather than ligand binding in a conventional active site. Traditional medicinal chemistry approaches, which depend on identifying a stable pocket to anchor an inhibitor, have therefore struggled. Molecular glue degraders sidestep this problem entirely: rather than needing to block VAV1’s function directly, they recruit the protein to the cellular degradation machinery, eliminating it altogether. The therapeutic effect derives from the loss of the protein rather than from inhibition of its enzymatic activity, opening targets that were previously considered inaccessible.

The peripheral restriction strategy embedded in ISM9546’s design deserves particular attention. Many immune-cell signaling proteins, including VAV1, are expressed in the central nervous system, where they participate in neuronal development, synaptic plasticity, and glial cell function. A degrader that crosses the blood-brain barrier could theoretically interfere with these processes, raising safety concerns that have complicated other immunomodulatory programs. By engineering the molecule to minimize brain exposure, confirmed through both in vitro transport assays using MDCKII-MDR1 cell lines and in vivo tissue distribution studies, the design team addressed a risk category that regulators and clinicians scrutinize closely, particularly for chronic therapies intended for long-term use in conditions like inflammatory bowel disease.

The pharmacokinetic profile described for ISM9546 also reflects a nuanced understanding of immunology. A balanced elimination half-life serves two purposes simultaneously. On one hand, sustained peripheral exposure maintains continuous target engagement, which is necessary because degraded VAV1 must remain suppressed for the therapeutic benefit to persist. On the other hand, the ability to allow prompt recovery of adaptive immune responses if treatment is interrupted provides a critical safety valve. Patients with autoimmune disease remain vulnerable to infections, and a degrader whose effects reverse quickly upon discontinuation offers clinicians more flexibility than agents with irreversible mechanisms or very long residence times. This reversibility-by-withdrawal property is inherent to the degradation mechanism itself, since new VAV1 protein is continuously synthesized by cells and will reaccumulate once the degrader clears.

The preclinical efficacy data reported for ISM9546 highlight several methodological points worth noting. The Disease Activity Index used in the acute colitis model is a composite measure that typically incorporates body weight change, stool consistency, and fecal blood, providing a longitudinal, non-invasive readout of disease severity. Observing significant improvement within days of treatment initiation suggests rapid target engagement and a swift downstream effect on inflammatory pathways, which is consistent with the expectation that removing a proximal signaling node like VAV1 should quickly curtail cytokine-driven inflammation. The approximately 60 percent improvement in histopathology scores provides an independent, tissue-level confirmation, as histological assessment captures cellular infiltration, epithelial damage, and architectural disruption that behavioral or clinical scores may miss. The comparison against a reference VAV1 degrader adds an important benchmarking dimension, indicating that the compound’s profile was evaluated not merely against placebo but against an active comparator in the same target class.

The proteomics profiling cited in the safety evaluation addresses one of the most significant liabilities of the degradation modality broadly. Because molecular glues work by recruiting proteins to E3 ligases, there is always a possibility that endogenous proteins other than the intended target will also be pulled into the degradation complex, a phenomenon known as neosubstrate degradation. Off-target protein depletion can produce unexpected toxicities that only become apparent in later development stages. A clean degradation profile, in which quantitative proteomics shows no significant depletion of proteins other than VAV1, is therefore a meaningful differentiator, though it is worth remembering that such profiling is typically performed in a limited set of cell types and conditions, and the full off-target landscape may only emerge through broader clinical exposure.

The 14-day toxicity studies in both rodents and non-human primates represent a standard early safety package, though the duration means they capture acute and subchronic effects rather than the chronic safety questions that will matter most for a therapy intended for long-term autoimmune disease management. Non-human primate data are particularly informative for degraders because cross-species differences in E3 ligase biology and protein turnover rates can affect both efficacy and toxicity predictions. The reported broad safety margins suggest the compound tolerated doses well above those expected to be pharmacologically active, which provides headroom for clinical dosing decisions.

The design workflow that produced ISM9546 illustrates how generative chemistry platforms have matured. The described process began with structural information about the target-ligand complex, from which the team defined a pharmacophore capturing the essential interaction points between the glue molecule, the target protein, and presumably the E3 ligase interface. Fragment growing, the de novo expansion of a small starting scaffold, was then guided by the generative models within Chemistry42, which reportedly integrates more than 40 generative architectures alongside predictive modules for properties such as permeability, metabolic stability, and brain penetration. The multi-objective nature of the optimization is notable: the team had to simultaneously improve potency, maintain selectivity, and reduce central nervous system exposure, objectives that frequently conflict in medicinal chemistry. Iterative design-make-test-analyze cycles, in which computational predictions are tested experimentally and the results feed back into the next design round, remain the practical engine of this approach, with the AI systems accelerating the exploration of chemical space rather than replacing experimental validation.

The broader productivity claims surrounding this nomination merit contextualization. Thirty-five preclinical candidates since 2021, with eleven in the first ten months of 2026 alone, represents an unusually high output for a single company, and the company attributes this to its AI-driven platform. Independent verification of such productivity claims is difficult, since the quality, novelty, and eventual clinical success of candidates matter more than raw counts. The company’s own pipeline provides a partial test case: Rentosertib, its most advanced internally discovered asset, has progressed to Phase III trials for idiopathic pulmonary fibrosis, which would represent a significant milestone for the field of AI-discovered drugs if it ultimately succeeds. The company’s reported financial performance, driven substantially by licensing and collaboration deals with established pharmaceutical partners, suggests that external parties are assigning real value to the platform’s output, though deal values reflect strategic assessments rather than validated clinical efficacy.

For patients with inflammatory bowel disease and rheumatoid arthritis, the therapeutic landscape already includes biologics targeting tumor necrosis factor, interleukin pathways, and integrins, as well as oral small molecules such as JAK inhibitors. However, a substantial fraction of patients fail to respond adequately to existing options or lose response over time, and safety concerns such as infection risk with immunosuppressive agents remain relevant across the class. A VAV1 degrader would represent a mechanistically distinct approach, intervening upstream of many of the cytokine pathways that current drugs target individually. Whether this translates into superior clinical outcomes, better safety, or advantages in specific patient subgroups will only be established through clinical trials, which have not yet begun for this candidate.

The path from preclinical candidate nomination to approved medicine remains long and uncertain. Typical development timelines span many years, with attrition rates high at every stage. For ISM9546, the immediate next steps would involve formal preclinical development activities, including formulation work, additional toxicology, and the preparation of an investigational new drug application or equivalent regulatory filing. The sub-nanomolar degradation potency and favorable preclinical safety profile provide a foundation, but human pharmacokinetics, immune function in diseased versus healthy populations, and long-term tolerability all remain to be demonstrated. The nomination of a preclinical candidate is best understood as the transition from discovery research to development-stage risk, not as the arrival of a therapy.

Within the larger arc of targeted protein degradation, the VAV1 program adds to a growing list of intracellular immune targets being pursued with degrader approaches, including kinases, transcriptional regulators, and signaling adaptors that were once dismissed as undruggable. The convergence of generative AI design tools with this modality suggests that the bottleneck in the field may be shifting from target identification and hit discovery toward clinical validation and the biology of degradation itself, including questions about tissue selectivity, degradation kinetics in humans, and the consequences of eliminating scaffold proteins that may have roles beyond their best-characterized functions. How these questions resolve will shape not only the fate of this particular candidate but the trajectory of the molecular glue field as a whole.

Subject of Research: AI-driven molecular glue innovation: Insilico nominates ISM9546, an oral, peripheral-restricted VAV1 degrader as PCC for autoimmune diseases

Article Title: AI-driven molecular glue innovation: Insilico nominates ISM9546, an oral, peripheral-restricted VAV1 degrader as PCC for autoimmune diseases

Article References: AI-driven molecular glue innovation: Insilico nominates ISM9546, an oral, peripheral-restricted VAV1 degrader as PCC for autoimmune diseases. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: AI-driven, molecular, glue, innovation, Insilico, nominates, ISM9546, oral, peripheral-restricted, VAV1, degrader, autoimmune

The research, conducted by Fuli Qiao and Mehrdad Mahdavi of the Department of Computer Science and Engineering at Penn State, builds on Low-Rank Adaptation, or LoRA, one of the most widely used parameter-efficient fine-tuning techniques in modern artificial intelligence. Instead of retraining the billions of weights inside a large language model, LoRA freezes the original weights and injects small pairs of low-rank matrices into each layer. Only these small matrices are updated during training, which slashes memory and computation costs by orders of magnitude. The catch, the authors note, is that conventional LoRA fixes the rank to the same value across every layer and every task, a one-size-fits-all choice that leaves a crucial question unexplored: how much capacity does each layer actually need for each new task?

Qiao and Mahdavi’s answer is a method that treats rank as a dynamic resource rather than a fixed hyperparameter. Their approach, which the team calls CL-Rank, relies on a subspace similarity metric to measure how orthogonal — that is, how non-overlapping — the low-rank subspaces occupied by different tasks are within a given layer. When a new task arrives, the algorithm evaluates the geometry of the learned subspaces and adaptively increases the layerwise rank for the new task where overlap threatens to interfere with previously learned representations. By steering new knowledge into directions of parameter space that are roughly orthogonal to old knowledge, the method minimizes interference while still giving each new task enough expressive capacity to generalize well.

The underlying intuition echoes a classical idea from neuroscience and machine learning known as the stability–plasticity dilemma. A learning system must be plastic enough to absorb new information but stable enough to retain old skills. Since the pioneering work on catastrophic interference in connectionist networks in the late 1980s and the advent of modern continual learning benchmarks, researchers have tried an arsenal of remedies: replaying stored examples, regularizing important weights, growing new network branches, or constraining gradient updates to null spaces of prior tasks. Many of these approaches are either memory-hungry, brittle to task order, or impractical for models with billions of parameters. What distinguishes the new work is that it tackles the problem entirely within the cheap, low-rank adapter framework, requiring no access to old data and no freezing of the base model.

Technically, the method tracks the principal subspaces spanned by the low-rank updates in each layer and computes a similarity score between the subspace of an incoming task and those of earlier tasks. If the new task’s gradient updates are projected into directions that overlap heavily with prior subspaces, the rank of the adapter in that layer is expanded, giving the optimizer room to find solutions that spare old representations. Layers whose subspaces remain naturally disjoint need no expansion, so the parameter budget is spent only where it matters. The paper’s supplementary analyses reveal that the learned rank distributions differ markedly across layers and modules — for example, the variation in ranks among the value projection modules of encoder layers is larger than among query projections, while encoder layers as a whole show more consistent rank allocations than decoder layers. This suggests that different parts of a transformer genuinely serve distinct roles, and that a uniform rank silently wastes capacity in some layers while starving others.

To test the idea, the researchers ran experiments on T5 and Llama-2-7b language models, using GPU servers with DeepSpeed for efficient training. For T5 experiments they employed four NVIDIA A6000 GPUs with a learning rate of 1e-3 and a batch size of 32, while the larger Llama-2-7b runs used four NVIDIA A100 GPUs at a learning rate of 1e-4. The evaluation spanned standard natural language processing continual learning benchmarks built from 15 datasets, including the classic text classification benchmark of Zhang and colleagues, the GLUE and SuperGLUE suites, and the IMDB movie review corpus, with six different task sequence orders to control for ordering effects. The team also pushed into harder territory with challenging mathematical reasoning benchmarks, GSM8K and MATH, each split into sequential tasks and tested in both orders to probe how task sequencing shapes forgetting.

The results show that the adaptive-rank method matches or beats strong baselines on three fronts at once: it mitigates forgetting of earlier tasks, improves accuracy on the tasks being learned, and preserves strong generalization to unseen data — all in a memory-efficient manner. On the math benchmarks, where sequential training on GSM8K and MATH typically causes steep declines in earlier-task accuracy, the method with orthogonal projection consistently outperformed the O-LoRA baseline in final average testing accuracy, forgetting less on the first task while generalizing better on the second. Notably, the study observes that fully fine-tuning a T5-large model on MATH yields only about 3.0 percent accuracy, and about 4.2 percent on GSM8K, consistent with prior work — underscoring how difficult these mathematical tasks are and how much room remains for continual learning approaches that squeeze more capability out of small adapters.

Why should this matter beyond the benchmark tables? Continual learning is arguably the missing ingredient between today’s static AI assistants and the adaptive, lifelong learning systems that both researchers and companies envision. Every time a deployed model needs a new skill — a new language, a new domain, a new tool — retraining it from scratch or even fully fine-tuning it is prohibitively expensive. Adapter-based continual learning offers a path to modular, on-demand skill acquisition, but only if adding skills does not corrupt existing ones. By showing that the humble LoRA rank, tuned per layer and per task, is a powerful lever for balancing stability and plasticity, the Penn State work gives practitioners a new, inexpensive knob to turn. Because the method requires no stored examples of past tasks, it also sidesteps the privacy and storage concerns that plague replay-based approaches.

The findings also carry a broader scientific message: the internal geography of large models is far from uniform, and adaptation schemes that respect that heterogeneity can outperform uniform ones. The observation that encoder and decoder layers, and even the query and value modules within them, prefer different rank allocations suggests that future parameter-efficient methods may benefit from treating adapters as structured, layer-aware components rather than interchangeable plug-ins. The authors have released their code publicly, and all datasets used in the experiments are openly available, making the results straightforward for other groups to verify and extend. The work was partially supported by the National Science Foundation under award EFMA-2318101, with experiments designed and conducted by Fuli Qiao under the supervision of Mehrdad Mahdavi.

As large language models continue to spread through science, medicine, and industry, the question of how to teach them new things without breaking old ones is shifting from academic curiosity to engineering necessity. This study does not solve continual learning outright — no single method does — but it demonstrates that a careful, geometry-aware allocation of low-rank capacity can deliver competitive anti-forgetting performance at a fraction of the cost of full retraining. In a field where progress often comes from scaling up, it is a reminder that sometimes the most powerful improvements come from scaling the right things in the right places, one layer at a time.

Subject of Research: Adaptive layerwise LoRA ranks for continual learning in large language models

Article Title: Learning Without Forgetting for Continual Learning in LLMs Through Adaptive LoRA Ranks

Article References: Qiao, F., & Mahdavi, M. (2026). Learning Without Forgetting for Continual Learning in LLMs Through Adaptive LoRA Ranks. Machine Learning, 115(9), Article 212. https://doi.org/10.1007/s10994-026-07145-4

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07145-4

Keywords: continual learning, large language models, LoRA, catastrophic forgetting, parameter-efficient fine-tuning, adaptive ranks, subspace orthogonality, stability-plasticity, GLUE, GSM8K, MATH, T5