Miroslav Lzicar

Synthesis-on-demand libraries expose vast chemical spaces through reaction templates and compatible synthons, so the key question is which candidates to evaluate under a small oracle budget. SynthonBench introduces reaction-native, budgeted search tasks where optimizers propose valid reaction/synthon tuples and are scored with top-k recall, enrichment, regret, and Pareto metrics. Accepted to the ICML 2026 GenBio Workshop.

Make-on-demand design spaces are increasingly defined by reactions and synthons rather than enumerated molecules, which makes search, audit, and visualization a grounding problem. This work presents SynthonOR for query-conditioned factorized retrieval and SynthonMap for route-aware sampling, embedding, and product-mass cartography. Accepted to the ICML 2026 GenBio Workshop.

Neural thickets suggest that many task experts may lie dense around pretrained weights, but density does not mean those experts are easy to discover. This audit tests whether nearby experts can be found through tiny learned low-rank routes under deployable selection and matched total query budgets, producing a mixed but useful probe of selector-sensitive neighborhoods. Accepted to the ICML 2026 Weight-Space Symmetries Workshop.

A lightweight Python package for reaction-grounded synthon retrieval that stays intentionally simple: generic RDKit fingerprints combined with a binary OR.

A compact synthon-conditioned transformer for navigating makeable chemical space, grounded in vendor enumerations rather than hallucinated SMILES.

Introducing a cellular-automata-native benchmark, leaderboard, and dataset suite that scores reasoning systems on metadata-rich ARC-AGI style episodes.

Ultra-lean BitBIRCH rewrite that clusters hundred-million scale molecular libraries with on-the-fly compression, optional C++ accelerators, and a batteries-included CLI.

A novel approach to virtual screening using 3D shape and electrostatics in a vector space.

A CHEESE case study where 3D electrostatic similarity prioritized NS1, a tetrazolic neurosteroid candidate, before synthesis. NS1 later showed strong activity and advanced to in vivo validation in an osteoarthritis pain model.

Our team at Deep MedChem won the DYRK1B challenge at 9ADD 2026 with a CHEESE ShapeSim + LightGBM workflow. This short note explains what mattered and includes the final presentation.

A quick note on adapting ISONET’s Gumbel–Sinkhorn edge alignment to molecular substructure search, with links to the source notebook and paper.

There are two distinct technologies to search in databases of molecules. This post explains both with intuitive LEGO illustrations and presents a practical guide on when to use each.

A recent paper claims an established virtual screening benchmark to be 'fundamentally compromised' due to 'egregious data leakage, rampant duplication, and pervasive analog redundancy', rendering it 'unfit for its intended purpose' and 'calling into question previous results based on its use'. The catch: the authors canonicalized to non-isomeric SMILES, effectively collapsing stereoisomers into duplicates.

This is my billion-scale vector engine for 3D shape & electrostatic molecular similarity. It turns ligand-based screening into nearest-neighbour lookups—so you can *search, rank, and visualise* huge chemical spaces in seconds.

Fingerprints have long been the standard for representing molecules, but neural network embeddings are opening doors to new possibilities. This post explores where these AI-driven representations excel, from better capturing molecular similarity with metrics to addressing modalities that traditional fingerprints often miss.

I tried, and failed, to get ViTs to work on the ARC challenge.