Company of the week: Iambic Therapeutics

Company Overview and Technology

Iambic Therapeutics is a San Diego-based clinical-stage biotechnology company using physics-informed artificial intelligence to discover and develop novel small-molecule drugs, primarily in oncology. Founded in 2019 as Entos by two computational chemistry professors — CEO Tom Miller from Caltech and CTO Fred Manby from the University of Bristol — the company rebranded to Iambic Therapeutics in October 2023 and has since emerged as one of the most closely watched players in the AI drug discovery space.

The company's core thesis is that integrating physics principles directly into AI architectures — rather than relying on purely data-driven or purely physics-based approaches — produces more accurate, data-efficient models for drug design. This philosophy has yielded a suite of proprietary AI tools spanning protein structure prediction, quantum chemistry, clinical property forecasting, and generative molecular design, all tightly integrated with an automated wet laboratory that executes design-make-test-analyze (DMTA) cycles on a weekly cadence.

In the five years since its founding, Iambic has raised over $346 million from investors including Sequoia, Coatue, NVIDIA, Regeneron Ventures, Qatar Investment Authority, and Mubadala Capital. It has progressed its lead program — IAM1363, a highly selective HER2 inhibitor — from conception to Phase 1/1b clinical trials in under 24 months, a timeline that would be exceptional by traditional drug discovery standards and is virtually unprecedented for an AI-native company. In February 2026, Iambic signed a landmark multi-year collaboration with Takeda worth up to $1.7 billion in milestone payments plus royalties, one of the largest AI drug discovery deals to date.

The Problem Iambic Is Solving

Traditional small-molecule drug discovery is slow, expensive, and failure-prone. The average drug takes 10–15 years from target identification to approval and costs over $2 billion, with approximately 90% of clinical candidates failing. The bottleneck is not a lack of chemical ideas but rather an inability to accurately predict how molecules will behave — how they bind to protein targets, how they are metabolized, whether they cross biological barriers like the blood-brain barrier, and whether they will be safe and effective in humans.

AI promises to compress these timelines by replacing slow empirical cycles with rapid computational prediction. However, the field faces a fundamental tension: purely data-driven machine learning models can interpolate well within known chemical space but struggle to generalize to novel targets and scaffolds, while purely physics-based simulations (molecular dynamics, quantum mechanics) are accurate in principle but computationally prohibitive at the scale needed for drug discovery.

Iambic's approach attempts to resolve this tension by embedding physical laws — symmetry constraints, quantum mechanical features, thermodynamic principles — directly into neural network architectures. The result, the company argues, is models that are both more accurate and more generalizable than either approach alone.

The Iambic Platform: Five Technology Pillars

Iambic's platform comprises five named AI tools, each addressing a different challenge in the drug discovery pipeline, integrated with an automated high-throughput chemistry and biology laboratory.

NeuralPLexer: Predicting How Drugs Bind to Proteins

NeuralPLexer is Iambic's flagship AI model and arguably its most important scientific contribution. It predicts the three-dimensional structure of protein-ligand complexes — how a drug molecule physically sits within its protein target — directly from the protein's amino acid sequence and the molecule's chemical structure.

What makes NeuralPLexer distinctive is its ability to predict conformational changes: how proteins change shape when a drug binds. This is critical because many drug targets only reveal their binding sites (so-called "cryptic pockets") upon conformational rearrangement, and allosteric drugs work precisely by exploiting these shape changes. Traditional structure prediction tools like AlphaFold2 excel at predicting static protein structures but cannot model these dynamics.

The technical architecture uses a two-module approach: a coarse-grained Contact Prediction Module that predicts residue-level contact maps, followed by an atomistic Equivariant Structure Denoising Module that iteratively samples all heavy-atom coordinates through a diffusion process incorporating biophysical constraints. In plain terms, the model first roughly maps which parts of the protein will be near the drug, then refines the exact atomic positions of every atom while respecting the physical rules of molecular geometry.

NeuralPLexer 3 replaced the diffusion framework with flow matching — a related but more computationally efficient generative approach — and incorporates physics-inspired priors from a globular polymer model with harmonic connectivity terms. On the PoseBusters benchmark, it achieved 97% accuracy on high-confidence predictions and 99% accuracy on ligand stereochemistry.

NeuralPLexer code is available on GitHub under a CC BY-NC-SA 4.0 license and has been deployed on NVIDIA's BioNeMo platform. This open-source strategy — rare among drug discovery companies — serves both as scientific validation and as a distribution channel that brings the broader computational chemistry community into Iambic's ecosystem.

OrbNet: AI-Accelerated Quantum Chemistry

OrbNet addresses the problem of calculating molecular interaction energies. In drug design, knowing how tightly a molecule binds to its target is essential. The gold standard for computing binding energies is density functional theory (DFT), a quantum mechanical method. DFT calculations are computationally expensive — a single protein-ligand binding energy calculation can take hours to days — making large-scale virtual screening impractical.

OrbNet is a graph neural network that predicts molecular properties with DFT-level accuracy at approximately 1,000× reduced computational cost. Its key innovation is incorporating quantum mechanical features — specifically, symmetry-adapted atomic orbital features from cheap semi-empirical electronic structure calculations — as inputs to the neural network. Rather than learning purely from molecular geometry, the network receives physically meaningful representations of electronic interactions, improving data efficiency and transferability to new chemical space.

OrbNet Denali was trained on 2.3 million DFT calculations covering 17 elements (including common drug atoms like carbon, nitrogen, oxygen, sulfur, and halogens) and matched the accuracy of modern DFT functionals on standard benchmarks at a fraction of the cost.

Enchant: Predicting Clinical Outcomes Early

Perhaps the most ambitious technology in Iambic's stack, Enchant is a multi-modal transformer model designed to predict clinical and preclinical drug properties — pharmacokinetics, ADME (absorption, distribution, metabolism, excretion), toxicity, and efficacy — from the earliest stages of drug discovery.

The fundamental challenge Enchant addresses is the "data wall" between preclinical and clinical drug development. Abundant experimental data exists from laboratory assays and animal studies, but clinical data on how drugs behave in humans is scarce and expensive to generate. Most drug candidates fail in clinical trials precisely because their real-world pharmacokinetics were not accurately predicted from preclinical data.

Enchant attempts to bridge this gap by training a multi-modal transformer on dozens of public and proprietary data sources simultaneously, learning to translate between different data modalities (in vitro assays, animal pharmacokinetics, human clinical data) rather than treating them as independent prediction tasks.

Enchant v2 was trained using NVIDIA HGX B200 GPU clusters and is fine-tuned weekly on experimental data from Iambic's own discovery programs, creating a feedback loop between computational predictions and wet-lab validation. If these claims hold up at scale, the ability to predict clinical pharmacokinetics from early-stage data could eliminate years of preclinical optimization and reduce late-stage clinical failures. However, this is also where Iambic's claims are hardest to independently verify, since the model's true predictive power will only become clear as more programs advance through clinical development.

ProPANE and Magnet: Optimization and Generation

ProPANE is a massively pre-trained graph neural network deployed across dozens of drug properties for lead selection and optimization. It predicts ADME endpoints — solubility, permeability, metabolic stability, bioavailability — and supports automated training, uncertainty quantification, and explainability features.

Magnet is Iambic's suite of generative AI technologies for de novo molecular design. It produces novel chemical structures optimized not just for biological activity but for practical synthesizability on Iambic's automated chemistry platform — an important constraint, since many generative models propose molecules that look good computationally but are impossible to actually make in the lab.

The Integrated Loop

These technologies form a tightly integrated design-make-test-analyze cycle. NeuralPLexer predicts binding structures; OrbNet evaluates binding energetics; Magnet generates novel molecules; ProPANE evaluates multi-parameter drug-likeness; Enchant predicts clinical viability. Automated nanoscale chemistry synthesizes thousands of unique compounds per week, covering >95% of standard medicinal chemistry transformations. High-throughput biology generates biochemical, metabolic, and cellular data. Automated data pipelines feed back into model retraining every week.

This weekly cadence — completing a full DMTA cycle in approximately 10 days — is a claimed key differentiator versus both traditional drug discovery (months per cycle) and AI-only companies that lack integrated wet lab capabilities. The weekly data generation also creates a proprietary flywheel: each experimental cycle produces new training data that improves the models, which in turn generate better predictions for the next cycle.

Key Assumptions and Limitations of the Technology

Iambic's technology rests on several assumptions worth examining:

Pipeline and Assets

Iambic's pipeline is focused on oncology, with a secondary thrust into neurology via a partnership with Lundbeck. The company has disclosed four programs at various stages of development.

IAM1363: A Selective, Brain-Penetrant HER2 Inhibitor (Lead Program)

IAM1363 is Iambic's most advanced program and serves as the primary proof-of-concept for the company's platform. It is a potent, irreversible, type II covalent tyrosine kinase inhibitor targeting HER2 — a well-validated oncology target overexpressed or mutated in breast, gastric, lung, and other cancers.

What distinguishes IAM1363 from existing HER2 inhibitors is its selectivity profile: >5,000-fold selectivity for HER2 over EGFR, the closely related receptor whose inhibition causes dose-limiting skin and gastrointestinal toxicities with current therapies like tucatinib and neratinib. IAM1363 also demonstrates brain penetrance approximately 10-fold greater than approved HER2 tyrosine kinase inhibitors — critical because brain metastases are common in HER2-positive cancers and represent a major unmet need.

Phase 1/1b clinical data were presented at the ESMO Congress in Berlin (October 2025). In patients dosed at ≥960 mg once daily, IAM1363 showed a 28% partial response rate (5 of 18 patients) in those with measurable systemic disease and a 33% partial response rate (1 of 3 patients) with measurable intracranial tumors. Activity was observed across HER2-amplified, HER2-overexpressing, and HER2-mutant tumors — including breast, gastric, NSCLC, renal cell, and ovarian cancers. The safety profile was favorable, with two well-tolerated doses advancing into dose optimization.

Dose escalation was completed in June 2025, and the trial has expanded from U.S. sites to the EU, with further expansion to the UK and APAC. In October 2025, Iambic announced a collaboration with Jazz Pharmaceuticals to evaluate IAM1363 in combination with zanidatamab (Ziihera®, a HER2-bispecific antibody) plus capecitabine in HER2-positive breast cancer patients previously treated with Enhertu® (T-DXd).

IAM-C1: First-in-Class CDK2/4 Inhibitor

IAM-C1 is a selective dual inhibitor of CDK2 and CDK4 while sparing CDK1, CDK6, and CDK9. This selectivity profile is first-in-class. Approved CDK4/6 inhibitors (palbociclib, ribociclib, abemaciclib) inhibit CDK6 along with CDK4, causing dose-limiting neutropenia, and they do not address CDK2, which is a key resistance mechanism.

The rationale is that many cancers develop resistance to CDK4/6 inhibitors through CDK2-mediated bypass, and CDK6 inhibition causes the most problematic toxicities. By hitting CDK2 and CDK4 while sparing CDK6, IAM-C1 aims to address both intrinsic and acquired resistance while improving tolerability. The program targets HR+/HER2- metastatic breast cancer and is currently in IND-enabling studies with targeted clinical entry in 2026.

KIF18A Allosteric Inhibitor

Iambic's KIF18A program targets a kinesin motor protein essential for proper chromosome segregation during cell division. Tumor cells with chromosomal instability (CIN) — a hallmark of many aggressive solid tumors — are critically dependent on KIF18A for survival, while normal cells are not. This creates a therapeutic window: inhibiting KIF18A should selectively kill chromosomally unstable tumor cells while sparing healthy tissue.

Iambic has developed allosteric inhibitors with single-digit nanomolar potency in biochemical and cell growth assays. Preclinical data presented at AACR 2025 showed potent tumor growth inhibition via mitotic arrest. The program is in IND-enabling studies.

Lundbeck Partnership: Migraine GPCR Program

In September 2024, Iambic announced a strategic research collaboration with Lundbeck for the discovery of a small-molecule therapeutic targeting an undisclosed GPCR for migraine treatment. The program involves an allosteric agonist modality and is in discovery stage.

Pipeline Summary

Partnerships

Iambic has executed five major partnerships in approximately 18 months, transitioning from a pure pipeline company to a dual-model organization that advances wholly-owned programs while monetizing its platform through technology collaborations.

Takeda (February 2026)

The Takeda collaboration is Iambic's most significant deal and one of the largest AI drug discovery partnerships announced to date. Under the multi-year agreement, Iambic will deploy its full AI platform to advance Takeda's high-priority small molecule programs, initially in oncology and gastrointestinal and inflammation therapeutic areas. Takeda also receives non-exclusive access to NeuralPLexer for broader drug discovery efforts. Financial terms include undisclosed upfront, research cost, and technology access payments, with success-based milestones potentially exceeding $1.7 billion plus royalties on net sales.

Revolution Medicines (July 2025)

A $25 million technology collaboration in which Iambic trains bespoke versions of NeuralPLexer using Revolution Medicines' proprietary structural and molecular libraries. Revolution also gains access to ProPANE for lead optimization. Both companies retain rights to exclusive targets.

Jazz Pharmaceuticals (October 2025)

A research collaboration and drug supply agreement in which Jazz provides zanidatamab at no cost for evaluation in combination with IAM1363 in HER2-positive breast cancer patients.

Lundbeck (September 2024)

A strategic research collaboration for small molecule discovery in migraine, with upfront payment, performance-based milestones, and royalties (amounts undisclosed).

NVIDIA (Ongoing Since 2023)

Iambic's relationship with NVIDIA is both financial (NVIDIA invested in the Series B) and technological. NeuralPLexer has been integrated into NVIDIA's BioNeMo platform, and Iambic leverages NVIDIA DGX Cloud infrastructure. Scientific advisor Anima Anandkumar was formerly NVIDIA's Senior Director of AI and a Caltech colleague of Tom Miller's. Iambic was prominently featured in CEO Jensen Huang's GTC24 keynote in March 2024.

Funding History and Financial Position

Iambic has raised over $346 million across multiple financing rounds, assembling an investor syndicate spanning top-tier venture capital, sovereign wealth funds, strategic pharma investors, and technology companies.

Beyond equity financing, Iambic has secured meaningful non-dilutive income through its partnership model. The Takeda deal alone includes undisclosed upfront and research payments plus up to $1.7B+ in milestones plus royalties. The Revolution Medicines collaboration is worth up to $25M. CEO Miller stated in February 2026 that Iambic has enough cash to last into 2028.

Leadership and Governance

Founders and Executive Team

Scientific and Clinical Advisory Board

The inclusion of Derek Lowe — whose widely-read "In the Pipeline" blog is known for its skeptical but rigorous assessment of drug discovery claims — is a notable credibility signal.

Red Team vs. Blue Team Analysis

The following sections present a Red Team vs. Blue Team analysis of Iambic Therapeutics, focusing on three critical dimensions: (1) Scientific Validity of the Platform, (2) Commercial and Pipeline Prospects, and (3) Competitive Positioning. The "Blue Team" represents an optimistic or bullish perspective, emphasizing strengths and opportunities. The "Red Team" offers a skeptical or cautious perspective, probing potential weaknesses, risks, and challenges.

Scientific Validity of the Platform

Blue Team Perspective: Scientific Rationale and Evidence

Unlike many AI drug discovery companies whose technology claims are difficult to verify, Iambic's core technology has been validated through peer-reviewed publication in Nature Machine Intelligence and public benchmarking. NeuralPLexer 3 achieves 78% success rate on the PoseBusters benchmark versus AlphaFold 3's 73%, and its unique ability to predict conformational changes upon ligand binding addresses a genuine gap that no competing tool currently fills. The open-source release on GitHub allows independent verification — this is not vaporware.

The most compelling evidence for the platform is IAM1363 itself. Progressing from program initiation to Phase 1 clinical trial in under 24 months — versus an industry average of 4–5 years — is a concrete, verifiable demonstration of platform value. Early Phase 1/1b data showing a 28% partial response rate in heavily pretreated patients, activity across multiple HER2-driven tumor types, and intracranial responses with a favorable safety profile suggests that the AI-designed molecule is not just fast but actually differentiated.

Iambic's physics-informed approach — embedding quantum mechanical features and biophysical constraints into neural architectures — represents a genuine scientific differentiation that is difficult to replicate. Competing approaches tend to be either purely data-driven (limited by training data, prone to poor generalization) or purely physics-based (computationally expensive). Iambic's hybrid approach requires deep expertise in both computational chemistry and machine learning — exactly the combination its founders bring. This intellectual moat is reinforced by proprietary datasets (2.3 million DFT calculations for OrbNet, weekly experimental data from the wet lab) and the continuous learning loop between models and experiments.

Key Scientific Strengths:

Red Team Perspective: Scientific Questions and Risks

Despite the encouraging proof-of-concept, several scientific questions remain. While the Phase 1/1b data for IAM1363 are encouraging, they are early. A 28% partial response rate in 18 evaluable patients is promising but far from definitive — sample sizes are small, durability of responses is unknown, and the competitive HER2 landscape is crowded and rapidly evolving. Enhertu (trastuzumab deruxtecan) has transformed HER2-positive cancer treatment, and patients in Iambic's trial are heavily pretreated, meaning those who respond may represent a selected subpopulation.

The speed-to-clinic claim requires context. HER2 is among the most well-characterized targets in oncology — decades of structural, biochemical, and clinical data exist. Designing a selective HER2 inhibitor, while non-trivial, is fundamentally different from discovering drugs against novel or poorly characterized targets. The true test of the platform will come with programs like KIF18A or CDK2/4, where less prior knowledge exists. Until those programs advance, the IAM1363 timeline may overstate the platform's general applicability.

Enchant's core promise — predicting human pharmacokinetics from early-stage data — is the highest-value and highest-risk claim in the technology stack. The history of drug development is filled with failed attempts to predict clinical outcomes from preclinical data. While Enchant's benchmark performance (0.74 Spearman correlation on Obach PK data) is a significant improvement over prior state-of-the-art, benchmark performance on retrospective datasets does not guarantee prospective accuracy in novel chemical series. Validation will require years of clinical data across multiple programs — data that does not yet exist.

Scientific Concerns:

Commercial and Pipeline Prospects

Blue Team Perspective: Business Model and Market Timing

The company's evolution from pure pipeline to dual pipeline-plus-platform model is strategically sound. Technology partnerships (Takeda, Revolution Medicines, Lundbeck) provide non-dilutive revenue and validate the platform's utility, while wholly-owned programs (IAM1363, IAM-C1, KIF18A) capture the full economic upside of successful drugs. The Takeda deal alone, if milestones are substantially achieved, could fund the company for years without further dilution.

Iambic is riding a wave of pharma interest in AI drug discovery — Eli Lilly, Novartis, Sanofi, and Takeda have all signed major AI deals in 2024–2025. The company's investor base (Sequoia, Regeneron Ventures, QIA, Mubadala, ARK) provides both capital and strategic connections. With cash runway into 2028 and no immediate need for an IPO, Iambic has the luxury of executing from a position of financial strength.

The AI drug discovery market is projected to grow from approximately $2–5 billion in 2025 to $8–20 billion by 2030 at a ~25–30% CAGR. Within this, the small molecule segment represents roughly 58% of the market, and oncology is the leading therapeutic area (~24% share) — both squarely in Iambic's wheelhouse. HER2-driven cancers alone represent a multi-billion-dollar market opportunity.

Red Team Perspective: Commercial Risks

The headline $1.7 billion in the Takeda deal reflects success-based milestone payments — the actual upfront and near-term payments are undisclosed but almost certainly a small fraction of the headline number. Industry analysis of AI drug discovery partnerships suggests that upfront payments typically represent only about 2% of announced deal values. The gap between headline "biobucks" and actual economics is a systemic feature of the industry that inflates perceived validation.

Despite over $15 billion in announced partnership deals and more than 30 AI-designed drugs entering clinical trials, no AI-discovered drug has yet received regulatory approval. The most advanced AI-designed drug — Insilico Medicine's rentosertib for idiopathic pulmonary fibrosis — has Phase IIa data but is years from potential approval. Multiple high-profile AI drug programs have been discontinued (Recursion's REC-994 and REC-2282, Exscientia pipeline setbacks, BenevolentAI's baricitinib COVID trial failure). The industry as a whole has yet to prove that AI fundamentally improves clinical success rates rather than simply producing faster failures.

AI can accelerate drug design, but it cannot compress the fundamentally time-bound aspects of drug development: clinical trial enrollment, patient follow-up, regulatory review, manufacturing scale-up. Even if Iambic designs perfect molecules instantly, Phase 2 and Phase 3 trials will take years. Many clinical failures result from biological complexity that no computational prediction can fully anticipate.

Commercial Risk Factors:

Competitive Landscape

Key Competitors

Blue Team Perspective: Where Iambic Wins

Iambic occupies a specific niche: physics-informed AI for small molecule drug discovery with integrated wet lab capabilities. This distinguishes it from Isomorphic Labs (pure ML/deep learning tradition, no integrated chemistry lab), Recursion (biological data scale and phenotypic screening rather than structure-based design), Insilico (end-to-end but without depth in quantum chemistry), and Schrödinger (philosophically similar but public company, decades-old platform, broader customer base).

The fastest-to-clinic comparison is instructive: Iambic brought IAM1363 from program start to Phase 1 in under 24 months. No other AI company has publicly claimed a comparable speed-to-clinic milestone for a wholly-owned asset. NeuralPLexer's conformational dynamics capability addresses a genuine gap — if AlphaFold 3 does not close this gap, Iambic retains a meaningful technical edge in structure-based drug design.

Red Team Perspective: Competitive Threats

Isomorphic Labs has $600 million in funding, Nobel Prize-winning AlphaFold technology, and deals worth ~$3 billion with Eli Lilly and Novartis. If AlphaFold 3 closes the conformational dynamics gap, Iambic's NeuralPLexer advantage narrows considerably. Recursion operates at far greater scale with 60+ petabytes of biological data and acquired Exscientia's precision chemistry capabilities. Genesis Molecular AI raised $300M+ and claims its Pearl model outperforms AlphaFold 3 on protein-ligand predictions. Schrödinger has decades of physics-based modeling expertise and a profitable software business funding its pipeline.

The competitive landscape is intensifying rapidly, and Iambic — as a private, pre-revenue company — faces well-funded competitors on multiple fronts. The window of technological differentiation may narrow as competitors invest in conformational dynamics and physics-informed approaches. Being first matters, but staying ahead requires continuous innovation.

Key Milestones Timeline (2024–February 2026)

Conclusion

Iambic Therapeutics presents a compelling and relatively differentiated entry in the crowded AI drug discovery landscape. Unlike many competitors whose claims rest primarily on in silico benchmarks or partnership announcements, Iambic has progressed a wholly-designed molecule into clinical trials and generated early clinical data showing biological activity — a concrete milestone that only a handful of AI-native companies can claim.

The company's physics-informed approach represents a genuine intellectual contribution validated through top-tier publication and public benchmarking. NeuralPLexer's ability to predict conformational changes fills a real gap in the structural biology toolkit, and the integrated DMTA loop with weekly cycle times creates a proprietary data flywheel that compounds over time. The addition of Enchant — if its clinical prediction capabilities prove out prospectively — could represent a step-change in how early drug candidates are triaged.

The key open questions are whether IAM1363's early clinical signals translate into durable, clinically meaningful benefit in larger trials; whether the platform's demonstrated speed on the well-characterized HER2 target generalizes to novel targets (IAM-C1 and KIF18A will be the test cases); and whether Enchant's ambitious clinical prediction claims survive real-world validation. The Takeda partnership provides both financial validation and a critical external test of the platform's utility across different therapeutic areas.

The risks are real: no AI-discovered drug has been approved, the competitive landscape is intensifying rapidly, and the fundamental challenges of clinical translation — patient heterogeneity, biological complexity, regulatory timelines — remain stubbornly resistant to computational shortcuts. The gap between announced deal values and actual near-term economics is also worth keeping in mind.

With approximately $346 million raised and cash runway into 2028, five major partnerships signed, a clinical-stage lead program showing early activity, and two additional programs approaching clinical entry, Iambic is well-positioned to test its thesis over the next 2–3 years. The company's trajectory over that period — particularly the maturation of IAM1363 clinical data and the clinic entry of IAM-C1 and KIF18A — will determine whether Iambic's physics-informed AI approach represents a genuine advance in drug discovery or another promising technology that struggles to overcome the irreducible complexity of human biology.

Disclaimer: The author is not a lawyer or financial adviser. The content presented in this article is for informational purposes only and does not constitute investment advice, legal advice, or a recommendation to buy or sell any securities. Readers should conduct their own due diligence and consult with qualified professionals before making any investment decisions.