The Therapeutic Dilemma and a Computational Solution
The remarkable therapeutic potential of psychedelics like psilocybin appears intrinsically linked to their ability to induce profound, often challenging, altered states of consciousness. This creates a significant barrier: the need for extensive psychological support, medical supervision, and the exclusion of many patients due to cardiovascular risks or psychiatric vulnerabilities. What if we could separate the neuroplastic, healing effects from the intense psychedelic experience? The Institute's Computational Psychopharmacology Unit is tackling this challenge head-on by using artificial intelligence to design a new class of drugs: non-hallucinogenic psychoplastogens. These compounds aim to trigger the same growth-promoting intracellular pathways as classic psychedelics, but without strongly activating the subjective, perceptual components.
Deconstructing the Signal: From Receptor to Experience
Our approach is based on a key insight from receptor pharmacology: different drugs binding to the same receptor (e.g., the serotonin 2A receptor) can activate different downstream signaling pathways—a phenomenon known as 'functional selectivity' or 'biased agonism.' A compound might preferentially activate the pathway leading to neuroplasticity (e.g., the PLC-IP3 pathway that elevates intracellular calcium and triggers BDNF release) while minimally activating the pathway linked to hallucinations (potentially the arrestin recruitment pathway linked to specific patterns of cortical desynchronization). Using high-throughput screening and molecular dynamics simulations, we have mapped the precise 'fingerprints' of known psychedelics at the 5-HT2A receptor—how they twist the receptor, which amino acids they interact with, and which intracellular proteins they recruit.
Generative AI and In Silico Drug Design
We feed this data into a generative adversarial network (GAN) trained on millions of known chemical structures and their biological activity data. The AI is tasked with designing novel molecular structures that optimally fit the 'neuroplasticity bias' fingerprint at the 5-HT2A receptor. It also optimizes for other crucial properties: blood-brain barrier penetration, metabolic stability, and a clean off-target profile. The AI generates thousands of virtual candidate molecules per hour. These are then screened through predictive models for their hallucinogenic potential, trained on behavioral data from animal models (like head-twitch response in mice) correlated with human psychedelic effects.
Promising virtual hits are then synthesized by our robotics-enabled chemistry lab. The synthesized compounds undergo rapid testing in our cellular assays measuring neurite outgrowth and in electrophysiology setups. The most successful, dubbed 'neuroplastins,' are then moved into animal behavioral studies. Our lead candidate, NP-101, shows robust increases in dendritic spine density and synaptic proteins in rodent prefrontal cortex, reverses chronic stress-induced behavioral deficits in models of depression, and yet produces no head-twitch response or disruptions in sensory-motor gating—proxies for the hallucinogenic experience.
This AI-driven pipeline represents a paradigm shift. Instead of randomly screening natural products or making incremental changes to existing psychedelics, we are rationally engineering molecules from first principles to achieve a specific therapeutic goal. The development of non-hallucinogenic psychoplastogens could democratize access to rapid-acting, plasticity-promoting treatments, allowing them to be administered in conventional clinical settings and potentially for a wider range of neurological conditions, from depression and PTSD to stroke recovery and cognitive decline. The future of psychotropic medicine may be designed not in the jungle, but in the silicon neural networks of our computers.