Subscribe for the Newsletter

Mobile Navigation

Closing the Loop Between AI Design and Clinical Translation

Closing the Loop Between AI Design and Clinical Translation

Nov 20, 2026PAO-11-26-CL-02

BigHat Biosciences has advanced BHB810, the first internally developed therapeutic to emerge from its integrated machine-learning and experimental platform, into the clinic. The milestone brings the company’s closed loop of computational design, proprietary experimental data, and laboratory validation into human testing, while a new $75 million Series C financing supports the next stage of its artificial intelligence platform and therapeutic pipeline. BigHat’s progress suggests that durable differentiation among AI-driven drug developers may depend on uniting sophisticated models with rapid experimentation, translational capabilities, and responsibility for the therapeutic outcome.

From Hypothesis to Clinical Testing

When Pharma’s Almanac first spoke with BigHat Biosciences in 2022, the company was building an integrated platform intended to shorten the distance between computational protein design and experimental evidence. In contrast to many of the new AI drug design startups, BigHat differentiated itself through the direct connection of an AI-based design system with high-throughput wet-lab experimentation and, increasingly, in vivo evaluation and therapeutic development within the same company. Machine learning (ML) can propose potentially improved molecules, but those designs still have to be produced, tested, and refined against the properties that determine whether a protein can become a viable therapeutic. Each cycle generates new evidence that is used to refine both the molecules and the models.

Four years later, BigHat has reached a milestone that places that original thesis under a more consequential test. The company has dosed the first patients in a clinical trial of BHB810, a cadherin 17 (CDH17)-directed antibody–drug conjugate (ADC) for gastric cancer and other advanced gastrointestinal tumors. It has also raised a $75 million Series C financing to advance BHB810, prepare a second program for the clinic, and continue expanding the platform behind them.

“Four years is a very long time,” Greenside said as she reflected on the interval between our conversations. “There’s been a lot of progress in all areas: pipeline, partnerships, and platform.”

That progress demonstrates that the company has carried its approach beyond computational predictions, laboratory optimization, and partner programs into the development of a therapeutic now being evaluated in humans. BHB810 is the first internally developed candidate to carry the design strategy BigHat initially described into patients.

Data Becomes the Differentiator

The basic limitation that motivated BigHat has not changed. No sufficiently reliable simulator can determine whether a computationally designed protein will function as intended in a biological system, exhibit the required pharmaceutical properties, and ultimately perform as a drug. The resulting designs still must be produced and tested experimentally.

What has changed is the scale and speed of that process. BigHat can now move approximately 3 ,000 molecules through its platform each week . The company measures between 10 and 20 properties across assays addressing affinity, biophysics, developability, and cell-based function. Those data return to the models and inform the next design cycle.

“Data is the key to unlock the best of AI,” Greenside said. As increasingly capable AI models and agents become more broadly available, she argues, access to relevant, high-quality experimental data becomes a more important source of differentiation. The value lies not simply in possessing a model but in generating the evidence needed to improve its predictions for therapeutic design.

BigHat has consequently expanded the biological range of its platform. Its initial work centered on solid tumor oncology, but additional functional assays have opened applications in immunology and enabled the company to evaluate more than binding and basic developability. The aim is to incorporate more of the biological behavior that a molecule must exhibit into the design-and-test cycle itself.

As the experimental system has scaled, AI has taken on a broader role than proposing protein sequences. BigHat is using agents to help coordinate the full design–build–test cycle, including translating design objectives into experimental plans, structuring incoming results, monitoring data quality, and using those findings to select the next molecules to produce. Scientists continue to define the therapeutic goals and determine which tradeoffs matter, but fewer manual handoffs now separate a model’s proposal from the experiment and the subsequent round of design. The result is a tighter learning system whose speed depends on both the number of sequences it can propose and how efficiently evidence from the laboratory shapes what happens next.

Owning the Therapeutic Outcome

BigHat’s development since 2022 also reflects an organizational choice. Rather than providing models, software, or candidate sequences and leaving experimental validation to others, the company has built the capacity to test whether its designs produce better therapeutics.

“We’re not just giving a partner a model or just relying on a model,” Greenside said. “We’re actually using the model to make a better therapeutic, and we’re owning the outcome.”

That responsibility extends from in vitro testing to in vivo evaluation in BigHat’s internal vivarium and, with BHB810, into clinical development. Greenside sees this continuity as one response to the gap between the considerable expectations surrounding AI in drug discovery and the limited number of AI-designed medicines that have reached patients.

“There’s a big translational gap between the potential of these methods and actually seeing them work,” she said. “We try to close that gap within one company.”

External collaborations have provided additional opportunities to apply the platform across different targets, modalities, and therapeutic areas. BigHat has disclosed partnerships with Amgen, Merck, Johnson & Johnson, AbbVie, and Eli Lilly and Company. It has completed three projects with Merck and a strategic collaboration with Johnson & Johnson, while its relationship with Lilly now includes participation in Lilly TuneLab as well as therapeutic collaborations.

Those programs complement BigHat’s internal pipeline. Partners bring problems and biological contexts beyond the company’s primary oncology and immunology focus, while BigHat contributes the capacity to design, build, and experimentally evaluate large sets of protein variants. “We learn as much from them as I think they learn from us,” Greenside said.

Designing BHB810 as a Complete Drug

BHB810 illustrates how BigHat applies that capability to a therapeutic problem. CDH17 is expressed in several gastrointestinal cancers, including gastric, colorectal, and pancreatic tumors, but no CDH17-directed therapy has yet been approved. Multiple drug developers are pursuing the target, making the design of the therapeutic critical to achieving the best possible outcome for patients and succeeding in an increasingly competitive marketplace .

Most approved ADCs use a full-length immunoglobulin G (IgG) antibody. BHB810 instead uses a smaller IgG-like structure containing the variable domain of a heavy-chain-only antibody (VHH) fused to an IgG1 fragment crystallizable region (VHH-Fc). The resulting construct is approximately half the size of a conventional IgG. BigHat selected the smaller format by showing that it could penetrate solid tumors more readily and deliver more payload to CDH17-expressing cells. The Fc supports bivalent target engagement and greater systemic persistence than a VHH alone.

While this format has been studied in humans in other diseases, BHB810 is one of the first VHH-Fc based ADCs. Most ADCs use the common IgG architecture. BigHat therefore sought to pair that structural choice with extensive optimization of the properties governing the molecule’s behavior.

“Affinity does not always line up with the functional property you care about, like internalization,” Greenside explained. A molecule can bind its target very tightly without delivering its payload into the cell as efficiently as required. The behavior of the antibody–receptor complex, including its internalization and intracellular trafficking, can matter more than binding strength alone.

BigHat optimized affinity, internalization, and developability concurrently rather than treating them as successive, independent problems. That distinction is central to the company’s account of BHB810. The goal was not to discover the strongest binder and then adapt it into an ADC but to identify the molecular design most likely to transport a cytotoxic payload into the tumor and then into the target cell.

The antibody format was only one part of that search. BigHat evaluated thousands of molecular variants and paired selected designs with different payload options. The final ADC incorporates a highly stable, site-specific linker-payload technology licensed from Synaffix and carries four payload molecules per antibody construct. BigHat selected the combination to preserve payload stability during circulation while supporting release after internalization.

“All these parts play together,” Greenside said. “We were able to combine different potential payload choices with different molecular designs and formats and explore a much wider combinatorial set of those pieces together.”

That work also influenced decisions extending beyond the molecule. “There’s the linker. There’s the payload. There’s the clinical strategy,” she said. “Every piece is important.”

Choosing Gastric Cancer

BigHat’s decision to begin clinical development in gastric cancer emerged from its comparative preclinical work rather than from CDH17 expression alone. The company evaluated BHB810 across nearly 30 patient- and cell-derived tumor models representing gastric and colorectal cancers. According to Greenside, approximately 80% of the relevant gastric cancer models showed a strong response, leading the company to prioritize that indication while retaining potential expansion opportunities in colorectal and pancreatic cancers. Up to about 70% of patients with gastric cancer express CDH17 at levels that can be targeted with BHB810.

Data presented by BigHat at the American Association for Cancer Research annual meeting included 13 primary gastric patient-derived xenograft (PDX) models. The company reported complete or near-complete tumor clearance across those models, including tumors with low or heterogeneous CDH17 expression. These findings supported the premise that the molecule’s penetration and internalization properties could remain useful when antigen distribution was uneven.

The development package also included a study in nonhuman primates. BigHat reported that no maximum tolerated dose was reached within the tested range, which extended to 5 mg/kg administered every two weeks. When considering the smaller molecule size of BHB810 compared with typical IgG ADCs, this is significantly higher than the doses that are typically safe in primates. These findings supported advancement into human testing but do not on their own establish the therapeutic window that will apply in patients.

“We’re seeing really strong safety and really strong early signs of efficacy in our preclinical program,” Greenside said. “And then, as usual, there’s no way to know for sure until you go into the clinic.”

Taking Molecular Risk Deliberately

Selecting a VHH-Fc ADC as BigHat’s first clinical candidate introduced a deliberate degree of format risk. VHH-based therapeutics have reached the market, and Fc-containing biologics have extensive clinical precedent, but the combination remains uncommon among ADCs. Greenside characterized BHB810 as innovative enough to test the platform’s design capabilities without beginning with the company’s most structurally ambitious concepts.

“Sometimes opportunities are somewhat risky, and I think that’s the beautiful thing,” she said.

That approach reflects the types of opportunities the company is currently prioritizing. BigHat is initially pursuing targets with established biological relevance for which molecular engineering may improve the therapeutic window, targeting profile, or functional activity. Over time, Greenside expects the company to assume more biological risk as well, using purpose-designed molecules to test mechanisms that conventional antibody formats cannot adequately interrogate.

The next internal program already broadens the test. BHB299 is an avidity-driven T cell engager (TCE) designed to target carcinoembryonic antigen-related cell adhesion molecule 6 (CEACAM6)-expressing solid tumors and recruit T cells through CD3. The program is approaching completion of preclinical development, and BigHat plans to begin clinical testing in 2027.

The contrast between the two lead assets is intentional. BHB810 tests whether an ADC built around the smaller VHH-Fc format can produce a useful therapeutic profile in solid tumors. BHB299 asks whether avidity and multispecific design can create sufficiently selective T cell engagement against another solid tumor target. Each program challenges the platform in a different way.

The Clinic Creates a New Feedback Loop

The first patients with advanced gastrointestinal cancer have now received BHB810 in the dose-escalation portion of a registered phase I/II study. The initial phase will evaluate safety, tolerability, pharmacokinetics, and preliminary antitumor activity before any potential dose expansion.

For BigHat, clinical development adds forms of evidence that preclinical systems cannot fully reproduce. Human data will test assumptions about exposure, tumor delivery, toxicity, efficacy, and immunogenicity. They will also reveal whether the engineered relationships among molecular size, affinity, internalization, linker stability, and payload delivery behave as expected in patients.

Greenside does not expect one early-stage trial to yield a general solution to clinical prediction. However, she does expect the findings to influence how BigHat evaluates future designs. Evidence that a particular format or engineered property is safe and functionally useful could justify incorporating it into more ambitious molecules. Unexpected behavior could redirect the platform just as meaningfully.

“The clinic is ultimately the final test,” she said. “We’re going to take advantage of where we see positive safety and efficacy data to lean into those design challenges that start to be proven out by us and start taking on more ambitious designs as we see certain successes or learn what is or isn’t working.”

Those human data will arrive much more slowly than results from BigHat’s weekly laboratory cycles, but they will carry far greater consequences.

Financing the Next Iteration

BigHat’s $75 million Series C brings the company’s total funding to $223 million. DFJ Growth and Premji Invest co-led the round, with participation from new and existing financial, pharmaceutical, and strategic investors.

The financing will support both sides of BigHat’s strategy. On the therapeutic side, the company plans to advance BHB810 and BHB299 toward key clinical readouts while continuing to develop a broader proprietary pipeline in oncology and immunology. On the platform side, it intends to increase experimental throughput, expand the range of assays and data types generated internally, and continue developing the computational systems that use those data.

This dual allocation is consistent with Greenside’s view that the platform and pipeline should strengthen one another. Internal programs create demanding therapeutic design problems and eventually produce clinical evidence. Partnerships expose the platform to a broader range of targets and molecular requirements. Expanded experimental capabilities allow both categories of work to generate richer data.

The financing therefore supports more than the advancement of two assets. It gives BigHat additional time and capacity to determine whether the relationship it has constructed among data generation, AI-guided design, and therapeutic development can produce repeatable results.

Building a Platform and a Clinical-Stage Company

BigHat’s progress since 2022 has brought it closer to the company Greenside described during our first conversation. It has expanded its partnerships, increased the scale of its experimental platform, developed an internal pipeline, and advanced its first program into clinical testing. Reaching the clinic changes the standard against which that progress will be judged.

BHB810 must now show whether the properties optimized in the laboratory produce a tolerable and potentially effective therapy. BHB299 must demonstrate that the platform can address a different format and mechanism. The company must also continue increasing the speed and breadth of its design system without compromising the quality of the experimental evidence that distinguishes it.

Greenside’s ambition is to preserve both identities as BigHat moves forward. “We want to continue to be a leading platform and an effective clinical-stage company,” she said. “That’s a dream really well worth chasing aggressively.”

STAGING