Genesis Molecular AI (formerly known as Genesis Therapeutics) has presented results for its PEARL model — Place Every Atom at the Right Location — which the company says advances protein–ligand co‑folding by modeling protein flexibility and small protein adjustments that improve ligand fit. Founders mentioned in the discussion include Evan Feinberg and CTO Sergey Edunov. According to Genesis, PEARL can predict how a ligand and a protein move together in ways that static approaches typically miss.
Why this matters
Accurate protein–ligand structures are critical in small‑molecule drug discovery because correct binding poses determine interactions and efficacy. The community often treats 2 Å root‑mean‑square deviation (RMSD) as the cutoff for a "good pose," but Genesis and several researchers argue this threshold is too permissive. Evan Feinberg contends that roughly 1 Å RMSD is needed to ensure the molecular core and key interactions are placed reliably.
Benchmarks under scrutiny
Genesis criticizes the field’s reliance on the 2 Å benchmark, arguing it can hide important errors because some interactions — for example hydrogen bonds — only tolerate about a 0.6 Å range. They attribute part of the community’s fixation on looser thresholds to a historical separation between method developers and practical users in drug discovery.
OpenBind results on EV‑A71 (802 complexes)
A notable claim by Genesis is PEARL’s performance on the OpenBind benchmark, specifically on a dataset of 802 previously unseen co‑complexes with the EV‑A71 target. EV‑A71 presents a difficult case because ligands induce conformational changes in the protein (“induced fit”), closing off the ligand’s entry path — a scenario that is hard for traditional docking to handle. According to Genesis, PEARL modeled these induced‑fit movements without long molecular dynamics simulations and outperformed public models across multiple metrics. The company also emphasizes that PEARL was evaluated without fine‑tuning on that target and that template PDB data for EV‑A71 became available only after PEARL’s training cutoff.
Agentic discovery and SAPPHIRE
Genesis frames their progress not only as increased model accuracy but as passing a threshold that enables "agentic" drug discovery loops. Their internal system, codenamed SAPPHIRE, is described as capable of iterating like a chemist: reasoning about poses, forming hypotheses, consulting literature, using internal tools, and proposing next‑round candidates. Combined with automated lab partnerships such as the one with Incyte, they envision continuous design–test cycles running around the clock.
Technical context and research trends
The discussion noted that while transformer‑based architectures have dominated modern LLM research, diffusion methods have emerged as a useful primitive for 3D structure problems. Sergey Edunov, who joined Genesis from Meta (where he led Llama 2 training and Llama 3 pretraining), highlighted the novel diffusion architecture work Genesis is pursuing for 3D prediction.
Where co‑folding may be headed
The article’s author, who has tracked ML methods for protein–ligand interactions for years, finds PEARL’s results impressive and sees similar progress at other closed teams such as Isomorphic and Deep Origin. These developments suggest the field may approach a point where protein–ligand binding prediction is practically solved for certain tasks. The author also stresses that academia will need to adapt benchmark standards and foster tighter integration of ML, large‑scale compute, and real‑world discovery workflows to broaden impact.
Conclusion
Genesis Molecular AI presents PEARL as a model that can handle induced‑fit protein movements, exceed public models on a challenging OpenBind EV‑A71 benchmark of 802 complexes, and motivate a reassessment of common RMSD thresholds (advocating ~1 Å over 2 Å). Paired with agentic workflows like SAPPHIRE and industrial lab partnerships, the company views these advances as enabling more continuous, automated small‑molecule discovery pipelines.
(Note: Genesis has released a PEARL technical report that discusses limitations of conventional benchmarks in more detail.)
