A clinical-stage biotech focused on neurological disorders used Systematic Search and Multi-Perspective Analyst to break out of local optima in molecular design — generating 48 candidates in its first search round, pruning them to 3 viable finalists, and compressing lead optimization from 18 months to 12 while cutting R&D costs by 40%.
Overview
Helion Therapeutics, a 350-person clinical-stage biotech developing treatments for treatment-resistant neurological disorders, was stuck. Its lead optimization program for a novel NMDA receptor modulator had been running for 11 months with diminishing returns — the medicinal chemistry team kept exploring modifications within familiar structural neighborhoods, burning through budget on analogs that offered marginal improvements. By deploying a Systematic Search (Tree of Thoughts architecture) paired with a Multi-Perspective Analyst (Ensemble architecture), Helion explored modification pathways that no human chemist on the team would have prioritized, identified three candidates with superior binding profiles, and compressed its remaining lead optimization timeline from 18 months to 12.
The Challenge
Helion's lead compound, HLN-4012, showed promising activity against treatment-resistant depression in early assays but had two persistent problems: moderate metabolic instability that limited its projected half-life, and a solubility profile that would complicate oral bioavailability. The medicinal chemistry team had spent 11 months running structure-activity relationship (SAR) studies — systematically modifying the compound's functional groups and testing the resulting analogs for improved properties.
The team was experienced. The lead chemist, Dr. Rana Mishra, had 18 years in CNS drug design. But experience can become a constraint. The team's modification strategy followed well-established medicinal chemistry heuristics: replace metabolically labile groups with known bioisosteres, introduce fluorine at predictable positions, adjust lipophilicity through standard substitution patterns. These are sound strategies. They are also the strategies that every trained medicinal chemist applies, which means the team was exploring the same structural neighborhoods that any comparable team would explore.
After 11 months, the program had evaluated 127 analogs. Sixteen showed measurable improvements in one property (usually metabolic stability) but at the cost of another (usually potency or selectivity). The team was trapped in a local optimum — a region of chemical space where every incremental modification produced trade-offs rather than net gains. Breaking out of a local optimum requires exploring modifications that appear counterintuitive from the perspective of established heuristics. That is precisely what experienced chemists are trained not to do.
The timeline pressure was severe. Helion's Series C funding carried a milestone clause: demonstrate a development candidate with acceptable ADMET properties within 24 months of the funding close. Eleven months were gone. The remaining 13 months had to cover lead optimization, candidate selection, and IND-enabling studies. At the current pace, the team projected needing 18 more months for lead optimization alone — five months past the funding deadline.
Dr. James Okafor, Helion's VP of Drug Discovery, framed the problem directly: "We were not making bad decisions. We were making the same good decisions that every medicinal chemist would make. The compound space we had not explored was the space where our intuition said 'that won't work.' We needed a way to search that space systematically without abandoning scientific rigor."
The Solution
Systematic Search (Tree of Thoughts Architecture)
The Systematic Search architecture approaches molecular optimization the way a chess engine approaches move selection — by exploring multiple modification pathways in parallel, evaluating each pathway several steps deep, and pruning branches that fail defined criteria before investing further resources.
Helion configured the system to generate modification trees rooted in HLN-4012's core scaffold. At each branching point, the system proposed 6 to 10 structural modifications — not randomly, but informed by property prediction models for metabolic stability, aqueous solubility, blood-brain barrier permeability, hERG liability, and NMDA receptor binding affinity. Critically, the system was not constrained by medicinal chemistry heuristics. It evaluated modifications that a human chemist would typically deprioritize: unusual ring replacements, non-obvious stereochemical inversions, scaffold hops that preserved pharmacophoric geometry while radically changing the molecular framework.
Each branch was evaluated three steps deep. A modification that degraded one property at step one but opened a pathway to improved properties at step two or three was retained for further exploration — something a traditional SAR campaign, which evaluates modifications one step at a time, would miss entirely. The system generated 48 candidate structures in its first complete search cycle, each with a predicted multi-property profile and a complete modification trail documenting the reasoning at every branching decision.
The modification trail proved unexpectedly valuable. Regulatory submissions for novel therapeutics require complete documentation of the design rationale. The Systematic Search architecture generated that documentation automatically — every branch explored, every branch pruned, every decision justified against quantitative property predictions.
Multi-Perspective Analyst (Ensemble Architecture)
Forty-eight candidates is a starting point, not an answer. Helion needed to reduce that set to a number that its wet-lab capacity could handle — ideally three to five compounds for synthesis and experimental validation. This is where the Multi-Perspective Analyst architecture delivered its value.
The Ensemble architecture evaluated each of the 48 candidates through five independent analytical lenses: a pharmacokinetics agent assessing ADMET predictions, a synthetic feasibility agent scoring the difficulty and cost of each synthesis route, a patent landscape agent checking for freedom-to-operate conflicts, an off-target liability agent predicting selectivity issues, and a clinical translatability agent evaluating how the compound's profile would map to real-world dosing regimens.
Each agent scored every candidate independently. The system then aggregated scores, identified consensus top performers, and — more importantly — flagged candidates where agents disagreed strongly. A compound that scored well on pharmacokinetics but poorly on synthetic feasibility might still be valuable if the synthesis challenge was solvable. A compound that scored well everywhere except off-target liability was likely a dead end regardless of its other properties.
The Multi-Perspective Analyst compressed what would normally be two to three months of cross-functional review meetings into a structured evaluation completed in nine days. The final output ranked all 48 candidates with composite scores, individual agent assessments, areas of consensus, areas of disagreement, and recommended next steps for each. Helion's team selected three candidates for synthesis.
The Results
Helion synthesized all three candidates within eight weeks. Two of the three confirmed the predicted property improvements in experimental assays — a hit rate that Dr. Mishra called "remarkable" given that typical SAR campaigns see confirmation rates of 20-30% for computationally predicted properties.
- 48 candidates generated in the first systematic search cycle, exploring structural modifications that the medicinal chemistry team had not considered in 11 months of manual SAR work.
- 3 viable finalists selected through multi-perspective evaluation, with 2 of 3 confirming predicted improvements in wet-lab validation (67% confirmation rate versus the 20-30% historical baseline).
- 6 months compressed from the lead optimization timeline — from a projected 18 remaining months to 12, bringing the program back within its funding milestone window.
- 40% reduction in lead optimization R&D costs, primarily from eliminating synthesis and testing of low-probability analogs. Helion estimated it avoided approximately 80 unnecessary analog syntheses at an average cost of $15,000 each.
- Complete modification trail for all 48 candidates, providing ready-made design rationale documentation for regulatory submissions — a deliverable that would normally require 4-6 weeks of dedicated effort from the chemistry team.
"The system explored modifications I would have vetoed in a design meeting — and two of them turned out to be our best candidates. That is a humbling and profoundly useful experience. The AI did not replace our chemists. It showed them the parts of chemical space where their own expertise was creating blind spots." — Dr. James Okafor, VP Drug Discovery, Helion Therapeutics
Key Takeaways
Local optima are an expertise problem, not a competence problem. Helion's chemists were making sound decisions within the modification space they explored. The AI's value was not better judgment within that space — it was systematic exploration of the space outside it.
Multi-step evaluation finds candidates that single-step SAR misses. Modifications that look unpromising at step one can open pathways to superior compounds at step two or three. The Tree of Thoughts architecture's ability to evaluate branches multiple steps deep is what distinguishes it from conventional virtual screening.
Independent analytical perspectives prevent groupthink in candidate selection. The Ensemble architecture's five agents surfaced trade-offs and conflicts that a single integrated model would have averaged away. Disagreement between agents was as informative as agreement.
Systematic search produces regulatory-ready documentation as a byproduct. Every branch explored and pruned is a documented design decision. For industries where regulators expect complete rationale trails, this is not a secondary benefit — it is a primary one.
Ready to Explore Systematic Search for Your R&D Pipeline?
Drug discovery timelines are measured in years and budgets in millions. Systematic AI search does not replace scientific expertise — it extends it into the regions of solution space where human intuition creates blind spots. Schedule a consultation to discuss how Systematic Search and Multi-Perspective Analyst architectures can accelerate your lead optimization and expand your molecular design horizons.