Hans Algorithm and Its Role in Lymphoma Classification

The Hans algorithm is a simple, widely used method for sorting diffuse large B-cell lymphoma (DLBCL) into two biologically distinct subtypes based on staining patterns under a microscope. Developed in the early 2000s, it uses just three protein markers to approximate the results of expensive gene expression profiling, and it remains a workhorse of lymphoma diagnosis worldwide. Yet its accuracy is imperfect, its prognostic value has been questioned in the era of modern immunochemotherapy, and several rival algorithms have tried to dethrone it. Understanding where the Hans algorithm works well and where it falls short matters for anyone navigating a DLBCL diagnosis.

Why Lymphoma Subtypes Matter

DLBCL is the most common aggressive lymphoma, but the name hides a critical distinction. Gene expression profiling studies revealed that what was being called one disease actually encompasses at least two biologically separate entities. One subtype, germinal center B-cell-like (GCB) DLBCL, resembles normal germinal center B cells and tends to respond well to chemotherapy. The other, activated B-cell-like (ABC) DLBCL, resembles B cells that have been activated outside the germinal center and carries a considerably worse prognosis.¹ These two subtypes differ in the expression of hundreds of genes, which translates into different signaling pathways, different vulnerabilities to drugs, and different survival curves after treatment.

The gold standard for identifying these subtypes is gene expression profiling (GEP), a technology that reads thousands of genes simultaneously from tumor tissue. But GEP is expensive, requires specialized equipment and fresh or specially preserved tissue, and is not available in most pathology labs around the world. That gap between the biological reality and everyday clinical practice is exactly the space the Hans algorithm was designed to fill.

How the Algorithm Works

The Hans algorithm relies on immunohistochemistry (IHC), a technique pathologists already use daily. You take a thin slice of tumor tissue, apply antibodies that bind to specific proteins, and look at the staining pattern under a microscope. The Hans algorithm asks about three proteins in a specific order: CD10, BCL6, and MUM1 (also called IRF4).

The decision tree is straightforward. If CD10 is positive, the case is classified as GCB regardless of the other two markers. If CD10 is negative, you check BCL6. If BCL6 is also negative, the case is non-GCB. If BCL6 is positive, you look at MUM1. A positive MUM1 result tips the case into the non-GCB category, while a negative MUM1 keeps it in GCB. The whole process takes a few stained slides and a pathologist’s eye, making it feasible in virtually any hospital with a basic pathology lab.

In the original validation, CD10 expression was the dominant classifier for the GCB group. In one large study, 41 of 53 GCB-classified cases were assigned on the basis of CD10 positivity alone, with the remaining 12 classified as GCB because they were BCL6-positive and MUM1-negative.² That heavy reliance on a single marker is both a strength (simplicity) and a vulnerability (if the CD10 stain is ambiguous, the downstream classification can wobble).

Accuracy Compared with Gene Expression Profiling

The Hans algorithm was never meant to perfectly replicate GEP. It was meant to be a practical stand-in, and in the original study it achieved roughly 86% concordance with the GEP-based classification.³ That sounds respectable, but it means about one in seven cases gets placed in the wrong subtype.

More recent comparisons using the Lymph2Cx assay, a commercially available GEP platform, paint a similar picture. One study of 156 cases found an agreement rate of 86% after excluding cases the Lymph2Cx assay called unclassified. The Hans algorithm classified about a third of cases as GCB and two-thirds as non-GCB, while Lymph2Cx identified an intermediate/unclassified group comprising roughly 12% of cases that the Hans algorithm has no way to recognize.⁴ That intermediate group is clinically meaningful because those patients may behave differently from either pure GCB or pure ABC cases, yet the Hans algorithm forces them into one bin or the other.

A broader analysis found overall concordance between GEP and IHC-based classification of about 72%. GEP confirmed roughly 85% of cases that IHC called GCB, but only 58% of cases IHC called non-GCB were categorized as ABC by GEP.⁵ In other words, when the Hans algorithm says GCB, gene profiling usually agrees. But when it says non-GCB, a substantial fraction of those cases are actually something other than ABC at the molecular level. That asymmetry matters when treatment decisions hinge on subtype.

The Prognostic Question in the Rituximab Era

The original DLBCL survival differences between GCB and ABC subtypes were established in patients treated with CHOP chemotherapy alone, before rituximab became standard. Adding rituximab (creating the R-CHOP regimen) improved outcomes dramatically, especially for the non-GCB group. This raised a pressing question: does the Hans algorithm still predict who does well and who does poorly when everyone gets rituximab?

The evidence is genuinely mixed. One study of R-CHOP-treated patients found no survival difference between GCB and non-GCB groups when classified by the Hans algorithm, with overall survival of 76% versus 77%.⁶ A separate analysis concluded that neither the individual proteins in the algorithm nor the algorithms themselves retained strong prognostic power in patients receiving immunochemotherapy.⁷

Other groups, however, have pushed back. One large series found that non-GCB patients had significantly worse progression-free and overall survival using the Hans, Choi, and Visco-Young algorithms alike, even under R-CHOP.⁸ The discrepancy likely reflects differences in cohort size, patient populations, staining protocols, and the inherent noise in IHC interpretation. Meanwhile, survival differences classified by gene expression profiling remain more consistently significant, which underscores that the biological distinction between GCB and ABC is real, even when the Hans algorithm’s proxy for it is too blunt to detect it reliably.⁵

Rival Algorithms and Why None Has Replaced Hans

Several alternative IHC algorithms have been proposed to improve on Hans. The Choi algorithm adds GCET1 and FOXP1 to the marker panel, and the Visco-Young algorithm uses a different decision tree. A five-marker algorithm incorporating GCET1, CD10, BCL6, MUM1, and FOXP1 achieved 93% concordance with GEP in one validation, outperforming the Hans algorithm’s 86%.³ That same algorithm also performed better for primary mediastinal large B-cell lymphoma, classifying all seven cases as GCB (consistent with the expected biology), whereas Hans classified only two of seven as GCB.

Despite these improvements on paper, none of the alternatives has unseated Hans in routine practice. The reasons are partly practical and partly cultural. GCET1 and FOXP1 antibodies are not stocked in every lab. Pathologists worldwide are already trained on the Hans decision tree. And most clinical trials that stratify by cell-of-origin still use either Hans or GEP, so outcomes data maps back to Hans more cleanly than to the newer algorithms. The Hans algorithm persists less because it is the best and more because it was first, it is cheap, and it is everywhere.

Where the Hans Algorithm Still Guides Treatment

Even if its prognostic value for first-line R-CHOP is debated, the Hans algorithm retains clinical relevance in specific treatment decisions. The most studied example is salvage therapy for relapsed or refractory DLBCL. In the bio-CORAL study, GCB subtype as classified by Hans was significantly associated with better progression-free survival in the R-DHAP salvage arm, an interaction that held in multivariate analysis.⁹ This suggests that different salvage regimens may work better depending on the Hans-classified subtype.

Lenalidomide, an immunomodulatory drug, has shown preferential activity in non-GCB DLBCL. A retrospective study of relapsed or refractory patients treated with lenalidomide found a response rate of 53% in non-GCB patients compared with just 9% in GCB patients. A subsequent phase II/III trial confirmed that the progression-free survival benefit of lenalidomide over alternative agents was substantially larger in non-GCB cases than in GCB cases.¹⁰ In these settings, the Hans classification directly influences whether a particular drug is likely to help.

CAR-T cell therapy, however, may be a different story. An analysis of patients with relapsed or refractory aggressive B-cell lymphomas treated with axicabtagene ciloleucel (axi-cel) found no association between Hans-defined cell-of-origin and outcomes.¹¹ CAR-T works by a fundamentally different mechanism, redirecting a patient’s own immune cells to attack the tumor, which may explain why the GCB versus non-GCB distinction becomes less relevant.

Technical Variability and the Reproducibility Problem

One underappreciated weakness of the Hans algorithm is that it is only as good as the staining quality in a given lab. IHC is not a standardized assay in the way that a molecular test is. Different antibody clones, different dilutions, different antigen-retrieval methods, and different thresholds for calling a stain “positive” all introduce variability. A Canadian harmonization study found that participating labs produced IHC results ranging from “weak” to “strong” analytical sensitivity, and the researchers had to adjust readout thresholds (using a cutoff of greater than 30% at moderate-to-strong staining intensity) just to get comparable results across sites.¹²

A broader study comparing nine different IHC-based classifiers on the same set of tumors found that concordance between algorithms was poor, with the Hans allocation sometimes differing sharply from the Choi or Tally classifier on the same tissue section.² When different algorithms applied to the same tumor give different answers, the problem is not just the algorithm. It is the fundamental challenge of turning a visual staining pattern into a binary molecular call. Even among experienced pathologists, the cutoff for “positive” BCL6 or MUM1 staining is subjective, and small differences in interpretation can flip a patient from GCB to non-GCB.

Digital Pathology and the Push Toward Automation

One response to the reproducibility problem is to remove the human eye from the equation. Digital pathology uses computer image analysis to quantify staining intensity and the percentage of positive cells, potentially standardizing the readout across institutions. An early proof-of-concept study applied digital image analysis to the three Hans markers and achieved strong performance for each individual stain, with area-under-the-curve values of 0.92 for CD10, 1.0 for BCL6, and 0.95 for MUM1 when predicting pathologist classifications. The algorithm-based classification agreed with pathologist classification in about 83% of cases.¹³

That 83% figure is not perfect, but the appeal of digital pathology is consistency. A computer will apply the same threshold every time, eliminating the lab-to-lab and pathologist-to-pathologist drift that plagues manual scoring. As these tools mature and are integrated into clinical workflows, they could make IHC-based subtyping more reliable without requiring the leap to full gene expression profiling.

The Hans Algorithm in CNS Lymphoma

Primary central nervous system lymphoma (PCNSL) is a rare form of DLBCL that arises in the brain or spinal cord, and its biology differs in several ways from systemic DLBCL. The Hans algorithm has been applied in this setting too, but the results highlight how context shapes interpretation. One study of 24 PCNSL patients found that all patients with GCB-type disease were still alive at a median follow-up of 22 months, whereas only 39% of non-GCB patients were alive, with a median survival of 11 months.¹⁴

However, the distribution of subtypes in CNS lymphoma skews heavily toward non-GCB. Genomic profiling found that roughly 72% of primary CNS lymphoma cases were non-GCB by Hans classification, compared with about 47% of systemic DLBCL cases.¹⁵ A larger series from India reported an even more extreme distribution, with about 80% of PCNSL cases classified as non-GCB, though in that cohort no significant survival difference between subtypes was observed.¹⁶ The inconsistency underscores that small cohort sizes and different treatment practices make it hard to draw firm conclusions about prognostic value in rarer lymphoma subtypes.

Connection to High-Risk Genetic Rearrangements

Beyond cell-of-origin, oncologists increasingly look for specific genetic events, particularly rearrangements of the MYC, BCL2, and BCL6 genes. Tumors harboring concurrent MYC and BCL2 rearrangements (so-called “double-hit” lymphomas) are especially aggressive. The Hans algorithm intersects with this biology because double-hit lymphomas tend to cluster in the GCB subtype. A screening study confirmed that double-hit lymphomas showed a significant positive correlation with GCB phenotype and with co-expression of MYC and BCL2 proteins.¹⁷

This is clinically useful because a GCB classification by Hans can serve as a flag to pursue further testing for MYC and BCL2 rearrangements, especially in resource-limited settings where running fluorescence in situ hybridization on every case is not feasible. In this way, the Hans algorithm functions not just as a subtyping tool but as a triage step that helps pathologists decide which tumors warrant more expensive molecular workup.

Primary Cutaneous Lymphoma and the Limits of Generalization

When the Hans algorithm is applied outside its original context of systemic DLBCL, results can diverge sharply from GEP. A study of primary cutaneous large B-cell lymphomas found near-perfect agreement between Hans and Lymph2Cx for one subtype (primary cutaneous follicle center lymphoma, where 100% of cases were GCB by both methods) but marked discordance for another (primary cutaneous DLBCL, leg type). In that group, the Hans algorithm classified 75% as non-GCB and 25% as GCB, while the Lymph2Cx assay classified only 18% as ABC, 39% as GCB, and a striking 43% as unclassified.¹⁸

Nearly half of the leg-type cases fell into a molecular category that the Hans algorithm cannot recognize, illustrating how the simple three-marker tree can fail in tumor types it was not designed for. The GCB and ABC subtypes that the algorithm was built to distinguish were defined in systemic DLBCL, and the biology of cutaneous lymphomas involves overlapping but not identical gene expression patterns. Applying Hans outside systemic DLBCL is common, but clinicians should interpret the results with extra caution.

Different Signaling Profiles Between Subtypes

The GCB versus non-GCB distinction is not just a label. The two subtypes rely on different molecular wiring. ABC-type DLBCL typically depends on chronic active B-cell receptor signaling and NF-κB activation, while GCB-type tends to involve different pathways. Preclinical work has shown that ABC and GCB cell lines display distinct JAK/STAT and B-cell receptor signaling profiles, even when key enzymes like SYK and STAT3 are active in both.¹⁹ These pathway-level differences explain why drugs targeting specific signaling nodes, such as BTK inhibitors or lenalidomide, tend to work better against one subtype than the other. They also explain why the subtyping question remains so clinically important, even as the best method for answering it continues to evolve.