Decoding the structure-function relationship for better medicines

Functional prediction

across chemical space

The most important compound in your program may be one you never tested. Our AI converts your phenotypic screening data into a computational discovery engine, driving compound prioritization, early-stage safety profiling, and hit-to-lead decisions.

10M+
Functional experiments
Largest phenotypic model to date
60×
Reduction in screening volume
vs. traditional HTS
80%
Hit discovery rate
Experimentally validated
<4 wks
From data to prioritized candidates
vs. 12+ month cycle
The challenge

Drug discovery operates on incomplete data

Each preclinical cycle takes months and millions, yet most candidates still fail later because teams are forced to advance compounds on partial evidence. Structure-based methods can suggest binding, but they do not show how compounds will behave in vivo.

01

Narrow screening misses the full picture

Traditional assays evaluate one phenotype in one cell type. Adverse effects, off-target activity, and context-dependent responses remain hidden until costly later stages.

02

Historical data can still drive discovery

Previous iterations generate functional data that are rarely reused when a program stalls or pivots. Those datasets may still contain mechanistic insight the next campaign can use.

03

Structure does not equal function

Binding affinity models cannot capture polypharmacology, pathway crosstalk, or cell-context-specific responses. The gap between target selection and therapeutic effect remains wide.

Our solution

What your screening cascade is not telling you

Our bulk foundation model scores small molecules by predicted functional response across viability, expression, morphology and more. Discovery teams use it to triage leads before committing to follow-up experiments, flag potential risks before in vivo studies, and classify mechanisms of action without target-dependent assays.

01

Profile compounds across the full biological space

Our model predicts phenotypic response across diverse cellular contexts, identifying off-target liabilities and context-dependent effects that single-assay workflows cannot. Teams can assess risk earlier, before committing to secondary assays or additional chemistry.

1,400+ cell lines
02

Recover value from prior screens

Our model learns across assays, allowing past screening data to improve predictions on new compounds and functional endpoints. Legacy data and shelved assets become the basis for the next round of candidate ranking.

Cumulative predictive power
03

Predict what docking cannot

We use chemical structure to predict functional assay performance, not binding affinity. The output is a confidence-ranked shortlist based on predicted activity, early risk signals, and target-agnostic MoA.

60× fewer compounds tested
12 months
Traditional hit identification
vs.
1 month
Reflector-enabled hit-to-lead
Fewer experiments. Better leads. More shots on goal.
Frequently asked questions

What discovery teams ask us

This is the scenario our platform was designed for. Traditional screening evaluates a compound in a single disease-relevant cell type. Our model simultaneously predicts phenotypic response across all cell contexts in our corpus, flagging selectivity issues, off-target toxicity, and context-dependent effects that single-assay workflows cannot detect.
In a retrospective analysis of approved MEK inhibitors, our model identified an off-target toxicity signal in hepatocytes that was not flagged until Phase II clinical trials, at which point the sponsoring company had committed over $40M.
The cost of late-stage failure
Going back to target validation after a failed clinical result costs tens of millions. Functional screening across cell contexts flags these signals during lead optimization, when course-correction is still possible.
Yes. Our foundation model was pre-trained on over 10 million functional experiments, so even a modest archived screen provides sufficient data for fine-tuning. We can re-score the original library against new phenotypic endpoints, identify structurally novel candidates that may have been deprioritized, and screen unsynthesized analogs virtually.
A rare disease biotech provided a <400-compound proprietary screen from an archived program. We identified structurally novel candidates that their internal pipeline had not prioritized, reopening a discovery track that had been considered exhausted.
Why this matters
Most platforms require massive proprietary datasets to deliver results. If you have deep biological expertise but limited screening data, that is precisely where transfer learning from a phenotypically diverse foundation model creates the most value.
Structure-based methods model molecular binding. They do not capture protein expression dynamics, pathway crosstalk, polypharmacology, or cell-context-dependent effects. We model how the cell actually responds to treatment, not what a docking simulation predicts it might do. The two approaches identify complementary candidate sets with minimal overlap.
Collaborators that have integrated our platform alongside existing structural approaches report that the two methods identify complementary candidate sets, suggesting that functional screening captures an entirely different axis of biological relevance.
Complementary, not competing
Our most productive collaborations are with teams that already use structural and target-based tools. Reflector adds the phenotypic layer that structure-based methods cannot provide. The result is compound prioritization that accounts for both binding affinity and functional consequence.

Yes. Our bulk foundation model has already learned shared structure across diverse assay types, including proliferation, morphology, expression, and viability readouts. It transfers knowledge even to assays it has never directly observed. A 100-compound screen with a robust functional readout provides sufficient signal for fine-tuning.

A biotech program used a proprietary assay with no public equivalent. Our model still predicted off-target liabilities across 20+ pharmacological targets and flagged a selectivity concern their internal screening had missed.
Technical note

We represent assays as points in a continuous phenotypic space, not as discrete categories. A novel assay is never truly “unseen” by the model. It is positioned relative to assays already in the training corpus, enabling meaningful predictions from day one.

This is the scenario our platform was designed for. Traditional screening evaluates a compound in a single disease-relevant cell type. Our model simultaneously predicts phenotypic response across all cell contexts in our corpus, flagging selectivity issues, off-target toxicity, and context-dependent effects that single-assay workflows cannot detect.

In a retrospective analysis of approved MEK inhibitors, our model identified an off-target toxicity signal in hepatocytes that was not flagged until Phase II clinical trials, at which point the sponsoring company had committed over $40M
The cost of late-stage failure

Going back to target validation after a failed clinical result costs tens of millions. Functional screening across cell contexts flags these signals during lead optimization, when course-correction is still possible.

Yes. Our foundation model was pre-trained on over 10 million functional experiments, so even a modest archived screen provides sufficient data for fine-tuning. We can re-score the original library against new phenotypic endpoints, identify structurally novel candidates that may have been deprioritized, and screen unsynthesized analogs virtually.

A rare disease biotech provided a <400-compound proprietary screen from an archived program. We identified structurally novel candidates that their internal pipeline had not prioritized, reopening a discovery track that had been considered exhausted.
Why this matters

Most platforms require massive proprietary datasets to deliver results. If you have deep biological expertise but limited screening data, that is precisely where transfer learning from a phenotypically diverse foundation model creates the most value.

 

Structure-based methods model molecular binding. They do not capture protein expression dynamics, pathway crosstalk, polypharmacology, or cell-context-dependent effects. We model how the cell actually responds to treatment, not what a docking simulation predicts it might do. The two approaches identify complementary candidate sets with minimal overlap.

Collaborators that have integrated our platform alongside existing structural approaches report that the two methods identify complementary candidate sets, suggesting that functional screening captures an entirely different axis of biological relevance.
Complementary, not competing

Our most productive collaborations are with teams that already use structural and target-based tools. Reflector adds the phenotypic layer that structure-based methods cannot provide. The result is compound prioritization that accounts for both binding affinity and functional consequence.

 
Supported by
Our scientific roots
Supported by
Our scientific roots

Our model is getting better every week.

We are scaling phenotypic prediction into an infrastructure layer for drug discovery. If you have built ML systems, understand perturbation biology, or know what it takes to close a partnership deal and you want to do it at the stage where your decisions shape the company, reach out.

Your dataset could train the next breakthrough.

We collaborate with academic and industry labs running functional screens. You get predictive analysis and co-publication opportunities. We get the biological diversity that makes our model generalize.