PharmAgents2 hands drug discovery to agents: it proposes a new LpxH allosteric site and the compounds pass wet-lab checks
Synopsis
PharmAgents2 moves drug discovery from single-task models to an end-to-end autonomous workflow: it distills transferable decision knowledge from published literature and uses parallel agents to explore target, hypothesis and strategy spaces. On the antibacterial target LpxH it autonomously hypothesized a previously unreported allosteric site and screened against it, yielding AIPher-L1 with experimentally confirmed inhibitory activity and binding; JAK1 work and secondary-target exploration for the known molecule AIPher-X were supported by cell assays and SPR experiments respectively.
Interpretation
Drug discovery is no longer carried only by individual prediction models; an agent system advances a research objective all the way to candidate molecules ready for experimental testing. Earlier AI tools each strengthened a single step such as protein structure prediction or molecular design, while organizing and advancing the overall process still relied on human judgment. On 121 questions across nine decision scenarios in PharmAgents-Bench v0.1, PharmAgents2 outperformed the evaluated baselines in most categories, and it achieved the highest process scores on all three end-to-end tasks.
The system turns decision knowledge into a retrievable skill library and lets parallel agents share one research context to widen exploration. Paper2Skill synthesizes methodologically related papers into skills that record the links between research objectives, methodological choices and the conditions under which those choices apply, then merges them with expert-authored skills. Across ten topics, directly synthesized skills lowered conditional perplexity on both construction and held-out papers, with further reductions after perplexity-guided completion; in paired evaluations, Direct-guided plans were preferred over no-skill plans in 87.9% of comparisons and Refined-guided plans over Direct-guided plans in 65.0%.
In real research scenarios, the candidates the system delivered were then tested in human-executed wet-lab experiments. Researchers purchased and tested 63 candidates in the LpxH case and 34 in the JAK1 case. AIPher-L1 inhibited KpLpxH with an IC50 of 8.38 μM and showed binding affinities of 1.40 μM by SPR and 0.87 μM by MST; at 10 μM, AIPher-J1 showed inhibitory activity comparable to the positive controls in two cytokine-signaling pathways.
Perspective
The demonstrated scope is three real research scenarios: LpxH inhibitor discovery, selective JAK1 inhibitor discovery, and secondary-target exploration starting from the known bioactive molecule AIPher-X; the benchmark covers 121 decision questions and three end-to-end tasks. Readers should treat this as evidence for a prototype system, not a guarantee of a reproducible workflow for arbitrary targets.
In the LpxH case, binding-site identification is ongoing and only preliminary evidence supports the proposed allosteric mechanism; in the JAK1 case, experiments to validate selectivity and the proposed binding modes are underway; the SPR binding result for AIPher-X and Target X appears as a placeholder in the source, with no specific value reported. The system-level account does not isolate the contribution of any particular skill or computational tool.