New model · Massachusetts Institute of Technology
ShEPhERD-2
3D generative model that designs drug-like molecules from explicit shape, electrostatic and pharmacophore profiles
· 5 claims checked against 1 source
What it is
ShEPhERD-2 is a 3D generative model, developed at MIT, that generates molecular structures conditioned on explicit interaction profiles. An interaction profile combines a target's shape, electrostatics, and directional pharmacophores into a single design specification.
How it works
Small-molecule binding is governed by how a molecule's shape, charge distribution, and pharmacophore geometry line up with a target site in three dimensions. ShEPhERD-2 treats that combination, the interaction profile, as the design input itself, rather than optimizing a docking score or a 2D fingerprint. Given a profile, the model generates 3D molecular structures meant to reproduce it.
What is new
The model unifies shape, electrostatics, and directional pharmacophores into one conditioning signal that the authors describe as a sufficient and transferable design specification. Compared with its predecessor, ShEPhERD, the new version generates low-strain, drug-like molecules more efficiently and with greater fidelity to the target interaction profile.
- Single, transferable interaction profile as the design spec
- Greater fidelity to target profiles than predecessor ShEPhERD
- Pharmacophore prioritization for weighting which interactions matter most
- Positioned by the authors as a general framework, not a task-specific tool
For medicinal and computational chemists, this offers a way to specify what a molecule should do geometrically and electrostatically, then let the model search for structures that satisfy it, an approach closer to how chemists reason about binding than optimizing an abstract score.
Limits
This is a preprint result. Independent benchmarking against real hit-finding campaigns will show how well profile fidelity translates into actual binders.
Availability
ShEPhERD-2 is available as a bioRxiv preprint titled 'Interaction Profiles as a Universal Language for Generative Molecular Design with ShEPhERD-2.' Code and weight availability are not stated in the preprint.
The evidence
Each statement above was written with the sentence from the source that supports it, and a program (not a model) checked that every quote and every figure appears in the pages it read. Here they are.
- 01
ShEPhERD-2 is a 3D generative model that generates molecular structures conditioned on explicit interaction profiles
“We introduce ShEPhERD-2, a 3D generative model that generates molecular structures conditioned on explicit interaction profiles” S1
- 02
An interaction profile comprises shape, electrostatics, and directional pharmacophores
“an interaction profile, comprising shape, electrostatics, and directional pharmacophores, is a sufficient and transferable design specification” S1
- 03
ShEPhERD-2 generates low-strain, drug-like molecules more efficiently and with greater fidelity to target interaction profiles than its predecessor
“ShEPhERD-2 generates low-strain, drug-like molecules more efficiently and with greater fidelity to target interaction profiles than its predecessor” S1
- 04
The model enables precise control over molecular design through pharmacophore prioritization
“enabling precise control over molecular design through pharmacophore prioritizatio” S1
- 05
The work is presented as a bioRxiv preprint titled 'Interaction Profiles as a Universal Language for Generative Molecular Design with ShEPhERD-2'
“Interaction Profiles as a Universal Language for Generative Molecular Design with ShEPhERD-2” S1
