Intelligent design forgene therapiesdrug deliverylipid nanoparticlesionizable lipidsmRNA deliverygene therapiesgene therapies

Ionizable lipids are the critical component of lipid nanoparticles for in vivo mRNA delivery, but discovery is bottlenecked by combinatorial library enumeration. Existing ML approaches rank pre-enumerated libraries rather than generating novel structures.

We introduce synthesis-constrained discrete diffusion — the first deep generative model for ionizable lipids that embeds reaction constraints directly into the diffusion process. Three components — scaffold conditioning, region-aware noise, and property conditioning via FiLM with classifier-free guidance — yield 99% chemical validity, 100% scaffold integrity, 62% novelty, and best predicted transfection potency 2× the training mean (in silico).

The framework inverts the dominant generate-then-filter paradigm: every molecule we produce is synthesizable by construction. Top candidates are being synthesized for experimental validation at CHOP.

Clostridioides difficile infection is an urgent public health threat with limited preventative options. We developed an mRNA-lipid nanoparticle vaccine targeting C. difficile toxins and virulence factors.

The multivalent vaccine elicited robust, long-lived systemic and mucosal antigen-specific humoral and cellular immune responses across animal models, independent of changes to the intestinal microbiota.

Vaccination protected mice from lethal infection in both primary and recurrent models, and including non-toxin cellular and spore antigens improved decolonization of toxigenic C. difficile from the gastrointestinal tract.

Adjuvants are critical for improving the quality and magnitude of adaptive immune responses to vaccination. LNP-encapsulated nucleoside-modified mRNA vaccines showed great efficacy against SARS-CoV-2, but the mechanism of action of the platform was not well characterized.

Using influenza virus and SARS-CoV-2 mRNA and protein subunit vaccines, we showed that the LNP formulation has intrinsic adjuvant activity promoting strong T follicular helper cell, germinal center B cell, long-lived plasma cell, and memory B cell responses associated with durable, protective antibodies in mice — outperforming the widely used MF59-like adjuvant AddaVax.

The adjuvant activity relies on the ionizable lipid component and on IL-6 induction, identifying LNPs as a versatile adjuvant that enhances both traditional and next-generation vaccine platforms.

Lipid nanoparticles are used to deliver siRNA and COVID-19 mRNA vaccines. The main factor known to determine their delivery efficiency is the pKa of the LNP containing an ionizable lipid. We report a method that can predict LNP pKa from the structure of the ionizable lipid, using theoretical, NMR, fluorescent-dye binding, and electrophoretic mobility measurements of protonation in both the lipid and the formulated particle.

The pKa of the ionizable lipid was 2–3 units higher than that of the LNP, primarily due to proton solvation energy differences between the particle and the aqueous medium.

We used this to explain a wide range of delivery efficiencies in vitro and in vivo for intramuscular and intravascular administration, and found that more negatively charged LNPs show higher off-target systemic expression in the liver — an effect that can be minimized through lipid and particle design.

Therapeutic peptides combine the advantages of small molecules and antibodies, offering target flexibility and low immunogenicity, yet successful translation requires evaluating multiple developability properties beyond binding alone. As chemically modified peptides become more common in drug design, no unified platform supports systematic property assessment across both canonical sequences and SMILES-based representations.

Leveraging the generalizability of large foundation models trained on protein and chemical data, we introduce PeptiVerse, a universal therapeutic peptide property prediction platform. It accepts either amino acid sequences or chemically modified peptide SMILES, delivers state-of-the-art performance across diverse prediction tasks, and provides both a web interface and an open-source implementation.

By unifying property prediction across representations, PeptiVerse supports early-stage peptide therapeutic development campaigns and property-aware generative design workflows.

Protein function is often controlled by ligands that bias the direction of state transitions — agonists and antagonists — rather than stabilizing a single conformation. This matters especially for clinically relevant GPCRs, where therapeutic efficacy depends on functional directionality. Structure-based design optimizes binding to static conformations and cannot represent non-reversible, directional effects.

We introduce TD3B, a sequence-based generative framework that designs binders with specified agonist or antagonist behavior via a directional transition control objective. It combines a target-aware Direction Oracle, a soft binding-affinity gate, and amortized fine-tuning of a pre-trained discrete diffusion model.

This enables targeted agonist and antagonist generation decoupled from binding affinity — unattainable by equilibrium-based or inference-only guidance baselines.

The computational design of protein-based binders offers a route to ‘undruggable’ targets, but effective binder design usually relies on stable three-dimensional structures or structure-influenced latent spaces.

We introduce PepMLM, a target sequence-conditioned designer of de novo linear peptide binders. A masking strategy that positions cognate peptide sequences at the C terminus of target protein sequences finetunes the ESM-2 protein language model to fully reconstruct the binder region, achieving low perplexities matching or improving upon validated peptide–protein pairs.

PepMLM-derived peptides show sequence-specific binding to cancer and reproductive targets including NCAM1 and AMHR2, and enable targeted degradation of proteins across contexts from Huntington's disease to live viral infections — without requiring structural input.

Cross-pair amortization of alchemical free-energy estimation requires a transferable within-pair conditioning coordinate. We identify the cumulative Fisher arc length ℓ_alch — the path length along the alchemical path under the Fisher metric — as the unique (up to affine rescaling) parameterization-invariant coordinate whose equal increments correspond to equal Kullback–Leibler distances between adjacent intermediates. The canonical coupling parameter λ fails this property because the Fisher information varies across pairs.

This information-geometric framing makes a falsifiable prediction: the advantage of ℓ_alch over raw λ should track the quality of the underlying single-topology atom mapping. Substituting ℓ_alch for λ at training time in an otherwise identical conditional flow-matching estimator confirms this — improvements on real-chemistry corpora concentrate in the cleanly-mapped operating regime, with synthetic and Lennard-Jones diagnostics verifying the mechanism in controlled settings.

Which parameters of an adaptive optimizer should retain exact second-moment state, and which can be compressed? Memory-efficient methods such as Adafactor, GaLore, and 8-bit Adam commit at design time. On transformer LM training, AdamW second moments concentrate on a small fraction of coordinates, with one dominant tensor typically carrying most of the adaptive mass.

DOSA treats state allocation as an online statistical decision, promoting the candidate update rule with the largest lower confidence bound on a smoothness-model predicted-descent score. The main result is an LCB-certificate separation theorem: under heavy-tail concentration, any uniform-width sketch meeting the same identification certificate as per-tensor adaptive widths requires asymptotically more memory.

A swept blend improves validation perplexity over a factored-Adafactor baseline across model scales and datasets, and the closed-form long-horizon surrogate rule recovers about half the swept gain without tuning on DistilGPT2/WikiText-2.