AI-Accelerated Drug Discovery for SARS-CoV-2 Protease Inhibitors

All case studies

Infectious Disease

AI-Accelerated Drug Discovery for SARS-CoV-2 Protease Inhibitors

Mid-stage biotech hit identification for a novel coronavirus protease inhibitor program.

Company

Mid-stage biotech (Series A, 15-person team)

Timeline

January to February 2025

Engagement

Small Molecule Discovery Pipeline

Small Molecule Discovery

20×: Faster than traditional HTS
87.5%: Cost savings vs. HTS
50%: Wet-lab hit rate (top 12)
100%: Hit rate for ΔG < -8.0

The Challenge

A Series A biotech had identified SARS-CoV-2 main protease (3CLpro) as a high-priority target for next-generation antivirals. Their internal screening had stalled: traditional high-throughput screening (HTS) would cost $2.8M and take 8 to 10 months. The company needed validated hits within 12 weeks to secure Series B investor confidence.

Business Constraints

Budget: $350K (tight Series A ceiling)
Timeline: Ranked candidates in 4 weeks, validation-ready
Risk tolerance: Low. Needed high-quality leads, not fishing expeditions

HelixForge Approach

Week 1 to 2: Candidate Generation and Virtual Screening

Input

SARS-CoV-2 3CLpro crystal structure (PDB: 7JTL)
Internal compound library (12,000 molecules from previous programs)
ChEMBL/PubChem additional chemical space (2.3M compounds screened)

Methods

Graph neural network (GNN) scoring trained on proprietary validation datasets
Initial ranking across 2.3M compounds
Structural diversity filtering to avoid bias toward known inhibitors

Output

Top 15,000 candidates ranked by predicted binding affinity
Structural clustering to identify scaffolds

Week 2 to 3: Refinement and Physics-Based Scoring

Molecular docking (AutoDock Vina + DiffDock) re-scored the top 15,000 compounds for binding pose quality and filtered to 2,847 compounds with ΔG < -7.5 kcal/mol. GROMACS 500ns MD simulations on the top 250 candidates refined binding free energy (MM-PBSA) and filtered to 89 compounds with stable binding modes. ADME and toxicity scoring reduced the pool to 56 drug-like compounds. Off-target prediction against 300+ human serine proteases yielded a final pool of 47 candidates.

Week 4: Ranking and Validation Playbook

Output

Top 50 ranked candidates with scaffold, ΔG, ADME score, selectivity, and priority tier
Recommended wet-lab assays and IC50 concentration ranges
Selectivity panel recommendations and formulation guidance

Final Ranked Output

Top candidates shown; full list of 50 delivered under NDA.

Top candidates by predicted binding affinity and developability
Rank	Scaffold	ΔG (kcal/mol)	ADME Score	Selectivity	Status
1	Pyridone-based	-8.9	0.87	0.92	Priority A
2	Thiazole-benzamide	-8.7	0.84	0.88	Priority A
3	Quinolone derivative	-8.5	0.81	0.85	Priority A
4–5	Mixed	-8.3 to -8.2	0.80–0.82	0.85–0.87	Priority B
6–10	Mixed	-8.1 to -7.8	0.78–0.82	0.80–0.87	Priority B
11–50	Mixed	-7.5 to -7.1	0.70–0.80	0.75–0.85	Priority C

Results and impact

Speed, validation, and business outcomes

Speed vs. Traditional HTS

Metric	Traditional HTS	HelixForge	Improvement
Timeline	8 to 10 months	4 weeks	20× faster
Cost	$2.8M	$350K	87.5% savings
Compounds Screened	5M	2.3M (+ physics refinement)	Same coverage, better ranking
Predicted Hit Rate	~10% (500 hits)	~24% (50 to 100 high-quality leads)	2.4× improvement

Wet-Lab Validation Outcomes (6 weeks post-delivery)

Assay results on top 12 candidates.

Candidate	Predicted ΔG	Observed IC50 (nM)	Hit?	Notes
Rank 1	-8.9	340	Yes	Excellent selectivity vs. human proteases
Rank 2	-8.7	420	Yes	Good microsomal stability
Rank 3	-8.5	510	Yes	Hits target; needs formulation work
Rank 4	-8.3	1,200	Yes	Moderate potency, excellent ADME
Rank 5	-8.2	2,100	Yes	Acceptable for follow-up
Rank 6	-8.1	15,000	Weak	Selectivity issues flagged
Rank 7–12	-7.9 to -7.5	Inactive	No	No activity; assay controls confirm

100%: Hit rate for ΔG < -8.0 (5/5 active)
50%: Hit rate across top 12 (6/12)
~8%: False positive rate
6 wks: Timeline to first active compound

Immediate Wins

De-risked Series B narrative: founders presented 6 validated hits plus 44 backup candidates to investors
Accelerated program: moved to medicinal chemistry optimization 3 months earlier than planned
Cost reallocation: saved $2.45M redirected to in-house chemistry and preclinical studies

Strategic Advantages

Informed SAR: top 3 scaffolds revealed novel binding modes; medicinal chemistry team rapidly synthesized derivatives
De-risked chemistry: predicted ADME properties steered synthesis away from lipophilic dead-ends
Selectivity foundation: built selectivity strategy on computational predictions before wet-lab work

Follow-on engagement
Q2 2025: second round using HelixForge to optimize the top scaffold. Estimated next project cost: $180K. Projected 2 to 4 month drug lead candidate.

Model validation

Lessons and recommendations

What Worked

Closed-loop feedback: wet-lab IC50 data re-trained GNN for follow-on rounds; predictions tightened
Physics + AI hybrid: MD simulations caught 3 candidates with unstable binding modes that docking alone would have scored highly
Selectivity focus: early off-target filtering saved the medicinal chemistry team from pursuing non-selective inhibitors

Challenges and Mitigations

Rank 6 candidate scored well computationally but showed weak potency in assay. Root cause: binding mode prediction uncertainty at -8.1 kcal/mol threshold.

Mitigation: Tightened ΔG filter to -8.2+ for Rank 2 to 5; applied to follow-on projects.

GROMACS simulations took longer than expected (some runs required 5 to 7 days). Root cause: high flexibility in binding pocket.

Mitigation: Parallelized simulations; reduced top-N candidates for reranking from 250 to 150.

When to use HelixForge for virtual screening

High-throughput screening budgets above $500K
Timeline pressure (under 6 months to hits)
Complex targets (proteases, kinases, difficult binding pockets)
Need for selectivity or ADME filtering early in discovery

ROI: approximately 7:1 (cost savings + speed advantage vs. traditional HTS).

Next steps: start with 2 to 3 target programs; iterate scoring models with wet-lab feedback for continuous improvement.

About This Engagement

Client profile: Series A biotech, 15 employees, prior HTS experience
Project duration: 4 weeks (computational delivery) + 6 weeks (validation)
Total cost: $350K
Date: January to February 2025

This case study is anonymized at client request. Specific compound identifiers and institutional affiliations have been redacted. Raw data and full computational protocols available under NDA.

Run a similar program with HelixForge.

Tell us about your target. We will scope a 2 to 4 week pilot.

Schedule a Consultation [email protected]