Generative AI punches above its data weight — until the data runs out

Doing more with less

Three preprints posted this week on ChemRxiv circle the same question from different sides: how far can generative AI get when labeled experimental data is scarce? Two of them report encouraging answers — a fine-tuned foundation-model workflow that finds nanoparticle-forming excipients, and a protein–ligand cofolding model that lands nanomolar kinase inhibitors prospectively. The third supplies the necessary counterweight: a benchmark showing that for ionic-conductivity prediction in polymer electrolytes, machine learning still performs about as well as its dataset lets it. Together they sketch a practical map of where data efficiency is real and where it is not. m:…0348 m:…4234 m:…9656

The pattern worth watching is where the reported wins come from. In the two success stories, AI is asked to propose and prioritize — candidates are then made and tested — whereas the benchmark evaluates pure prediction from composition alone. All three studies are preprints, so the usual caveats about peer review apply, and none of the findings has clinical or industrial validation. m:…0348 m:…9656

Excipients from a foundation model

The nanoparticle result introduces a generative AI framework for molecular and nanoparticle design in low-data regimes. The approach combines parameter-efficient fine-tuning of foundation models, using complementary positive and negative property prompts, with molecular graph-kernel Gaussian process regression that ranks generated candidates for synthesizability and drug-likeness. The authors report that fine-tuning with those paired prompts is what makes generative design workable when labeled examples are few. m:…0348

The test case was excipients for the anticancer drug sorafenib. The workflow identified lithocholic acid and sorafenib metabolite M6 as novel nanoparticle-forming excipients, and the resulting sorafenib nanoparticles are reported to be stable with increased in vitro potency. For drug-delivery researchers, the notable claim is the regime: useful discovery outputs from a low-data setting, which is where most formulation problems actually live. m:…0348

A prospective win for cofolding

The Boltz-2 result is a prospective validation story, which is the harder kind to tell in AI-for-chemistry. Rather than benchmarking retrospectively, the study coupled the Boltz-2 cofolding affinity model with a custom-synthesized library of stereochemically complex tetrahydroquinolines — accessed through a single-step multicomponent synthetic approach that deliberately bypasses the limits of standard commercial screening catalogs — and reports that the model successfully guided both high-throughput virtual screening and hit optimization. m:…4234

The headline result: nanomolar inhibitors of cyclin-dependent kinase 2, CDK2, discovered in that custom chemical space. For medicinal chemists, the implication is that cofolding affinity models can navigate novel, stereochemically rich regions of chemistry rather than only re-ranking familiar scaffolds. The usual limits apply: the digest notes no open-source code or dataset for the custom library, and these are preclinical, preprint-stage findings. m:…4234

Where data still rules

The conductivity benchmark studies a different problem — predicting ionic conductivity in polymer electrolytes for next-generation, low-cost batteries — and reaches a sobering conclusion about what ML can do with thin datasets. On the established OBELiX lithium dataset, a standard Random Forest/XGBoost pipeline achieves an R² of 0.79. But on a newly constructed dataset of 135 multivalent (magnesium, zinc, sodium) electrolyte systems, the same class of models manages only an R² of 0.47 under leave-one-out cross-validation. m:…9656

The authors' reading of the gap is what makes this worth citing alongside the two success stories: they attribute the performance difference primarily to data scarcity and heterogeneity of literature measurements, rather than to model architecture. That frames the field's real bottleneck as dataset construction — larger, standardized multivalent-electrolyte datasets — rather than better algorithms. For R&D strategy in non-lithium batteries, the message is to invest in measurement infrastructure before model sophistication. m:…9656