About Neuralocity

Neuralocity LLC builds a molecule data platform for early drug discovery. We package computationally generated and curated molecules into a searchable, scored catalog (pre-screened and ranked by target-class fit, safety risk, drug-likeness, novelty, and synthesis readiness), delivered as licensed datasets with provenance and export workflows.

What we do

Discovery teams spend too much time narrowing enormous chemical spaces into practical screening lists. Neuralocity compresses that work: customers browse a catalog of novel molecules with multi-dimensional annotations, build shortlists, and export licensed data for downstream computational and laboratory validation.

Our scoring emphasizes target-class relevance across 10 biological target classes, expressed as within-catalog ranks and percentiles, not surrogate assay potency claims. That framing keeps the product scientifically defensible and suitable for research-use licensing.

Our technology

At the core of the catalog is a generative AI model we built in-house: a large language model purpose-built and trained to generate novel molecules with drug-like properties.

Generation is only the first step, and we take a deliberately conservative approach to what actually makes it into the catalog. Every candidate is filtered before it is ever scored: molecules outside typical drug-like property ranges (molecular weight, lipophilicity, hydrogen-bond donors and acceptors) are discarded, molecules that look impractical to synthesize are discarded, and molecules flagged by medicinal-chemistry structural-alert rules for reactive or unstable functional groups and implausible ring systems are discarded. The goal isn’t just to generate molecules; it is to generate molecules a chemist could actually pick up and work with.

From there, every molecule is scored by a suite of trained machine learning models, not a single generic algorithm. A dedicated model is trained for each target class and mechanism (antagonist or agonist), each toxicity endpoint, and each drug interaction pathway, so every score reflects a model purpose-built for that specific biological question.

Target-class relevance models estimate activity likelihood across 10 biological target classes:

• Aggregation Target
• Apoptosis Regulator
• Enzyme
• Epigenetic
• GPCR
• Ion Channel
• Kinase
• Nuclear Receptor
• Protease
• Transporter

Toxicity classifiers, each its own trained model, screen for carcinogenicity, hepatotoxicity, cardiotoxicity, and other endpoints; drug-drug interaction models flag possible metabolic conflicts; and drug-likeness and synthetic accessibility scores round out the picture. Every model is evaluated against a held-out test set before it ships, and results are expressed as ranks, percentiles, and documented tiers rather than raw model probabilities.

A separate trained model predicts aqueous solubility for every catalog molecule: a graph neural network (AttentiveFP) learns the relationship between a molecule’s structure and its solubility directly from experimentally reported data, then predicts log S (a standard log-scale measure of solubility in water) for each candidate.

Beyond the catalog’s pre-computed, target-class-wide scores, molecule detail pages also support on-demand structure and binding prediction against a specific target a customer supplies. This runs Boltz-2, an open-source biomolecular structure and binding-affinity model, to predict the 3D complex between a candidate molecule and that target directly from sequence, no known crystal structure required. A customer can then optionally refine that result with AutoDock Vina, a well-established physics-based docking engine, which independently searches the same predicted pocket using an entirely different method. Because the two approaches work in fundamentally different ways, agreement between them is a meaningfully stronger signal than either result alone, and disagreement is itself informative rather than being hidden.

Who we serve

• Computational chemists at early-stage biotechs triaging compound lists
• Medicinal chemists evaluating shortlists for synthesis and ordering
• Discovery program leads who need defensible starting pools for partnering discussions
• CRO and academic teams that need focused molecule sets without building a full screening pipeline

How we are different

• In-house generative AI: a purpose-built LLM we designed and built ourselves to generate molecules with drug-like properties, not aggregation of public catalogs alone
• Filtered for practical use, not just generated: every candidate is screened for drug-like properties, synthesizability, and stability before it ever reaches the catalog, so what you browse is meant to be chemistry a chemist could actually work with
• Target-class intelligence with publishable methodology and provenance per catalog release
• Multi-dimensional safety overlay applied at catalog scale
• Cart-based purchasing as the commercial unit, with audit logging and entitlement controls

Contact

Questions about Neuralocity, licensing, or pilots may be sent to contact@neuralocity.com.